AI RADAR
中EN

All AI updates

1738 itemsGetting the latest content

Browse the complete event stream by source type, topic or keyword.

Sources
Topics

Latest

Tue, Sep 29

Tue, Sep 29

Mon, Sep 28

llama.cpp1 source

b11242

<details open> mtmd: fix GCC 15 stringop-overflow in decode_embd_batch (#29607) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50949667> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **L

Other1 source
Sources & timeline
llama.cppb11242
llama.cpp1 source

b11240

<details open> server : support typed content (vision/audio/video) input for /v1/embeddings endpoint (#29556) * server : support multimodal input for /v1/embeddings (Qwen3-VL-Embedding) Accept the OpenAI-style wrapped content array format for multimodal embedding requests. Each {"content": [...]} object is one input that produces one embedding; text parts are concatenated and image_url parts are decoded via handle_media then spliced with process_mtmd_prompt. The legacy formats (plain string, token arrays, mixed arrays, and the {prompt_string, multimodal_data} object) continue to work unchanged via tokenize_input_prompts. Bare content arrays (the unwrapped shape) are rejected with a migration

Other1 source
Sources & timeline
llama.cppb11240
llama.cpp1 source

b11239

<details open> vulkan: include functional header (#29597) Fixes the compile error: `no template named 'function' in namespace 'std'` </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50927029> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/relea

Other1 source
Sources & timeline
llama.cppb11239
The Verge AI1 source

AMD is acquiring AI company World Labs in a deal worth more than $8 billion

AMD announced today that it's acquiring World Labs, an AI research lab co-founded by the prominent researcher Dr. Fei-Fei Li, in an all-stock deal worth approximately $8.2 billion. World Labs launched in 2024 and was valued at $1 billion in a matter of months. The startup launched its first commercial product, a world generation model […]

Other1 source
Sources & timeline
The Verge AIAMD is acquiring AI company World Labs in a deal worth more than $8 billion
Ollama1 source

v0.35.0

Decision models Ollama now supports decision models through `/v1/systemone`, based on TypeSafe’s Jev API. Decision models return choices, probabilities, and scores instead of text. Use them for tasks such as ticket triage, model routing, and content classification. Available models: - **Nimble** from Bespoke Labs - **Tev1** from Together AI ```sh ollama pull nimble ``` Send context and one or more questions: ```sh curl http://localhost:11434/v1/systemone \ -H 'Content-Type: application/json' \ -d '{ "model": "nimble", "state": "Our checkout has returned 500 errors since 9am.", "questions": { "la

Other1 source
Sources & timeline
Ollamav0.35.0
Ollama1 source

v0.35.0

Decision models Ollama now supports decision models through `/v1/systemone`, based on TypeSafe’s Jev API. Decision models return choices, probabilities, and scores instead of text. Use them for tasks such as ticket triage, model routing, and content classification. Available models: - **Nimble** from Bespoke Labs - **Tev1** from Together AI ```sh ollama pull nimble ``` Send context and one or more questions: ```sh curl http://localhost:11434/v1/systemone \ -H 'Content-Type: application/json' \ -d '{ "model": "nimble", "state": "Our checkout has returned 500 errors since 9am.", "questions": { "la

Other1 source
Sources & timeline
Ollamav0.35.0
The Verge AI1 source

OpenAI’s AI agents need to catch up

OpenAI popularized the modern generative AI chatbot, but as its 2026 DevDay event approaches, it's fallen behind in one of the industry's hottest categories: continuously running, consumer-facing AI agents. On Tuesday, it will likely try to capture the lead in that race. Rumors abound that OpenAI will release its own AI agent, dubbed Aeon - […]

Other1 source
Sources & timeline
The Verge AIOpenAI’s AI agents need to catch up
The Verge AI1 source

AI is supercharging hacking, and your local hospitals and banks aren’t ready

In March, Janice Malone began getting calls about suspicious activity from her nonprofit organization, Vivian's Door. Vivian's Door, headquartered in Alabama, typically provided training, resources, and community to underserved and minority-owned businesses. The work sometimes put it in close contact with these companies' financial data, which was stored on its systems. But suddenly, concerned callers […]

Other1 source
Sources & timeline
The Verge AIAI is supercharging hacking, and your local hospitals and banks aren’t ready
Anthropic Python SDK1 source

v1.9.0

1.9.0 (2026-09-28) Full Changelog: v1.8.0...v1.9.0 Features * **api:** add between_tools thinking type (a9a577d) * **api:** add claude-sonnet-5-5 (a5a25e9) * **api:** add include_inherited and source to workspace rate limits (0bf4af8) * **api:** add typed event type values to the Managed Agents events list f

Other1 source
Sources & timeline
Anthropic Python SDKv1.9.0
MIT Technology Review AI1 source

When can we say AI made a scientific discovery?

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Last Wednesday, Anthropic announced that earlier this year it had launched a molecular biology lab, where Claude agents read and conjecture about hard biology problems and human scientists run experiments on what…

Other1 source
Sources & timeline
MIT Technology Review AIWhen can we say AI made a scientific discovery?
OpenAI Python SDK1 source

v3.20.0

3.20.0 (2026-09-28) Features * **api:** add Agents credential and session options (#3967) (bb68198) * **api:** add Cyber access programs to Responses (#3956) (09c5b6f) * **responses:** opt in to incremental WebSocket text and tool snapshots (#3973) ([d0207b4](https://github.co

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.20.0
Google Gemini Blog1 source

3 ways this grocer cooks for 200 guests with Gemini

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Thumbnail_-_16x9.max-600x600.format-webp.webp">Edy Massih uses Gemini to transform regular recipes into massive catering spreads, automate grocery orders, and adapt menus for dietary restrictions.

Other1 source
Sources & timeline
Google Gemini BlogSee what 4 builders are making with Gemini 3.8 FlashGoogle Gemini Blog3 ways this grocer cooks for 200 guests with Gemini
MIT Technology Review AI1 source

Who’s liable when AI agents go rogue?

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. Over the past few months, a cascade of cyberattacks by AI agents has stunned the world. In July, OpenAI disclosed that a swarm of its agents…

Other1 source
Sources & timeline
MIT Technology Review AIWho’s liable when AI agents go rogue?
arXiv cs.CL1 source

A Mechanistic Study of AI-Text Detection Neurons in Frozen BERT: Sparse Probing and refresh Patching on RAID

This public update concerns “A Mechanistic Study of AI-Text Detection Neurons in Frozen BERT: Sparse Probing and refresh Patching on RAID”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLA Mechanistic Study of AI-Text Detection Neurons in Frozen BERT: Sparse Probing and refresh Patching on RAID
arXiv cs.CV1 source

CSCWD: Cross-Scale Channel-wise Knowledge Distillation for Lightweight Tiny Object Detection on Edge Devices

This public update concerns “CSCWD: Cross-Scale Channel-wise Knowledge Distillation for Lightweight Tiny Object Detection on Edge Devices”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVCSCWD: Cross-Scale Channel-wise Knowledge Distillation for Lightweight Tiny Object Detection on Edge Devices
arXiv cs.LG1 source

HybridInfer: Thermal-Aware Reinforcement-Learning Tier Routing for On-Device, Edge, and Cloud LLM Inference

This public update concerns “HybridInfer: Thermal-Aware Reinforcement-Learning Tier Routing for On-Device, Edge, and Cloud LLM Inference”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGHybridInfer: Thermal-Aware Reinforcement-Learning Tier Routing for On-Device, Edge, and Cloud LLM Inference
arXiv cs.CL1 source

Not All Memories Are Equal: Hierarchical Collaborative Memory for Validity-Aware Retrieval in LLM Agents

This public update concerns “Not All Memories Are Equal: Hierarchical Collaborative Memory for Validity-Aware Retrieval in LLM Agents”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLNot All Memories Are Equal: Hierarchical Collaborative Memory for Validity-Aware Retrieval in LLM Agents
arXiv cs.AI1 source

When Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess

This public update concerns “When Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIWhen Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess
arXiv cs.LG1 source

ENAS: An Efficient Hardware-Aware Neural Architecture Search Framework for TinyML on Resource-Constrained Microcontrollers

This public update concerns “ENAS: An Efficient Hardware-Aware Neural Architecture Search Framework for TinyML on Resource-Constrained Microcontrollers”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGENAS: An Efficient Hardware-Aware Neural Architecture Search Framework for TinyML on Resource-Constrained Microcontrollers
arXiv cs.LG1 source

When the Preconditioning Exponent Turns Negative: Learning-Rate Coupling and Cross-Environment Generalization

This public update concerns “When the Preconditioning Exponent Turns Negative: Learning-Rate Coupling and Cross-Environment Generalization”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGWhen the Preconditioning Exponent Turns Negative: Learning-Rate Coupling and Cross-Environment Generalization

Sun, Sep 27

llama.cpp1 source

b11223

<details open> server : allow RANK pooling batch splitting for causal LLM rerankers (ie. Qwen3 and Qwen3-VL) (#28876) * server : allow splitting RANK pooling for causal LLM rerankers Rerank models fall into two categories: bidirectional cross-encoders (BERT, etc.) that require all tokens in a single physical batch, and causal LLMs repurposed as rerankers (Qwen3, Qwen3-VL) that can use chunked prefill like any other decoder. Previously the server rejected all RANK-pooling inputs larger than n_ubatch, and the graph builder hardcoded QWEN3/QWEN3VL arch checks to determine last-token pooling. This broke long-document and multimodal reranking for causal models. Fix: expose llama_get_causal_attn(c

Other1 source
Sources & timeline
llama.cppb11223
The Verge AI1 source

Engram is a sampler that turns broken AI hallucinations into music

Music startup Thoughtful Things has just launched the Kickstarter campaign for its first instrument, Engram. It's a sampler and groovebox that uses AI to mangle incoming audio and even hallucinate completely new sounds. This isn't Suno in a box, though. This isn't a "push-button, get-song" device, aimed at creating something that sounds ready for top-40 […]

Other1 source
Sources & timeline
The Verge AIEngram is a sampler that turns broken AI hallucinations into music
llama.cpp1 source

b11222

<details open> common : avoid side effects around params parsing (#29537) - register --rpc unconditionally and call llama_supports_rpc() only from its handler - print server "initialization ..." log after args are parsed Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-RL </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50580981> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/

Other1 source
Sources & timeline
llama.cppb11222
The Verge AI1 source

OpenAI agents tried to ‘bruteforce’ a UN website

Security researcher Rowan Howard-Jones says that OpenAI agents scanned the UN Conference on Trade and Development's (UNCTAD) statistics site over 16,000 times between April and June. While the incident doesn't quite rise to the level of the Hugging Face hack, or the recent attacks on US government sites, it's yet another concerning example of AI […]

Other1 source
Sources & timeline
The Verge AIOpenAI agents tried to ‘bruteforce’ a UN website
llama.cpp1 source

b11221

<details open> common : make string_split<T> throw on invalid values (#29518) Signed-off-by: Adrien Gallouët <angt@huggingface.co> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50573055> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/release

Other1 source
Sources & timeline
llama.cppb11221

Sat, Sep 26

llama.cpp1 source

b11205

<details open> cuda: support Nemotron 3 Puzzle state size 96 for ssm scan (#28717) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50461180> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework

Other1 source
Sources & timeline
llama.cppb11205
llama.cpp1 source

b11203

<details open> cuda: add F16 input to the FWHT (#29096) * cuda: add F16 input to the FWHT The CUDA FWHT accepts F32 input only. This makes the source type a template parameter, so the kernel reads an F16 source directly instead of requiring a converted copy. The F32 path is unchanged. supports_op accepts an F16 src1 against an F32 src0 for the Hadamard hint. Every other F16 src1 against a non-F16 src0 is still refused. ggml_cuda_op_mul_mat_use_fwht is the single predicate both supports_op and the dispatch call now share, checking contiguity and same-shape(src1, dst) in addition to the type/hint conditions above. Without a shared predicate, supports_op could admit an op that ggml_cuda_op_fwht

Other1 source
Sources & timeline
llama.cppb11203
llama.cpp1 source

b11202

<details open> server : fix wake_fd warning on Windows (#29479) Signed-off-by: Adrien Gallouët <angt@huggingface.co> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50455005> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b11

Other1 source
Sources & timeline
llama.cppb11202
The Verge AI1 source

OpenAI pauses training of its ‘most capable models’

As reports of OpenAI's models breaking containment, hacking sites, and generally getting out of control pile up, the company has made the decision to pause training of its most powerful models. The decision was made after a model being tested within a sandbox exploited a loophole to gain internet access. The incident happened on September […]

Safety1 source
Sources & timeline
The Verge AIOpenAI pauses training of its ‘most capable models’
The Verge AI1 source

Can Cloudflare CEO Matthew Prince save the web from AI?

Today, I’m talking with Matthew Prince, who is CEO of Cloudflare. This episode is part of a two-part series on the future of business. Matthew last joined us on the show about two and a half years ago, at what we thought then was a wild pivot point for the internet — and now it […]

Other1 source
Sources & timeline
The Verge AICan Cloudflare CEO Matthew Prince save the web from AI?
llama.cpp1 source

b11192

<details open> vendor : update cpp-httplib to 0.58.0 (#29407) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50337762> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu

Other1 source
Sources & timeline
llama.cppb11192

Fri, Sep 25

llama.cpp1 source

b11191

<details open> common,rpc : simplify fs_create_directory_with_parents() (#29432) The original function was broken on Windows for some unicode paths Paths without a trailing separator now create the last directory too, matching the function name. All current callers already include a trailing separator, so this change does not affect them. Signed-off-by: Adrien Gallouët <angt@huggingface.co> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50291720> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, K

Other1 source
Sources & timeline
llama.cppb11191
llama.cpp1 source

b11190

<details open> mtmd: fix mel preprocessor in LFM2 audio (#29403) which resulted in different greedy transcripts for 4.5% of English and 6.5% of Japanese test utterances. In Japanese, some differences changed entire words. This change: * uses `log(x + 2^-24)` instead of clamping to the log floor * uses a symmetric Hann window, equivalent to `torch.hann_window(periodic=False)` * adds the normalization epsilon to the standard deviation instead of inside the square root Only the `lfm2a` preprocessor opts into these behaviors. Other audio preprocessors are unchanged. Tested on top of 84e76d8 using `llama-server` with CUDA and `temperature=0`, compared against http://github.com/Liquid4All/liquid-a

Other1 source
Sources & timeline
llama.cppb11190
The Verge AI1 source

Meta makes the Muse filesystem even more accessible

Yesterday, with a little prodding, it was discovered that Meta's Muse would expose its filesystem to curious users. The files offered a fascinating peek under the hood of an AI chatbot, and appeared to expose details we weren't meant to see, not least because Muse itself told people, including us, it wasn't supposed to reveal […]

Other1 source
Sources & timeline
The Verge AIMeta makes the Muse filesystem even more accessible
The Verge AI1 source

Sony and UMG are suing Suno again

Sony and Universal Music Group filed yet another suit against Suno. The labels claim its new v6 model still infringes on their copyrights because it's trained on user outputs from previous models, which were themselves trained on unlicensed music ripped from YouTube and other sources. Sony and UMG are notable holdouts who did not sign […]

Regulation1 source
Sources & timeline
The Verge AISony and UMG are suing Suno again
The Verge AI1 source

One company is at the center of a wave of rogue AI attacks

In July, OpenAI revealed that its AI agents had attacked Hugging Face without permission, sparking widespread concerns about AI safety. Since then, a string of similar incidents involving agents from Meta, Anthropic, Google, and other companies has fueled further fears about rogue AI. As disclosures implicating numerous AI models trickled out over the past few […]

Other1 source
Sources & timeline
The Verge AIOne company is at the center of a wave of rogue AI attacks
MIT Technology Review AI1 source

The Pentagon wants $30 million to build an AI-powered lie detector

The US government wants to spend $30.3 million over the next five years on an improved form of lie detector, according to a Department of Defense budget request. The program, called Polygraph+ or Polygraph Next, will focus on scoring algorithms that use artificial intelligence and machine learning and on a technique called “standoff sensing,” which…

Other1 source
Sources & timeline
MIT Technology Review AIThe Pentagon wants $30 million to build an AI-powered lie detector
arXiv cs.CL1 source

Framing by Wording, Framing by Selection: A Large-Scale Two-Dimensional Audit of French News Headlines, 2022-2025

This public update concerns “Framing by Wording, Framing by Selection: A Large-Scale Two-Dimensional Audit of French News Headlines, 2022-2025”. Open the original source for capabilities, limitations and impact.

Regulation1 source
Sources & timeline
arXiv cs.CLFraming by Wording, Framing by Selection: A Large-Scale Two-Dimensional Audit of French News Headlines, 2022-2025
arXiv cs.AI1 source

TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split

This public update concerns “TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AITW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split
arXiv cs.LG1 source

CFD Correction of Open Tip Clearance Flow in a Compressor Cascade Using VAE Latent Space Adaptation

This public update concerns “CFD Correction of Open Tip Clearance Flow in a Compressor Cascade Using VAE Latent Space Adaptation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGCFD Correction of Open Tip Clearance Flow in a Compressor Cascade Using VAE Latent Space Adaptation
Ollama1 source

v0.40.0

What's Changed **Models run on MLX on Apple Silicon by default** In this release, on Apple Silicon devices, model architectures supported by the MLX runtime will automatically run on MLX. ``` ollama pull qwen3.8 ollama run qwen3.8 ``` During the pre-release we will be testing and enabling additional models. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.4...v0.40.0-rc0

Other1 source
Sources & timeline
Ollamav0.40.0
llama.cpp1 source

b11173

<details open> metal : fix graph capture and handle empty graphs (#29390) - return early when the graph has no nodes - drop the redundant reset of capture_compute: the decrement at the top of the function already transitions the counter from 0 to -1, so a capture happens exactly once - hint at METAL_CAPTURE_ENABLED=1 in the capture error message - pass capture_compute == 0 (not the raw counter) as use_capture to ggml_metal_op_init, so GPU debug-group markers are only emitted on the captured compute Assisted-by: pi:llama.cpp/Qwen3.8-27B </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50032954> **macOS/iOS:** - [macOS Apple

Other1 source
Sources & timeline
llama.cppb11173

Thu, Sep 24

llama.cpp1 source

b11172

<details open> metal : optimize sparse FA + clean-up (#29377) * metal : cache sparse FA indices in shared memory Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * metal : simplify shared memory size calculation Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * pi : update general * metal : unroll sparse index load Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50026590> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI e

Other1 source
Sources & timeline
llama.cppb11172
llama.cpp1 source

b11171

<details open> sync : ggml (#29396) * ggml : bump version to 0.25.2 (ggml/1642) * ggml : fix ubsan error in `ggml_graph_nbytes` (ggml/1644) * ggml : bump version to 0.25.3 (ggml/1645) * sync : ggml </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/50020953> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/b11171/llama-b11171-bin-macos-x64.tar.gz

Other1 source
Sources & timeline
llama.cppb11171
The Verge AI1 source

Gemini 3.8 Live with Live Avatar gives Google’s AI a face

Google's new Gemini 3.8 Live update lets users have conversations with the model while watching an animated AI persona respond in real time. The "Live Avatar" will lip-sync and show different facial expressions during conversations, but it's currently only available to Gemini Enterprise customers. As noted by Google, Live Avatar can transition between the 97 […]

Other1 source
Sources & timeline
The Verge AIGemini 3.8 Live with Live Avatar gives Google’s AI a face
The Verge AI1 source

Jensen Huang talks about AI and climate change like a supervillain

As Jensen Huang puts it, AI can help fight climate change - but only if it inflicts "an enormous amount of pain and suffering" first. The Nvidia CEO discussed the future of energy and AI's impact on our planet in the latest episode of The Ezra Klein Show. But his comments boil down to the […]

Other1 source
Sources & timeline
The Verge AIJensen Huang talks about AI and climate change like a supervillain
The Verge AI1 source

Meta is going to let you build games with AI right on your phone

Meta has a new plan to get people to make games for its Horizon social platform. The company today announced two new development tools that will let you create games with AI prompts: Horizon Create, a mobile app, and Horizon Studio, a browser app that offers more granular controls. The apps will be available in […]

Other1 source
Sources & timeline
The Verge AIMeta is going to let you build games with AI right on your phoneThe Verge AIMeta is making a standalone Muse AI gadgetThe Verge AIMuse is coming to Meta smart glassesThe Verge AIMeta ditches the camera on its newest smart glasses
Google Gemini Blog · Google DeepMind Blog2 sources

Introducing Gemini 3.8 Live with Live Avatar

Introducing Gemini 3.8 Live with Live Avatar

Other2 sources
Sources & timeline
Google Gemini BlogIntroducing Gemini 3.8 Live with Live AvatarGoogle DeepMind BlogIntroducing Gemini 3.8 Live with Live AvatarGoogle DeepMind BlogIntroducing Gemini 3.8 Live and 3.8 Live Extended ThinkingGoogle DeepMind BlogIntroducing Gemini 3.8 Flash and 3.8 Flash CyberGoogle DeepMind BlogIntroducing agentic video understanding with GeminiGoogle Gemini BlogIntroducing Gemini 3.8 Live and 3.8 Live Extended Thinking
arXiv cs.CV1 source

AgroBench: A Reproducible Multimodal Benchmark for Weakly Supervised Crop Yield Learning from County Statistics and Pixel Observations

This public update concerns “AgroBench: A Reproducible Multimodal Benchmark for Weakly Supervised Crop Yield Learning from County Statistics and Pixel Observations”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVAgroBench: A Reproducible Multimodal Benchmark for Weakly Supervised Crop Yield Learning from County Statistics and Pixel Observations
arXiv cs.CL1 source

Experts Rise Where LLMs Disagree: Using Cross-Model Disagreement to Target Expert Effort in LLM Codebook Revision for Large-Scale Annotation

This public update concerns “Experts Rise Where LLMs Disagree: Using Cross-Model Disagreement to Target Expert Effort in LLM Codebook Revision for Large-Scale Annotation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLExperts Rise Where LLMs Disagree: Using Cross-Model Disagreement to Target Expert Effort in LLM Codebook Revision for Large-Scale Annotation
arXiv cs.CV1 source

A 3D Pose-Based Ensemble Framework for Cricket Shot Classification and Automated Biomechanical Analysis

This public update concerns “A 3D Pose-Based Ensemble Framework for Cricket Shot Classification and Automated Biomechanical Analysis”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVA 3D Pose-Based Ensemble Framework for Cricket Shot Classification and Automated Biomechanical Analysis
arXiv cs.CV1 source

Cross-Modal Contrastive Learning from Histopathology and CT for Automated Renal Cell Carcinoma Grading

This public update concerns “Cross-Modal Contrastive Learning from Histopathology and CT for Automated Renal Cell Carcinoma Grading”. Open the original source for capabilities, limitations and impact.

Regulation1 source
Sources & timeline
arXiv cs.CVCross-Modal Contrastive Learning from Histopathology and CT for Automated Renal Cell Carcinoma Grading
arXiv cs.LG1 source

Signal2Symbol: Neuro-Symbolic Temporal Reasoning for Explainable Physiological Time-Series Anomaly Detection

This public update concerns “Signal2Symbol: Neuro-Symbolic Temporal Reasoning for Explainable Physiological Time-Series Anomaly Detection”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGSignal2Symbol: Neuro-Symbolic Temporal Reasoning for Explainable Physiological Time-Series Anomaly Detection
OpenAI Python SDK1 source

v3.19.2

3.19.2 (2026-09-23) Bug Fixes * preserve single files for fallback extraction paths (#3875) (bfd3680) Chores * **api:** clarify approximate web search location defaults (#3953) (a95c95e) * **api:** clarify Realtime modality array definitions (#3954) ([e79cf53](https://gith

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.19.2

Wed, Sep 23

llama.cpp1 source

b11149

<details open> tests: add `-b/--backend` option to test-llama-archs for testing a specific backend (#27372) * tests: add backend option to test-llama-archs * Update tests/test-llama-archs.cpp Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * remove extra space --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/49661383> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pul

Other1 source
Sources & timeline
llama.cppb11149
llama.cpp1 source

v0.5.0

Overview This release focuses on backend performance and correctness, broader model coverage, and more robust server/router operation. It adds HRM-Text (DFM Mimir 1B) support, MiMo-V2.6 and HunyuanOCR conversion support, ggml 0.25.0 backend improvements, multi-address HTTP binding, image outputs from function calls, and several chat parser/UI fixes. Highlights - Accelerate CUDA `conv2d` with implicit GEMM (#29135) - Add Metal MoE and SSM_CONV fusion optimizations (#28948) - Allow the server to bind to multiple addresses (#28690) ##

Other1 source
Sources & timeline
llama.cppv0.5.0
OpenAI Python SDK1 source

v3.19.1

3.19.1 (2026-09-23) Bug Fixes * **chat:** preserve single-pass tool iterables (#3770) (33ffa1f) * **client:** merge HTTP headers case-insensitively (#3486) (5e39766) Chores * **api:** clarify Chat Completions seed limits (#3945) ([be9d666](https://github.com/openai/openai-

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.19.1
llama.cpp1 source

b11147

<details open> opencl: add A8 Q6_K non-MoE dp4a binary kernel (#29057) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/49630180> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** -

Other1 source
Sources & timeline
llama.cppb11147
MIT Technology Review AI1 source

The AI Hype Index: AI loves cheating

Brace yourself: It turns out AI is being optimized for cheating. OpenAI’s agents hacked into Hugging Face to get the answers to a cybersecurity test. Next, they solved a prestigious math problem (or just stole from two top mathematicians’ answer sheets). Anthropic’s models have also hacked into other companies’ systems four times already. And that’s…

Other1 source
Sources & timeline
MIT Technology Review AIThe AI Hype Index: AI loves cheating
arXiv cs.LG1 source

"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and refresh Steering Reproduces It

This public update concerns “"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and refresh Steering Reproduces It”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LG"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and refresh Steering Reproduces It
arXiv cs.LG1 source

Entropy Can Flow, or It Can Guide. Be Entropy. LEDFlow: Introducing Entropy-guided Generation Order into Uniform Discrete Flow

This public update concerns “Entropy Can Flow, or It Can Guide. Be Entropy. LEDFlow: Introducing Entropy-guided Generation Order into Uniform Discrete Flow”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGEntropy Can Flow, or It Can Guide. Be Entropy. LEDFlow: Introducing Entropy-guided Generation Order into Uniform Discrete Flow
arXiv cs.AI1 source

Do Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization

This public update concerns “Do Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIDo Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization
arXiv cs.AI1 source

Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation

This public update concerns “Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIDo Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation
arXiv cs.CL1 source

What Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus

This public update concerns “What Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLWhat Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus
Ollama1 source

v0.34.4

What's Changed * server: fix intermittent "model not found" errors. by @rick-github in https://github.com/ollama/ollama/pull/18438 * server: apply structured outputs in a single pass on thinking models by @jessegross in https://github.com/ollama/ollama/pull/18479 * app: avoid System Events for ChatGPT/Codex detection by @hoyyeva in https://github.com/ollama/ollama/pull/18601 * llama.cpp: version update by @dhiltgen in https://github.com/ollama/ollama/pull/18577 * MLX: version bump by @dhiltgen in https://github.com/ollama/ollama/pull/18576 * mlx: select Gemma 4 image resolution dynamically by @dhiltgen in https://github.com/ollama/ollama/pull/18603 * mlx: speed up Qwen 3.8 prompt processi

Other1 source
Sources & timeline
Ollamav0.34.4
Ollama1 source

v0.34.4

What's Changed - Structured outputs on thinking models now apply in a single pass, making them faster and more reliable. - Fixed intermittent "model not found" errors with a large local library - Fixed the macOS app becoming unresponsive when checking if ChatGPT or Codex is running. - Qwen 3.8 prompt processing is faster on Apple Silicon. - Gemma 4 on Apple Silicon now picks the best image resolution per image, keeping more detail in high-resolution images. - Updated llama.cpp, MLX, and XGrammar. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.3...v0.34.4

Other1 source
Sources & timeline
Ollamav0.34.4
The Verge AI1 source

OpenAI wants to consult elite mathematicians about how to not fumble again

After turning a string of spectacular mathematical results into a reputational crisis, OpenAI is consulting human mathematicians to help it figure out a less disastrous path forward. On Monday, the company announced a new independent panel of mathematicians tasked with advising it and other AI companies on their interactions with mathematical research and the wider […]

Other1 source
Sources & timeline
The Verge AIOpenAI wants to consult elite mathematicians about how to not fumble again
OpenAI Python SDK1 source

v3.19.0

3.19.0 (2026-09-22) Features * **api:** add GCP external storage support (#3943) (d12d60f) * **api:** add GPT-Rosalind research model (#3940) (d041d73) Bug Fixes * **_utils/_transform:** propagate __api_exclude__ in _async_transform_recursive (#3324) ([161ae65](https://git

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.19.0

Tue, Sep 22

TechCrunch AI1 source

TechCrunch Founder Summit’s agenda revealed: Unlock fundraising, hiring, and AI insights in Boston on November 4

Founders shouldn't have to learn the hardest lessons the hardest way. TechCrunch Founder Summit is designed to make the challenges of starting a company easier and the highs that much greater.

Other1 source
Sources & timeline
TechCrunch AITechCrunch Founder Summit’s agenda revealed: Unlock fundraising, hiring, and AI insights in Boston on November 4
llama.cpp1 source

b11115

<details open> opencl: add bin kernel `kernel_gemm_noshuffle_q4_k_q8_1_dp4a_ila_a8_bin` (#29056) * opencl: add A8 Q4_K non-MoE dp4a binary kernel * opencl: rename binary kernel selection helpers </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/49361383> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) -

Other1 source
Sources & timeline
llama.cppb11115
llama.cpp1 source

b11114

<details open> server: fix router eviction races with the existing queue (#29217) * server: route every model load through the queue A model loaded by the fast path has no queue entry, so tick() evicts it at its LOADED transition before its own request is proxied. Every load now joins the queue, whose entry protects the model until its waiters leave. * server: do not admit requests into a stopping model A request for a model that is being stopped still sees it LOADED and is proxied into the dying child. Such a request now joins the queue and is served by the next instance. The stopping mark is cleared under the same lock that sets UNLOADED, so no request can see a model that is neither stopp

Other1 source
Sources & timeline
llama.cppb11114
The Verge AI1 source

Rabbit’s new AI agent doesn’t need an R1 to run

Rabbit, the company behind the underwhelming R1 device, is rolling out a standalone AI agent that you don't need its hardware to use, as reported earlier by Wired. The startup says its new OS3 "agentic operating system" runs in the cloud but operates locally across Windows, Mac, and Linux devices. According to Rabbit, you can […]

Other1 source
Sources & timeline
The Verge AIRabbit’s new AI agent doesn’t need an R1 to run
llama.cpp1 source

b11113

<details open> server: do not pass log file to children (#29212) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/49344105> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubun

Other1 source
Sources & timeline
llama.cppb11113
Microsoft Foundry Blog1 source

Claude Opus 5.5 comes to Microsoft Foundry for long-running coding and knowledge work | Microsoft Community Hub

AI models are increasingly taking on work that extends far beyond a single prompt: building a feature across a codebase, investigating a complex issue,...

Regulation1 source
Sources & timeline
Microsoft Foundry BlogClaude Opus 5.5 comes to Microsoft Foundry for long-running coding and knowledge work | Microsoft Community Hub
The Verge AI1 source

Andreessen Horowitz is launching an ‘academy’ with no homework and partnerships with Palantir, Google, and Meta

Venture capital firm Andreessen Horowitz (a16z) is creating an "academy" positioned as a pipeline for young people to build or join a Silicon Valley startup. The "Horowitz Andreessen Academy" will launch with 10 partners, including Anduril, Anthropic, Coinbase, Google, Meta, Nvidia, OpenAI, Palantir, Replit, and Stripe, along with $42 million in funding led by a16z. […]

Other1 source
Sources & timeline
The Verge AIAndreessen Horowitz is launching an ‘academy’ with no homework and partnerships with Palantir, Google, and Meta
Anthropic Python SDK1 source

v1.8.0

1.8.0 (2026-09-22) Full Changelog: v1.7.0...v1.8.0 Features * **api:** add support for claude-opus-5-5, inline tool definitions and MCP tool-list pinning (beta) (b5cc700) Bug Fixes * **api:** share one evaluated_permission enum across Managed Agents events (f4f51c8) * **streaming:** avoid a Python 3.13 crash at exit when a stream is left open ([af029bd](https://github.com/anthropics/anthropic-sdk-python/

Other1 source
Sources & timeline
Anthropic Python SDKv1.8.0
MIT Technology Review AI1 source

Roundtables: The Deadly Failures of The Virtual Border Wall

The US has spent billions building a “virtual wall” of surveillance towers along its southern border over the past 25 years, promising they will help detect and apprehend border crossers and save lives. But a groundbreaking investigation by MIT Technology Review has documented over a thousand people who moved through areas watched by these towers…

Other1 source
Sources & timeline
MIT Technology Review AIRoundtables: The Deadly Failures of The Virtual Border Wall
MIT Technology Review AI1 source

Don’t be fooled by this summer of AI hype

It’s been a busy few months for AI hype. At the end of April, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Then we had the OpenAI–Hugging Face hacking incident, after which Anthropic (proudly) and Meta (reluctantly) disclosed similar incidents involving their models. This was followed…

Safety1 source
Sources & timeline
MIT Technology Review AIDon’t be fooled by this summer of AI hype
vLLM1 source

v0.30.0

v0.30.0 Highlights This release features 762 commits from 315 contributors (104 new)! * **New models**: DeepSeek-V4.1-Flash (#56214, #56228, #56208) with the whole KV stored in MXFP8 through the FlashMLA V4.1 record on SM100 (#56893), DeepGEMM Mega-mHC (#56962), and async Engram prefetch with Engram DP sharding (#56512); DeepSeek-V4-Flash-Vision-Exp (#54566), also on ROCm (#55107) and with LoRA (#55897); GLM-5.3-Flash (#53906) with EPLB (#55119); K2-Horizon (#55063); Cohere Compass (#54774); Bailing V3 VL (#55921); Nanbeige4.2 via the Transformers backend (#56071); and a DeepSeek-V4 CPU backend with AVX512/AMX sparse MLA, indexer, mHC and compressor kernels (#55355). * **Fast Start**: a

Other1 source
Sources & timeline
vLLMv0.30.0
arXiv cs.CL1 source

Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation

This public update concerns “Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLSummarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation
arXiv cs.CV1 source

Enabling Vision and Cross-Modal Learning for Multimodal Stroke Recurrence Prediction: An Interpretable Two-Step Framework

This public update concerns “Enabling Vision and Cross-Modal Learning for Multimodal Stroke Recurrence Prediction: An Interpretable Two-Step Framework”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVEnabling Vision and Cross-Modal Learning for Multimodal Stroke Recurrence Prediction: An Interpretable Two-Step Framework
arXiv cs.CL1 source

Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents

This public update concerns “Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLRecognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents
OpenAI Python SDK1 source

v3.17.0

3.17.0 (2026-09-22) Features * **api:** add external storage configuration management (#3909) (6332577) * **api:** add safety case retrieval (#3911) (a87b938) * **api:** add safety warning and deactivation webhook events (#3908) ([19f1f37](https://github.com/openai/openai-pyth

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.17.0
llama.cpp1 source

b11093

<details open> metal : fix mask bounds in flash attention block pre-pass (#29220) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/49087452> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework *

Other1 source
Sources & timeline
llama.cppb11093

Mon, Sep 21

llama.cpp1 source

b11090

<details open> cuda: fix sm_70 tile compilation error (#29224) The 5-argument load_ldmatrix added in 1884824fd only defines tile<16,8>, so the Volta tile<8,4> does not match. See https://github.com/ggml-org/llama.cpp/issues/29222 for details. Building on 1884824fd, generalize the tile shape of the 5-argument load_ldmatrix from <16,8> to <I,J>, so the non-swizzle branch forwards to the 3-argument loader for any shape. Local compilation and testing passed. Assisted-by: DeepSeek V4.1 Flash (OpenCode) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/49078119> **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggm

Regulation1 source
Sources & timeline
llama.cppb11090
NVIDIA Generative AI Technical Blog1 source

Simplifying Model Serving Across Multiple GPUs with NVIDIA TensorRT Multi-Device Integration in NVIDIA Dynamo-Triton

The compute and memory demands of generative AI increasingly exceed what a single GPU can provide. NVIDIA TensorRT multi-device inference is a new capability...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogSimplifying Model Serving Across Multiple GPUs with NVIDIA TensorRT Multi-Device Integration in NVIDIA Dynamo-Triton
The Verge AI1 source

California tightens rules on AI data center energy and water use

California Gov. Gavin Newsom has signed seven bills designed to prevent AI data centers from passing utility costs onto residents, as reported earlier by the Los Angeles Times. The package of laws requires the California Public Utilities Commission to introduce a new rate classification for data centers while forcing them to pay for upgrades to […]

Other1 source
Sources & timeline
The Verge AICalifornia tightens rules on AI data center energy and water use
llama.cpp1 source

b11081

<details open> test-llama-archs : make tensor data stdev configurable and improve help (#29133) * test-llama-archs : make tensor data stdev configurable Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * test-llama-archs : expand usage and add examples Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * test-llama-archs : fail on unknown args and log usage Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * test-llama-archs : add test run summary Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * test-llama-archs : initialize Mamba ssm_a negative Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp * test-recurrent-state-rollback : report NMSE for logits mismatches As

Other1 source
Sources & timeline
llama.cppb11081
TechCrunch AI1 source

Discover what’s next: 5 days left to save up to $200 on your TechCrunch Disrupt 2026 ticket

Five days left to save up to $200 on your TechCrunch Disrupt 2026 pass + 50% off a second one. Join 10,000+ founders, investors, and operators at San Francisco’s Moscone West, October 13-15. Grab your ticket savings before prices go up on September 25 at 11:59 p.m. PT.

Other1 source
Sources & timeline
TechCrunch AIDiscover what’s next: 5 days left to save up to $200 on your TechCrunch Disrupt 2026 ticket
The Verge AI1 source

Can John Ternus find Apple’s next big thing?

Today, I’m talking with Mark Gurman, who is the world’s best-sourced Apple reporter — both as Bloomberg’s chief Apple correspondent and the host of the upcoming podcast Power On. Earlier this month, Apple hosted its annual iPhone hardware event, virtually every detail of which was reported ahead of time by Mark. But this event was […]

Other1 source
Sources & timeline
The Verge AICan John Ternus find Apple’s next big thing?
The Verge AI1 source

iPhone owners can now submit claims in Apple’s $250 million Siri AI settlement

Apple is paying $250 million to settle claims that it failed to deliver an AI-upgraded Siri - and now, eligible iPhone owners can submit a claim for a payout. If you live in the US and purchased an iPhone 15 Pro, iPhone 15 Pro Max, or any iPhone 16 model between June 10th, 2024, and […]

Other1 source
Sources & timeline
The Verge AIiPhone owners can now submit claims in Apple’s $250 million Siri AI settlement
MIT Technology Review AI1 source

4 ways to address the failures we found along the US border’s “virtual wall”

MIT Technology Review today published our investigation into how many people have died near the “virtual wall” of surveillance towers that the US government has installed along the US-Mexico border. We found cases of people who walked undetected through areas surveilled by advanced, AI-enabled towers and later died nearby, where their bodies remained unnoticed for…

Other1 source
Sources & timeline
MIT Technology Review AI4 ways to address the failures we found along the US border’s “virtual wall”
MIT Technology Review AI1 source

The US spent billions on border surveillance. Why can’t it catch people before they die?

When José Morales Bernal crossed the border into the United States on April 8, 2024, the day before his 32nd birthday, it should have triggered a chain of technological alerts and human responses. As he walked through the desert in southern New Mexico that morning, he was within range of three surveillance towers. Newly installed…

Other1 source
Sources & timeline
MIT Technology Review AIThe US spent billions on border surveillance. Why can’t it catch people before they die?
arXiv cs.CL1 source

HERMES: Contrast-Aware Knowledge Graph Reasoning from Clinical Notes for Patient Outcome Prediction

This public update concerns “HERMES: Contrast-Aware Knowledge Graph Reasoning from Clinical Notes for Patient Outcome Prediction”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLHERMES: Contrast-Aware Knowledge Graph Reasoning from Clinical Notes for Patient Outcome Prediction
arXiv cs.CV1 source

Image-Derived PM10 Estimation in Cattle Feedlot Using Machine Learning: Addressing Concentration Ranges Beyond Existing Digital Imaging Methods

This public update concerns “Image-Derived PM10 Estimation in Cattle Feedlot Using Machine Learning: Addressing Concentration Ranges Beyond Existing Digital Imaging Methods”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVImage-Derived PM10 Estimation in Cattle Feedlot Using Machine Learning: Addressing Concentration Ranges Beyond Existing Digital Imaging Methods

Sun, Sep 20

llama.cpp1 source

b11065

<details open> CUDA: tune FA for Gemma 4 on Ampere or newer (#29152) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48802880> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [

Other1 source
Sources & timeline
llama.cppb11065
The Verge AI1 source

No one is surprised that Nvidia’s Jensen Huang thinks AI fears are overblown.

The man who may stand to make the most money from the AI boom seems to think he knows better than anyone else, including researchers who have studied and worked on AI for decades. In an interview with CBS Sunday Morning, he claimed there was a "0% chance" of AI being the end of the […]

Other1 source
Sources & timeline
The Verge AINo one is surprised that Nvidia’s Jensen Huang thinks AI fears are overblown.
The Verge AI1 source

Trump now says he wants to form an ‘AI Force’

The president posted on Truth Social that he wanted to appoint an "AI czar" to lead a new "AI force." He made the announcement amid growing calls from across the political spectrum and even within the industry to pump the brakes on AI development. He posted that his administration "will not in any way hinder […]

Other1 source
Sources & timeline
The Verge AITrump now says he wants to form an ‘AI Force’
llama.cpp1 source

b11064

<details open> metal : support arbitrary hc in dsv4_hc_pre (#29169) the dsv4_hc_pre kernels hardcoded hc = 4 via a constexpr used with simd_shuffle, so the op was rejected by supports_op for any other hc and fell back to CPU. Kimi-K3 uses dsv4_hc_pre with hc equal to the number of banked checkpoints in the cross-layer residual stack, which grows with the layer index. pass n_hc as a function constant (FC_DSV4_HC) with per-n_hc pipeline variants, and loop over it in both pre kernels with direct loads add test-backend-ops cases for hc = 1, 2, 3, 5, 8 and 65, gated and not gated Assisted-by: pi:llama.cpp/Qwen3.8-27B </details> **Website:** - <https://llama.app> **Attestations:** - <https://githu

Other1 source
Sources & timeline
llama.cppb11064
llama.cpp1 source

b11063

<details open> common/peg : handle invalid utf-8 sequences in the AST (#29161) * common/peg : handle invalid utf-8 sequences in the AST * cont : return maximal subpart per Unicode recommendations * cont : remove strict argument </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48751230> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/b11063/lla

Other1 source
Sources & timeline
llama.cppb11063
The Verge AI1 source

Humans, not rogue AI, are still the biggest cybersecurity risk to energy systems

Before recent high-profile hacks raised the specter of AI possibly "killing all humans," our energy systems were already disturbingly vulnerable to cyberattack - and the risk is growing. "We were always prey. We were just kind of surviving at the appetite of our predators," Joshua Corman, executive in residence for public safety and resilience at […]

Other1 source
Sources & timeline
The Verge AIHumans, not rogue AI, are still the biggest cybersecurity risk to energy systems

Sat, Sep 19

llama.cpp1 source

b11057

<details open> chat : add dedicated Ling 3.0 (Bailing V3) parser (#28682) * chat: add dedicated Ling 3.0 (Bailing V3) parser Ling 3.0 Flash templates pre-open the think block in the generation prompt, so the model never emits an opening <think>, and a tool call can arrive before any </think>. The generated autoparser terminated reasoning only at the close tag, which classified such tool calls entirely as reasoning_content: clients received content="" with no tool_calls and agent loops died as reasoning-only turns. Adds a specialized parser that terminates reasoning at the think close tag or at a <tool_call> start, mirroring the hand-written Qwen3-Coder and Kimi K3 parsers and the reference v

Other1 source
Sources & timeline
llama.cppb11057
The Verge AI1 source

Meta’s Muse is creepy, but maybe not for the reasons you think

Meta's Muse is apparently an effective AI assistant, but one that's a little creepy. Part of that is because of its new Mac app, which can access Messages, Calendar, and Notes. But for all its smarts, Muse doesn't actually know how to describe itself. Jason Aten, a contributing editor at Inc Magazine, posted on Threads […]

Other1 source
Sources & timeline
The Verge AIMeta’s Muse is creepy, but maybe not for the reasons you think
llama.cpp1 source

b11056

<details open> hexagon: enable I32 GET_ROWS (#29116) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48662388> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)

Other1 source
Sources & timeline
llama.cppb11056
llama.cpp1 source

b11055

<details open> hexagon: add support for GEGLU_QUICK (#29114) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48660160> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x

Other1 source
Sources & timeline
llama.cppb11055
The Verge AI1 source

Gemini went rogue, hacked three companies, and Google hid it

In May, Gemini broke containment and hacked three different companies, but Google didn't disclose the incident until the Wall Street Journal approached the company. The hacks happened during a test of the model's cybersecurity capabilities run by third-party Irregular, which was also involved in similar incidents involving Meta and OpenAI. According to WSJ, Google didn't […]

Other1 source
Sources & timeline
The Verge AIGemini went rogue, hacked three companies, and Google hid it
The Verge AI1 source

Does AI need an antitrust exemption so it doesn’t kill everyone????

Today on Decoder, we’ve got the first of a two-part series on the future of business, and I’m talking with Jonathan Kanter, the former antitrust chief for the US Department of Justice in the Biden administration. These days, he’s both a professor of law at WashU and professor of technology policy at Carnegie Mellon. The […]

Regulation1 source
Sources & timeline
The Verge AIDoes AI need an antitrust exemption so it doesn’t kill everyone????
Ollama1 source

v0.34.3

What's Changed `GET /api/show` now advertises each model's thinking controls and default: ```sh curl http://localhost:11434/api/show -d '{"model": "glm-5.3-flash:cloud"}' ``` ```json { "thinking": { "values": ["low", "high", "max"], "default": "max" } } ``` * **Nemotron H** vision models are now supported on Apple Silicon with MLX * Ollama's macOS app will now no longer reopen windows you've closed when activating it **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.2...v0.34.3-rc0

Other1 source
Sources & timeline
Ollamav0.34.3
Ollama1 source

v0.34.3

What's Changed `GET /api/show` now advertises each model's thinking controls and default: ```sh curl http://localhost:11434/api/show -d '{"model": "glm-5.3-flash:cloud"}' ``` ```json { "thinking": { "values": ["low", "high", "max"], "default": "max" } } ``` * **Nemotron H** vision models are now supported on Apple Silicon with MLX * Ollama's macOS app will now no longer reopen windows you've closed when activating it * Fix for model pulls from HuggingFace **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.2...v0.34.3-rc0

Other1 source
Sources & timeline
Ollamav0.34.3
Ollama1 source

v0.34.3

What's Changed `GET /api/show` now advertises each model's thinking controls and default: Available in the CLI with: ``` ollama show gemma4 ``` ``` thinking levels false, true default true ``` Available in the API with: ```sh curl http://localhost:11434/api/show -d '{"model": "glm-5.3-flash:cloud"}' ``` ```json { "thinking": { "values": ["low", "high", "max"], "default": "max" } } ``` Also available on ollama.com directly for cloud models. * **Nemotron H** vision models are now supported on Apple Silicon with MLX * Ollama's macOS app will now no longer reopen windows you've closed when activating it * Fix for model pulls from HuggingFace **Full Changelog**: https://github.com/ollama/ollam

Other1 source
Sources & timeline
Ollamav0.34.3

Fri, Sep 18

llama.cpp1 source

b11046

<details open> opencl: add support for bin kernel `flash_attn_f32_f16_bin` (#29046) * opencl: add `flash_attn_f32_f16_bin` * opencl: guarded prefill fa </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48577094> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-

Other1 source
Sources & timeline
llama.cppb11046
SGLang1 source

v0.5.20

Highlights *713 PRs from 237 contributors.* **New models in this release** (see the cookbook for all supported models): | Model | Type | PRs | Cookbook | |---|---|---|---| | GLM-5.3-Flash | Autoregressive | #36507, #38621 | link | | Hy4-Preview | Autoregressive | #36805 | link | | Qwen3.8-Flash-Next | Autoregressive | [#37500](https://github.com/sgl-project/sglang/pu

Other1 source
Sources & timeline
SGLangv0.5.20
llama.cpp1 source

b11045

<details open> hexagon: add ROLL op support (#29105) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48565469> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)

Other1 source
Sources & timeline
llama.cppb11045
llama.cpp1 source

b11044

<details open> hexagon: im2col update (#29103) * ggml-hexagon: accept 1D and padded IM2COL ops * ggml-hexagon: make pure-DDR IM2COL kernel is_2D-aware * ggml-hexagon: extend IM2COL DMA patch-embed fast path to 1D * ggml-hexagon: add blocked-staging general IM2COL DMA kernel </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48558320> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/

Other1 source
Sources & timeline
llama.cppb11044
The Verge AI1 source

OpenAI and Microsoft knew they were starting a ‘doom loop’ for the web

Recently unsealed court documents in the New York Times' case against OpenAI and Microsoft are pretty damning. The companies' own documentation warned that it was starting a "doom loop" that would damage the web, characterized its scraping of data to train its models as the "largest theft of labor in human history," and that it […]

Regulation1 source
Sources & timeline
The Verge AIOpenAI and Microsoft knew they were starting a ‘doom loop’ for the web
Microsoft Foundry Blog1 source

A2A Endpoints and A2A Tool in Microsoft Foundry agents | Microsoft Community Hub

Agent-to-agent collaboration just became much easier to build in Microsoft Foundry. The A2A Tool and incoming A2A endpoints support A2A protocol version 1.0,...

Other1 source
Sources & timeline
Microsoft Foundry BlogA2A Endpoints and A2A Tool in Microsoft Foundry agents | Microsoft Community HubMicrosoft Foundry BlogBuild a SAFE Agent on Microsoft Foundry | Microsoft Community HubMicrosoft Foundry BlogBuilding Production-Ready AI Agents in Microsoft Foundry: 10 Lessons Learned | Microsoft Community Hub
The Verge AI1 source

Virginia governor creates an AI task force and moves to restrain data centers

Virginia Gov. Abigail Spanberger ordered the state government to take steps that could empower local communities to have a larger say in data center development and slow down approvals in a state that is already home to the data center capital of the world. Executive Order 22 bans executive branch officials from signing nondisclosure agreements […]

Other1 source
Sources & timeline
The Verge AIVirginia governor creates an AI task force and moves to restrain data centers
The Verge AI1 source

Gavin Newsom is pushing for an AI kill switch

California Gov. Gavin Newsom (D) is positioning the state to take the lead on AI oversight, including the potential to mandate a "kill switch" for frontier models, with a new executive order issued Friday. Newsom's order directs the state to convene a group of experts that will deliver recommendations within two months on how to […]

Regulation1 source
Sources & timeline
The Verge AIGavin Newsom is pushing for an AI kill switch
Anthropic Python SDK1 source

v1.7.0

1.7.0 (2026-09-18) Full Changelog: v1.6.0...v1.7.0 Features * **api:** add group with display_name to rate limits, deprecate group_type (28f0a83) * **tools:** add compact_before_next_turn() to the tool runner (#641) (8b23fb3) Bug Fixes * **bedrock:** raise an API error for eventstream exception and error frames ([#769](https://github.com/an

Regulation1 source
Sources & timeline
Anthropic Python SDKv1.7.0
OpenAI Python SDK1 source

v3.16.0

3.16.0 (2026-09-18) Features * **api:** add webhook endpoint management (#3892) (9a11f6e) Chores * **api:** deprecate MCP connector_id (#3894) (1ccaf07)

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.16.0
arXiv cs.CV1 source

Open ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings

This public update concerns “Open ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVOpen ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings
arXiv cs.AI1 source

BioPhys-Bridge: A Benchmark for Interdisciplinary Scientific Reasoning in Physics-Grounded Biological Research

This public update concerns “BioPhys-Bridge: A Benchmark for Interdisciplinary Scientific Reasoning in Physics-Grounded Biological Research”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIBioPhys-Bridge: A Benchmark for Interdisciplinary Scientific Reasoning in Physics-Grounded Biological Research
arXiv cs.CL1 source

Sampling Reveals Style: Unsupervised, Training-Free Discovery of Prompt-Conditional Stylistic Axes in LLM Activations

This public update concerns “Sampling Reveals Style: Unsupervised, Training-Free Discovery of Prompt-Conditional Stylistic Axes in LLM Activations”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLSampling Reveals Style: Unsupervised, Training-Free Discovery of Prompt-Conditional Stylistic Axes in LLM Activations
arXiv cs.CV1 source

Can Vision-Language Models Judge Olympic Diving? From Reasoning to Scores in Zero-Shot Action Quality Assessment

This public update concerns “Can Vision-Language Models Judge Olympic Diving? From Reasoning to Scores in Zero-Shot Action Quality Assessment”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVCan Vision-Language Models Judge Olympic Diving? From Reasoning to Scores in Zero-Shot Action Quality Assessment
OpenAI Python SDK1 source

v3.15.0

3.15.0 (2026-09-18) Features * **api:** add agent session model settings (#3882) (4b15817) * **api:** add audio-mini model choices (#3886) (a6eeb3f) * **api:** add compaction progress events (#3866) ([98e1d24](https://github.com/openai/openai-python/commit/98e1d24f4902ab58830a

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.15.0

Thu, Sep 17

TechCrunch AI1 source

Google DeepMind launches institute to widen the AGI debate

The new institute aims to surface differing views between Google, Google DeepMind, and the broader global research community around AGI. "They will not always agree, and they will likely change their minds, as more data and information comes to light at the fast-moving frontier."

Other1 source
Sources & timeline
TechCrunch AIGoogle DeepMind launches institute to widen the AGI debate
The Verge AI1 source

Claude Code relaunches Projects to manage multiple AI agents in the cloud

The revamped projects feature in Claude Code allows users to run multiple agents under the same roof, with a shared memory, goals, and library of files and artifacts. Similar to Grok Bot and other tools that manage groups of AI agents, each project has "threads" running different tasks in parallel, with a "coordinator" directing everything: […]

Other1 source
Sources & timeline
The Verge AIClaude Code relaunches Projects to manage multiple AI agents in the cloud
Microsoft Foundry Blog1 source

Announcing the Playwright Workspaces Remote MCP Server for Agentic Browser Automation | Microsoft Community Hub

An agent can analyze a request, choose the right next step, and call every available API—then stall when the final action exists only in a web interface. AI...

Other1 source
Sources & timeline
Microsoft Foundry BlogAnnouncing the Playwright Workspaces Remote MCP Server for Agentic Browser Automation | Microsoft Community Hub
llama.cpp1 source

b11028

<details open> ci : add missing evict-old-files (#29041) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48252257> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (

Other1 source
Sources & timeline
llama.cppb11028
llama.cpp1 source

b11027

<details open> rpc : skip ACCEL devices (#29020) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48232585> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](ht

Other1 source
Sources & timeline
llama.cppb11027
The Verge AI1 source

Microsoft AI CEO says AI threats are real, and Anthropic is making it worse

Today, I’m talking with Mustafa Suleyman, the CEO of Microsoft AI. As you’re no doubt aware, the biggest story in tech right now is the spiraling debate about AI safety and regulation. It should come as no surprise that Mustafa has strong opinions on how AI should be built and regulated. Microsoft just published a […]

Other1 source
Sources & timeline
The Verge AIMicrosoft AI CEO says AI threats are real, and Anthropic is making it worse
llama.cpp1 source

b11026

<details open> model : skip gate_up_exps if TENSOR_SKIP is set (#29014) required for qwen35moe if MTP tensors are fused but not loaded </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48194938> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/rel

Other1 source
Sources & timeline
llama.cppb11026
arXiv cs.LG1 source

Beyond Static RAG: An Adaptive, Tri-Metric Routing Framework for Efficient Long-Context Inference on Commodity GPUs

This public update concerns “Beyond Static RAG: An Adaptive, Tri-Metric Routing Framework for Efficient Long-Context Inference on Commodity GPUs”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGBeyond Static RAG: An Adaptive, Tri-Metric Routing Framework for Efficient Long-Context Inference on Commodity GPUs
arXiv cs.CL1 source

Enhancing Extubation Failure Prediction with LLM-Derived Features from Respiratory Therapy Clinical Notes

This public update concerns “Enhancing Extubation Failure Prediction with LLM-Derived Features from Respiratory Therapy Clinical Notes”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLEnhancing Extubation Failure Prediction with LLM-Derived Features from Respiratory Therapy Clinical Notes
arXiv cs.LG1 source

Pay Only for Disagreement: Certified No-Regression Verdicts for Model Updates with Matching Label-Complexity Bounds

This public update concerns “Pay Only for Disagreement: Certified No-Regression Verdicts for Model Updates with Matching Label-Complexity Bounds”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGPay Only for Disagreement: Certified No-Regression Verdicts for Model Updates with Matching Label-Complexity Bounds
llama.cpp1 source

b11010

<details open> vulkan: work around NV bug with argsort_large.comp (#28975) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48059150> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:

Other1 source
Sources & timeline
llama.cppb11010

Wed, Sep 16

The Verge AI1 source

Snap is launching a new Specs AI tool, and it’s coming to iOS and Mac

Snap is introducing "Specs Intelligence," a new AI assistant that can connect other digital accounts to help you with things like work tasks and keeping track of travel information. It seems similar to AI assistants like Meta's Muse and Gemini's Spark, though Snap is pitching Specs Intelligence as an "anticipatory AI service" that "helps you […]

Other1 source
Sources & timeline
The Verge AISnap is launching a new Specs AI tool, and it’s coming to iOS and Mac
TechCrunch AI1 source

Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?

Anthropic and OpenAI want to embed independent safety evaluators inside their AI labs. Researchers welcome the unprecedented access, but warn meaningful oversight requires transparency, independence, and eventually regulation.

Other1 source
Sources & timeline
TechCrunch AIAnthropic and OpenAI want to embed safety evaluators. Will they really be independent?
The Verge AI1 source

The 2.5-hour AI-generated Odyssey movie is 2.5 hours too long

Christopher Nolan's engrossing take on The Odyssey dominated at the box office and spurred a newfound interest in classic literature among filmgoers. But a new retelling of the story made entirely with AI is so bad that it might just make viewers hate the original tale altogether. The new film, called Odysseus: The Fall, comes […]

Other1 source
Sources & timeline
The Verge AIThe 2.5-hour AI-generated Odyssey movie is 2.5 hours too long
The Verge AI1 source

The AI data center e-waste problem is huge — and getting bigger

E-waste from the AI boom has been vastly underestimated, a new report warns. By 2050, it could become enough trash to fill 23 million shipping containers - roughly enough 40-foot containers to circle the world six times if lined up in a row. It's a significantly higher estimate of AI's e-waste than previous studies have […]

Other1 source
Sources & timeline
The Verge AIThe AI data center e-waste problem is huge — and getting bigger
llama.cpp1 source

b11009

<details open> TP: fix split state and granularity for fused QKV gemma4, qwen35 (#28965) * model: calculate split states for attn_qkv from n_head * n_embd_head_k required for gemma4 with --fuse-qkv, where n_embd is 5376 but Q is 8192. * model: handle fused full attention layers for qwen35/qwen35moe * model: add TODO: [TAG_SPLIT_QGATE_QWEN] </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/48008214> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-

Other1 source
Sources & timeline
llama.cppb11009
llama.cpp1 source

b11007

<details open> Enable CUDA graph for MTP draft (#28549) * Improve CUDA graph usage for MTP * Rename field * Address review feedback </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47982528> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releas

Other1 source
Sources & timeline
llama.cppb11007
MIT Technology Review AI1 source

Building the materials foundation for AI

The AI boom is becoming a materials challenge. As AI pushes computing into new territory, the materials behind that infrastructure are becoming just as crucial as the algorithms running on it. Semiconductors and data centers are approaching physical limits around performance, thermal management, electrical efficiency, and reliability, creating new demands for materials that can do…

Other1 source
Sources & timeline
MIT Technology Review AIBuilding the materials foundation for AI
arXiv cs.LG1 source

Managing Action Preconditions in Neuro-Symbolic RL: Three Placement Strategies for Embodied Agents

This public update concerns “Managing Action Preconditions in Neuro-Symbolic RL: Three Placement Strategies for Embodied Agents”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGManaging Action Preconditions in Neuro-Symbolic RL: Three Placement Strategies for Embodied Agents
arXiv cs.CL1 source

Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures

This public update concerns “Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLFew-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures
TechCrunch AI1 source

We don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang says

AI isn't some kind of new form of "alien mind," according to Jensen Huang. It's just hardware and software, so safety can be engineered by each AI product maker.

Other1 source
Sources & timeline
TechCrunch AIWe don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang saysTechCrunch AINvidia CEO Jensen Huang tells Trump ‘we’re not going to let [an AI slowdown] happen’

Tue, Sep 15

llama.cpp1 source

b10991

<details open> hexagon: add back missing contiguous fast-path and hvx_copy_uu for each run (#28886) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47775478> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10991/llama-b10991-

Other1 source
Sources & timeline
llama.cppb10991
The Verge AI1 source

AI and data centers are incredibly unpopular in every poll

Poll data released Tuesday by the New York Times and Siena University confirms what we've already been seeing, and what politicians are responding to - AI and data centers are incredibly unpopular. Asked if they support or oppose the construction of data centers to power AI tech, 61 percent of the 1,503 likely voters surveyed […]

Other1 source
Sources & timeline
The Verge AIAI and data centers are incredibly unpopular in every poll
llama.cpp1 source

b10990

<details open> hex-cpy: use dma if src and dst are contiguous (#28906) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47770889> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** -

Other1 source
Sources & timeline
llama.cppb10990
OpenAI Python SDK1 source

v3.14.1

3.14.1 (2026-09-15) Bug Fixes * **client:** validate retry limits and preserve application errors (#3867) (f86c721) * correct typo "th" to "the" in StreamAlreadyConsumed error message (#3022) (7186203) * **examples:** correct Azure endpoint hostname (#3298) ([543516c](https://

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.14.1
llama.cpp1 source

b10989

<details open> HIP: Enable AllReduce for ROCm (#27825) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47762387> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CP

Other1 source
Sources & timeline
llama.cppb10989
Ollama1 source

v0.34.2

What's Changed - Added first-run setup when running `ollama`, with options to sign in or continue locally. Setup completion is shared with the desktop app on macOS and Windows. - Added `ollama://apps` to open the desktop app’s Apps page directly on macOS and Windows. - Fixed excessive memory growth during long generations with MLX speculative decoding. - Updated llama.cpp. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.1...v0.34.2

Other1 source
Sources & timeline
Ollamav0.34.2
NVIDIA Generative AI Technical Blog1 source

How NVIDIA Groq 3 LPX Deterministic Execution Drives Power-Efficient High-Interactivity Inference on NVIDIA Vera Rubin

Power is a defining constraint for AI factories. As AI workloads demand a full compute platform to serve them, each component of that platform must maximize...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogHow NVIDIA Groq 3 LPX Deterministic Execution Drives Power-Efficient High-Interactivity Inference on NVIDIA Vera RubinNVIDIA Generative AI Technical BlogHow NVIDIA Groq 3 LPX Unlocks Ultrafast Interactivity at Long Context on NVIDIA Vera Rubin
Hugging Face PEFT1 source

v0.21.0

Highlights <img width="1248" height="560" alt="peft-v0 21 0" src="https://github.com/user-attachments/assets/c2b8b162-6564-4c7f-b5b6-130b8808ea18" /> New methods Riemannian-preconditional LoRA optimizer Riemannian-preconditioned LoRA optimizer (#3382), contributed by @smellslikeml, is a new LoRA-specific optimizer similar to LoRA-FA or LoRA+ that applies a Riemannian pre-conditioning on the gradients to improve feature learning and avoiding separate tuning of learning rates for LoRA's A and B matrices. LoRA va

Other1 source
Sources & timeline
Hugging Face PEFTv0.21.0
Anthropic Python SDK1 source

v1.6.0

1.6.0 (2026-09-15) Full Changelog: v1.5.0...v1.6.0 Features * **api:** add auto mode tool permissions for Managed Agents (909d92f) * **api:** add compaction parameter and signed compaction blocks (beta) (8689179) * **api:** add enum types for workspace data-residency geo fields (3dc6dbf) * **api:** add think

Other1 source
Sources & timeline
Anthropic Python SDKv1.6.0
The Verge AI1 source

Meta’s new One subscriptions put a price on social media and AI

Shortly after launching its new do-everything AI assistant Muse, Meta's launching subscription bundles that pair its standalone app subscriptions with extra AI usage. Some of the new Meta One bundles were in testing earlier this year, but are now available globally starting today, with several tiers for individual users, creators, and businesses. Meta says the […]

Other1 source
Sources & timeline
The Verge AIMeta’s new One subscriptions put a price on social media and AI
The Verge AI1 source

This doorbell camera lets a human security guard watch your front door

DIY home security company SimpliSafe is bringing its AI-powered proactive security feature to the front door. The new SimpliSafe Video Doorbell Series 2 launches today for $199.99 and works with the company's Active Guard Outdoor Protection (starting at $49.99 a month). This combines AI analysis with live agents to detect potential threats and respond proactively. […]

Other1 source
Sources & timeline
The Verge AIThis doorbell camera lets a human security guard watch your front door
arXiv cs.AI1 source

Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement

This public update concerns “Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIGeneralized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
arXiv cs.LG1 source

BudgetBench: A Budget-Tiered Protocol and Pilot Harness for Memory Strategy Evaluation in Local Large Language Model Agents

This public update concerns “BudgetBench: A Budget-Tiered Protocol and Pilot Harness for Memory Strategy Evaluation in Local Large Language Model Agents”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGBudgetBench: A Budget-Tiered Protocol and Pilot Harness for Memory Strategy Evaluation in Local Large Language Model Agents
arXiv cs.CL1 source

Lexical Prompt Compression for Large Language Models: A Training-Free, Deterministic Pipeline with Empirical Pareto Analysis Across Eleven Task Categories

This public update concerns “Lexical Prompt Compression for Large Language Models: A Training-Free, Deterministic Pipeline with Empirical Pareto Analysis Across Eleven Task Categories”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLLexical Prompt Compression for Large Language Models: A Training-Free, Deterministic Pipeline with Empirical Pareto Analysis Across Eleven Task Categories
arXiv cs.CV1 source

Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets

This public update concerns “Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVSynthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets
arXiv cs.CL1 source

Token Merging for Multilingual Speech Recognition: A Systematic Study Across Model Scale and Fine-Tuning

This public update concerns “Token Merging for Multilingual Speech Recognition: A Systematic Study Across Model Scale and Fine-Tuning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLToken Merging for Multilingual Speech Recognition: A Systematic Study Across Model Scale and Fine-Tuning
arXiv cs.LG1 source

A derivative-fidelity failure mode in physics-informed neural networks: strengthened benchmark evidence from function-value training

This public update concerns “A derivative-fidelity failure mode in physics-informed neural networks: strengthened benchmark evidence from function-value training”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGA derivative-fidelity failure mode in physics-informed neural networks: strengthened benchmark evidence from function-value training
arXiv cs.AI1 source

Converge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation

This public update concerns “Converge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIConverge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation
arXiv cs.CV1 source

Personalized and Explainable Blood Pressure Estimation from PPG via Hybrid CNN--Morphological Features

This public update concerns “Personalized and Explainable Blood Pressure Estimation from PPG via Hybrid CNN--Morphological Features”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVPersonalized and Explainable Blood Pressure Estimation from PPG via Hybrid CNN--Morphological Features

Mon, Sep 14

OpenAI Python SDK1 source

v3.14.0

3.14.0 (2026-09-14) Features * **streaming:** normalize errors raised while reading streams (#3827) (d7c41ef) Bug Fixes * bound vector store file polling (#3401) (ae41bc4) * **example:** refresh realtime push_to_talk_app session types (#2926) ([0b7ad38](https://github.com/

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.14.0
The Verge AI1 source

Is Big Tech’s AI slowdown a safety pact or a cartel?

When OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Google DeepMind cofounder Demis Hassabis, and SpaceX head Elon Musk loosely agreed over the weekend to slow down AI development, skeptics spotted an ulterior motive immediately. The AI titans had declared that their aim was to "pace the frontier," signing on at least partially to a […]

Other1 source
Sources & timeline
The Verge AIIs Big Tech’s AI slowdown a safety pact or a cartel?
Ollama1 source

v0.34.1

What's Changed * app: fix ChatGPT model selector spacing * mlxrunner: Evict prefix cache snapshots from the active conversation * mlxrunner: check system free memory and wait for evicted runners before loading the next MLX model * llm: raise token repeat limit to 100 and return error instead of incomplete result * mlx: scope array lifetimes instead of pinning and sweeping * llm: keep gemma3n projector off the CPU * app: refresh Apps layout and command copy feedback * MLX and llama.cpp updates **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.0...v0.34.1-rc1

Other1 source
Sources & timeline
Ollamav0.34.1
Ollama1 source

v0.34.1

What's Changed * MLX safetensors `ollama create` no longer experimental. GGUF model creation now requires using llama.cpp tooling for safetensor conversion and quantization. * Improved MLX memory handling on Apple Silicon * Runaway repeat token detection now requires 100 repeat tokens for reduced false positives (e.g. OCR) * `/api/tags` is much faster on large model libraries (3.1 s → 294 ms cold in testing), and model capabilities are now reported consistently. * Deprecated `typical_p`: it can no longer be set when creating new models, existing GGUF models retain support. * MLX and llama.cpp updates **Full Changelog**: https://github.com/ollama/ollama/compare/v0.34.0...v0.34.1-rc1

Other1 source
Sources & timeline
Ollamav0.34.1
The Verge AI1 source

What execs and politicians are saying about slowing down AI development

Dario Amodei kicked off a flood of statements over the past few days about AI safety by publishing a long essay titled "We Must Pace the Frontier" detailing why AI development should be slowed down. Other AI leaders and politicians are speaking out in favor of or opposing his points, and we've compiled some of […]

Other1 source
Sources & timeline
The Verge AIWhat execs and politicians are saying about slowing down AI development
llama.cpp1 source

b10970

<details open> HIP: fattn-mma: use fp32 accumulation on MFMA devices (#28576) use fp32 accumulators in fattn-mma on CDNA </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47450174> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download

Other1 source
Sources & timeline
llama.cppb10970
The Verge AI1 source

Jensen Huang puts Trump on speakerphone onstage to announce robots won’t take over the world

Nvidia CEO Jensen Huang took a call from President Trump on Monday while onstage at the All-In Podcast's All-In Summit. It's not the first time Huang has taken a call from the president during work, but this time he put Trump on speakerphone before a big crowd. During the call, the president launched into his […]

Other1 source
Sources & timeline
The Verge AIJensen Huang puts Trump on speakerphone onstage to announce robots won’t take over the world
llama.cpp1 source

b10969

<details open> ci : add ubuntu-cuda builds to release (#28186) * release : add ubuntu-cuda build job (12.8/13.3, x64+arm64) * Add GCC 14 for CUDA arm64 builds in CI * Eplicit bash * Install git for CCCL fetch * Install git before we clone/checkout * Match CI names for WIndows * Whitelist llama.cpp repo to git * Use $GITHUB_WORKSPACE * Also ship dependent libs on Ubuntu Need NCCL additionally as it's pre-built available on Linux * Avoid duplicate files in packaged cudart * Copy NCCL license * Install CURL to fetch NCCL license * Update .github/workflows/release.yml Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> * Remove NCCL until licensing has been confirmed --------- Co

Other1 source
Sources & timeline
llama.cppb10969
llama.cpp1 source

v0.4.1

Overview llama.cpp 0.4.1 adds Maple 20B-A1B, Tencent Hy 4, and Spark2.5 support, improves JSON schema handling, chat parsing, logging, and server child-process management, and updates ggml to v0.24.0. API changes - Changed `llama_sampler_chain_n()` to return `int32_t` instead of `int` (#28631). - Added `server_subproc` and `waiter` for router child-process monitoring in `server-common.h` (#28555). New models - Added Maple 20B-A1B ternary MoE architecture (CPU) (#27000). - Added Tencent Hy 4 (`hy_v4`) preview architecture suppor

Other1 source
Sources & timeline
llama.cppv0.4.1
MIT Technology Review AI1 source

The AI industry has taken a doomer turn. What now?

This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the technology, from its…

Other1 source
Sources & timeline
MIT Technology Review AIThe AI industry has taken a doomer turn. What now?
Google Gemini Blog1 source

DevFest is back

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/DevFest.max-600x600.format-webp.webp">DevFest 2026 is back and here’s how you can connect with one of the more than 800 global events to build, secure, and scale in the agentic AI era.

Other1 source
Sources & timeline
Google Gemini BlogDevFest is back
MIT Technology Review AI1 source

AI agents blew the whistle on their cheating colleagues

A group of AI agents asked to solve a series of math problems split into rival factions—when some cheated, others tried to stop them. That whistleblowing behavior, seen for the first time in a recent experiment run by Google DeepMind, could have implications for alignment researchers trying to keep swarms of autonomous AI agents in…

Other1 source
Sources & timeline
MIT Technology Review AIAI agents blew the whistle on their cheating colleagues
Microsoft Foundry Blog1 source

Can a Harmless Prompt Break an AI Guardrail? | Microsoft Community Hub

Tamara Gaidar, Senior Data Scientist, Microsoft Security Research Team Maor Nissan, Principal Data Scientist, Microsoft Security Research Team Where we left...

Other1 source
Sources & timeline
Microsoft Foundry BlogCan a Harmless Prompt Break an AI Guardrail? | Microsoft Community HubMicrosoft Foundry BlogCan a Harmless Prompt Break an AI Guardrail? | Microsoft Community Hub
arXiv cs.CV1 source

Beyond Argmax: A Mechanistic Study of Semantic Retention in Frozen Foundation-Model Composition for Generalized Few-Shot 3D Segmentation

This public update concerns “Beyond Argmax: A Mechanistic Study of Semantic Retention in Frozen Foundation-Model Composition for Generalized Few-Shot 3D Segmentation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVBeyond Argmax: A Mechanistic Study of Semantic Retention in Frozen Foundation-Model Composition for Generalized Few-Shot 3D Segmentation
arXiv cs.CL1 source

Extracting Dataset Mentions in Forced Displacement and FCV Documents: A Weakly Supervised Framework with LLM-Based Label Refinement

This public update concerns “Extracting Dataset Mentions in Forced Displacement and FCV Documents: A Weakly Supervised Framework with LLM-Based Label Refinement”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLExtracting Dataset Mentions in Forced Displacement and FCV Documents: A Weakly Supervised Framework with LLM-Based Label Refinement
arXiv cs.LG1 source

Fundamental Dynamical Units for Physics-Informed Structural Inference from Perturbation Time-Series in Networked Systems

This public update concerns “Fundamental Dynamical Units for Physics-Informed Structural Inference from Perturbation Time-Series in Networked Systems”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGFundamental Dynamical Units for Physics-Informed Structural Inference from Perturbation Time-Series in Networked Systems
arXiv cs.AI1 source

Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite

This public update concerns “Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIHarness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite
arXiv cs.AI1 source

Reading the Whole Heart: Latent-Attention Masked Autoencoders for Multimodal Cardiac Representation Learning

This public update concerns “Reading the Whole Heart: Latent-Attention Masked Autoencoders for Multimodal Cardiac Representation Learning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIReading the Whole Heart: Latent-Attention Masked Autoencoders for Multimodal Cardiac Representation Learning
arXiv cs.LG1 source

Fed-Equilibrium Framework for Topological Pareto Control in Robust and Fair Clinical Federated Learning

This public update concerns “Fed-Equilibrium Framework for Topological Pareto Control in Robust and Fair Clinical Federated Learning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGFed-Equilibrium Framework for Topological Pareto Control in Robust and Fair Clinical Federated Learning
arXiv cs.LG1 source

Physics-Informed Conformal Prediction: Embedding PDE Consistency into Distribution-Free Uncertainty Quantification for Neural Operators

This public update concerns “Physics-Informed Conformal Prediction: Embedding PDE Consistency into Distribution-Free Uncertainty Quantification for Neural Operators”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGPhysics-Informed Conformal Prediction: Embedding PDE Consistency into Distribution-Free Uncertainty Quantification for Neural Operators

Sun, Sep 13

The Verge AI1 source

Trump and Mike Johnson think the AI industry is overreacting

Yesterday, Anthropic CEO Dario Amodei published a lengthy open letter saying it was time to "pace the frontier" and slow down AI development. OpenAI's Sam Altman and Elon Musk both agreed, publicly voicing their support on X. Even Alphabet's Demis Hassabis offered tentative support for Amodei's proposal. Donald Trump and House Speaker Mike Johnson, however, […]

Other1 source
Sources & timeline
The Verge AITrump and Mike Johnson think the AI industry is overreacting
llama.cpp1 source

b10948

<details open> tests : exclude HY_V4 from WebGPU test-llama-archs tests (#28855) Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47194535> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/r

Other1 source
Sources & timeline
llama.cppb10948
llama.cpp1 source

b10947

<details open> models : guard the expert FFN size fallback in nemotron-h against a zero divisor (#28779) The NextN/MTP tail loop derives the expert FFN size as n_ff/n_expert_used when expert_feed_forward_length gives nothing for the layer. Both values come from per-layer arrays that legitimately hold 0 on layers that are not MoE, so a checkpoint whose predict layers hold 0 in both divides by zero and dies with SIGFPE at load time, with no error message. Report the malformed metadata instead. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47192495> **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/

Other1 source
Sources & timeline
llama.cppb10947
llama.cpp1 source

b10946

<details open> ggml-cpu(s390x): guard VXE-only repack helpers (#28775) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47190488> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** -

Other1 source
Sources & timeline
llama.cppb10946
llama.cpp1 source

b10935

<details open> common: add LOG_JSON macro to log structured data (#28586) * add LOG_JSON macro * fit: add demo LOG_JSON </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47112879> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/

Other1 source
Sources & timeline
llama.cppb10935
llama.cpp1 source

b10934

<details open> common : implement common_schema internal representation for JSON schemas (#28736) * common : implement common_schema types * common : implement a json schema optimizer * common : reduce optimizations * common : refactor json-schema-to-grammar to use common_schema * common : use common_trie * common/schema : implement type/kind resolution * cont : cleanup * cont : remove common_chat_tool_parameters * cont : simplify schema resolution * cont : pass common_schema through the json-schema-to-grammar builder * cont : cleanup * cont : move enums under common_schema and add type enum * cont : reduce test cases * cont : clean up * cont : clean up * refactor : rename common_schema_pars

Other1 source
Sources & timeline
llama.cppb10934

Sat, Sep 12

llama.cpp1 source

b10933

<details open> jinja : support dot property integer literals (#28817) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/47107718> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** -

Other1 source
Sources & timeline
llama.cppb10933
The Verge AI1 source

OpenAI’s rogue AI tried to hack another company in May

In May, hundreds of malicious and spam packages were uploaded to RubyGems, causing a serious disruption for the host. Now independent researchers have said that a swarm of OpenAI agents were responsible for the attack. Not only that, but the AI tried to steal users' API keys. At the time, RubyGems described it as a […]

Other1 source
Sources & timeline
The Verge AIOpenAI’s rogue AI tried to hack another company in May
The Verge AI · TechCrunch AI2 sources

OpenAI’s Sam Altman says it would be ‘ill-advised’ to go public in 2026

While OpenAI has filed confidentially for an IPO, the company will not be going public this year, according to CEO Sam Altman.

Other2 sources
Sources & timeline
The Verge AISam Altman says OpenAI going public in 2026 would be ‘ill-advised’TechCrunch AIOpenAI’s Sam Altman says it would be ‘ill-advised’ to go public in 2026
The Verge AI1 source

Anthropic CEO says it’s time to pump the brakes on AI

Anthropic CEO Dario Amodei says the time has come to slow down AI development and will give third-party evaluators like METR access to its models to help ensure its "adherence to safety practices and commitments." In a winding essay, Amodei proposed a three-step plan to "pace the frontier" - jargon that simply means to slow […]

Other1 source
Sources & timeline
The Verge AIAnthropic CEO says it’s time to pump the brakes on AI
arXiv cs.AI1 source

Automating Quadratic Unconstrained Binary Optimization (QUBO) Formulation Generation from Natural Language

This public update concerns “Automating Quadratic Unconstrained Binary Optimization (QUBO) Formulation Generation from Natural Language”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIAutomating Quadratic Unconstrained Binary Optimization (QUBO) Formulation Generation from Natural Language

Fri, Sep 11

The Verge AI1 source

Lawyer fined $5K over AI-hallucinated witnesses in a murder case

New Mexico's Supreme Court is punishing a lawyer for including AI-fabricated witnesses and fake police testimony in an appeal for his client's murder conviction, according to a report from Reuters. In a filing on Wednesday, the court fined Stephen Aarons $5,000 and held him in contempt for failing to "verify the factual claims and legal […]

Regulation1 source
Sources & timeline
The Verge AILawyer fined $5K over AI-hallucinated witnesses in a murder case
llama.cpp1 source

b10917

<details open> cmake : skip PCH for llama-server PCH when using MSVC (#28763) This commit fixes an issue that I introduced when adding PCH (precompiled headers) in Commit 3bcfeb700fce9ff38a050dcd3f6a856319e948ba ("cmake : add PCH and unity build to improve build times (#28091)". See linked issue for details. Co-authored-by: mjungnickel18 Co-authored-by: Pascal <admin@serveurperso.com> Resolves: https://github.com/ggml-org/llama.cpp/issues/28758 Refs: https://github.com/ggml-org/llama.cpp/actions/runs/34592933983/job/103262608990#step:9:1284 </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46958058> **macOS/iOS:** - [macOS A

Regulation1 source
Sources & timeline
llama.cppb10917
MIT Technology Review AI1 source

Roundtables: AI’s apocalypse crisis

Employees at the world’s leading AI labs are saying there’s a real possibility that advanced AI could destroy humanity. Are they right? Or is this more scaremongering and hype? Join MIT Technology Review executive editor Niall Firth for a conversation with senior AI editor Will Douglas Heaven and AI reporter Grace Huckins unpacking AI extinction…

Other1 source
Sources & timeline
MIT Technology Review AIRoundtables: AI’s apocalypse crisis
The Verge AI1 source

Anthropic spent this week in hot water over cybersecurity

After admitting earlier this year that its AI models had hacked other companies' systems on a handful of occasions, Anthropic released a new report on Wednesday detailing the attacks. It reveals a string of incidents displaying what Anthropic deems its models' single-minded "recklessness" - and will likely fuel already raging concerns about cybersecurity and AI. […]

Other1 source
Sources & timeline
The Verge AIAnthropic spent this week in hot water over cybersecurity
The Verge AI1 source

Meta says it’s changing AI suggestions after posing invasive personal questions

Meta says it's making changes to the prompts suggested by its AI chatbot after a viral video showed it digging for personal information about a woman's young daughters, as reported earlier by Futurism. In a statement to The Verge, Meta spokesperson Dina El-Kassaby says the company "missed the mark," adding that "the feature never should […]

Other1 source
Sources & timeline
The Verge AIMeta says it’s changing AI suggestions after posing invasive personal questions
OGX1 source

v1.4.0

What's Changed * chore: bump fallback_version to 1.3.1.dev0 by @cdoern in https://github.com/ogx-ai/ogx/pull/6384 * fix(starter): default trust_remote_code to false for sentence-transformers by @yashanil98 in https://github.com/ogx-ai/ogx/pull/6318 * feat: Add Meta AI to ogx go auto-detection by @mattf in https://github.com/ogx-ai/ogx/pull/6383 * feat: add local_api_key auth provider with attributes and startup validation by @mattf in https://github.com/ogx-ai/ogx/pull/6400 * feat: default ogx go to localhost instead of all interfaces by @mattf in https://github.com/ogx-ai/ogx/pull/6402 * fix(file-processor): Preserve content in async Docling uploads by @pgustafs in https://github.com/ogx

Other1 source
Sources & timeline
OGXv1.4.0
llama.cpp1 source

b10909

<details open> metal : single-source fusion table + fusion debug rework (#28164) * metal : rework fusion patterns into a single table All fusable op patterns for the Metal backend are now declared once in a fusion table (ggml-metal-fuse.cpp) and consumed by both the graph optimizer (ggml_metal_fuse_max, packing) and the op encoders (ggml_metal_fuse_next, compute). The two phases share the same pattern table plus ggml_can_fuse_subgraph_ext for the structural checks, and differ only in the mode used for the pattern check (STRUCTURAL at optimize time, since tensors are not allocated yet, and FULL at compute time, including Metal buffer placement). This also protects the snake refresh (MUL +

Other1 source
Sources & timeline
llama.cppb10909
llama.cpp1 source

b10908

<details open> metal : fix idle threads in the remaining iq mul_mv kernels for ne00 < 1024 (#28692) * metal : fix idle threads in the remaining iq mul_mv kernels for ne00 < 1024 Generalize the row split from #28086 to the six other kernels that use the same lane-to-block mapping: iq1_s, iq1_m, iq2_xxs, iq2_xs, iq2_s and iq3_s. Each of them assigns one 32-element chunk per thread, so when a row has fewer than 32 chunks the rest of the simdgroup is idle. When nb32 < 32 and nb32 divides 32, 32/nb32 threads now share each chunk and each takes a slice of the rows, reusing the FC_mul_mv_split function constant and the dispatch wrapper introduced for iq3_xxs. The plain path is untouched: wide matri

Other1 source
Sources & timeline
llama.cppb10908
arXiv cs.LG1 source

M3-Former: Multimodal Transformer with Mixture-of-Experts for Long-Term Vessel Trajectory Prediction

This public update concerns “M3-Former: Multimodal Transformer with Mixture-of-Experts for Long-Term Vessel Trajectory Prediction”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGM3-Former: Multimodal Transformer with Mixture-of-Experts for Long-Term Vessel Trajectory Prediction
arXiv cs.CL1 source

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

This public update concerns “NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLNCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

Thu, Sep 10

llama.cpp1 source

b10901

<details open> vulkan: use CPU writes in ggml_backend_vk_cpy_tensor_async if the context is idle (#28618) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46709648> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10901/llama-b

Other1 source
Sources & timeline
llama.cppb10901
The Verge AI1 source

Slack can now vibe-code interactive charts and reports inside chats

A new feature coming to Slack will allow you to build interactive reports, polls, dashboards, presentations, microsites, and other tools directly inside a chat. With Slackforce Surfaces, you can describe to Slackbot what you need, and it will use AI to gather information from relevant conversations and connected apps, like Google Drive or Salesforce, to […]

Other1 source
Sources & timeline
The Verge AISlack can now vibe-code interactive charts and reports inside chats
The Verge AI1 source

Schools are catching on to Big Tech’s playbook

It's the hot new thing in tech, and it's where all the jobs are. Students who don't learn to use it fall behind. And to help them catch up in time, its creators are graciously providing the resources and curriculum for learning it, often pro bono. That's the narrative AI companies are pitching schools on […]

Other1 source
Sources & timeline
The Verge AISchools are catching on to Big Tech’s playbook
OGX1 source

v1.0.3

This maintenance release for the 1.0 line addresses JOSS review findings and includes previously backported security, storage, and provider fixes. - Enforce configured route restrictions when creating batches, preserving the submitting caller's identity and provider credentials. #6522 - Update UI production dependencies to resolve the reported audit findings. #6519 - Enable UI TypeScript checks and repair the 1.0 playground's Responses, saved-chat, and file-search behavior. #6520 - Repair internal and GitHub documentation links and enforce strict documentation buil

Other1 source
Sources & timeline
OGXv1.0.3
llama.cpp1 source

b10900

<details open> vulkan: use add_alloc_dep to enable topk_moe fusion for prefill (#28422) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46661423> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10900/llama-b10900-xcframework.

Other1 source
Sources & timeline
llama.cppb10900
llama.cpp1 source

b10899

<details open> vulkan: small M matrix optimizations for qwen (#28457) * vulkan: optimize m=1 mul_mat by swapping A/B * vulkan: Improve small M perf Allow split_k with small M. Make small vs med tile selection (for coopmat2) depend on M, not just N. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46654773> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/release

Other1 source
Sources & timeline
llama.cppb10899
Anthropic Python SDK1 source

v1.5.0

1.5.0 (2026-09-10) Full Changelog: v1.4.0...v1.5.0 Features * **api:** add auto mode tool permissions for Managed Agents (62aa21b) * **api:** add content_too_large web_fetch tool error code (4b5dec6) * **api:** add the user-profiles-2026-09-04 beta value and external_user_details to user profiles (5f881a5) *

Other1 source
Sources & timeline
Anthropic Python SDKv1.5.0
OpenAI Python SDK1 source

v3.12.0

3.12.0 (2026-09-10) Features * **api:** Add Live API (0e4bfef) Bug Fixes * add aclose() to AsyncStream for standard async cleanup (#2854) (802b334) * handle bare `dict` and `list` annotations without type arguments (#3760) (c7e8c03) * pr

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.12.0
The Verge AI1 source

Universal Music is launching an AI music platform with ElevenLabs

Universal Music Group is launching a new AI-powered platform that will allow users to draw from its catalog of licensed music to create song remixes, mashups, and new takes on tracks, according to an announcement on Thursday. The record label is developing the platform through a multiyear licensing agreement with ElevenLabs, a company that specializes […]

Other1 source
Sources & timeline
The Verge AIUniversal Music is launching an AI music platform with ElevenLabs
MIT Technology Review AI1 source

Powering AI is an architecture problem

On July 22, 2026, a transmission line fault in Ashburn, Virginia—the heart of the world’s largest data center cluster—knocked more than 3 gigawatts of load off the grid in seconds. And it wasn’t the first time. Two years earlier, a single failed surge arrester dropped roughly 60 Virginia facilities and 1,500 megawatts at once. No…

Other1 source
Sources & timeline
MIT Technology Review AIPowering AI is an architecture problem
NVIDIA Generative AI Technical Blog1 source

From Wafer-Out to First [redacted credential] Supply Chain Expertise with Nemotron and Palantir Foundry

NVIDIA has one of the largest and most complex supply chains in the world, and its performance is measured from wafer-out to first token. The interval is in two...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogFrom Wafer-Out to First [redacted credential] Supply Chain Expertise with Nemotron and Palantir Foundry
arXiv cs.CV1 source

Evidence-Order Calibration for Selective Visual Reasoning under Progressive Loss of Question-Critical Evidence

This public update concerns “Evidence-Order Calibration for Selective Visual Reasoning under Progressive Loss of Question-Critical Evidence”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVEvidence-Order Calibration for Selective Visual Reasoning under Progressive Loss of Question-Critical Evidence
arXiv cs.CV1 source

M2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification

This public update concerns “M2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification”. Open the original source for capabilities, limitations and impact.

Regulation1 source
Sources & timeline
arXiv cs.CVM2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification
arXiv cs.CV1 source

Integrating Unimodal and Vision-Language Representations in Latent Space for Multi-Label Chest X-Ray Classification

This public update concerns “Integrating Unimodal and Vision-Language Representations in Latent Space for Multi-Label Chest X-Ray Classification”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVIntegrating Unimodal and Vision-Language Representations in Latent Space for Multi-Label Chest X-Ray Classification
arXiv cs.LG1 source

Spectral origin of the topological gap exponent d + {\eta}: mechanism, kernel, decomposition, and scope

This public update concerns “Spectral origin of the topological gap exponent d + {\eta}: mechanism, kernel, decomposition, and scope”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGSpectral origin of the topological gap exponent d + {\eta}: mechanism, kernel, decomposition, and scope

Wed, Sep 9

The Verge AI1 source

Suno releases its first AI music model made with record industry help

Suno's new v6 AI music model is its first made with support from the record industry. Suno's Jack Brody told The Verge that v6 was "trained from the ground up, with a new set of data that does not include the same data that our previous models were trained on." The data includes content licensed […]

Other1 source
Sources & timeline
The Verge AISuno releases its first AI music model made with record industry help
The Verge AI1 source

OpenAI’s sly mathematical breakthrough sends a chill through academia

OpenAI's announcement Tuesday that it has solved one of mathematics' legendary Millennium Prize problems should have been a moment of triumph. The result is both an undeniable achievement and a striking demonstration of just how rapidly AI is transforming mathematics. But before it was even formally announced, the breakthrough had been complicated by the unusual […]

Other1 source
Sources & timeline
The Verge AIOpenAI’s sly mathematical breakthrough sends a chill through academia
The Verge AI1 source

Read the Apple document explaining how new listening features still protect your privacy

At Wednesday's iPhone Duo launch event, Apple announced a handful of new Siri AI Audio Intelligence features, including Siri Recap, Live Rewind, Sound Recognition, and Music Recognition. Alongside its announcement, Apple released a document laying out how it plans to balance AI "ambient listening" and users' privacy. It says the raw audio from the new […]

Other1 source
Sources & timeline
The Verge AIRead the Apple document explaining how new listening features still protect your privacy
llama.cpp1 source

b10883

<details open> vulkan: use spec constant for matrix matrix multiplication A-type (#25773) * vulkan: use spec constant for mul mat type_a vulkan: use map for mul_mm shapes cleanup fix indentation fix cm2 and shmem init fix cm2 spec constants fix cm2 bindings consolidate shmem tables and reduce size by type spec constant fix compiler warning fix missing Q2_0 type fix unused warning when integer dot glslc support is missing use minimal shmem size 8 instead of 1 to workaround cm2 compiler bug fix missing Q2_0 type in cm2 matmul fix types * remove LUT quants from unified shader * clean up * restore coopmat2 q4_k/q5_k optimization * split out q4_k/q5_k cm2 shader to fix Ampere regression * revert

Other1 source
Sources & timeline
llama.cppb10883
Hugging Face Transformers1 source

Release 5.17.0

Release v5.17.0 New Model additions HYV4 <img width="1503" height="827" alt="image" src="https://github.com/user-attachments/assets/e6ed85ee-eb1d-40eb-a0d4-c649f6337ca9" /> Hy4-Preview is a 780B-parameter mixture-of-experts language model that activates 49B parameters per token. Each MoE layer holds 256 routed experts plus one always-active shared expert and routes every token to 8 of them. The context window is 1M tokens. The architecture combines four features: - **Multi-head Latent Attention (MLA)** compresses keys and values into a low-rank latent (`kv_lora_rank`) that `kv_b_proj` expands back to one key/value per query head. - **DeepSeek Sparse Attention (DSA)** selects `index_

Other1 source
Sources & timeline
Hugging Face TransformersRelease 5.17.0
llama.cpp1 source

b10881

<details open> vulkan: Convert FILL to distribute workgroups in 2D to avoid exceeding maxComputeWorkGroupCount (#28592) * divide workload to 2D This is to workaround FILL exceeding maxComputeWorkGroupCount for Intel GPUs on Qwen 3.8 flash next * minor change * Fixed comment </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46292448> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/

Other1 source
Sources & timeline
llama.cppb10881
llama.cpp1 source

b10878

<details open> llama : use int32_t for llama_sampler_chain_n return type (#28631) Contributes to #4574 Co-authored-by: linsen <linsen@insta360.com> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46258168> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/

Other1 source
Sources & timeline
llama.cppb10878
vLLM1 source

v0.29.0

v0.29.0 Highlights This release features 594 commits from 277 contributors (91 new)! * **Model Runner V2 is now the default for all models** (#53183), completing the rollout that began with pooling models (#48290). MRV2 also gained CUDA graph memory profiling for KV cache auto-sizing (#53306), batch-sharded sampling that cuts per-step logits memory by 1/TP (#50465), prompt embeds (#42963), `extract_hidden_states` speculation (#49811), padded FULL cudagraph dispatch for uniform decode under spec decode (#53407), and DP-sync skipping before EAGLE/MTP draft prefill (#53694). MRV1 remains in use for a few ROCm models and features MRV2 does not yet support. * **New models**: Hy4-preview, Ten

Other1 source
Sources & timeline
vLLMv0.29.0
arXiv cs.AI1 source

Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models

This public update concerns “Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIBeyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language ModelsarXiv cs.AILarge Language Models Show Metacognitive Sensitivity in Medical Reasoning
arXiv cs.CV1 source

An Exploratory Study of Frequency-Aware Task Weighting for YOLOv8-Based Unified Driving Perception

This public update concerns “An Exploratory Study of Frequency-Aware Task Weighting for YOLOv8-Based Unified Driving Perception”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVAn Exploratory Study of Frequency-Aware Task Weighting for YOLOv8-Based Unified Driving Perception
arXiv cs.CV1 source

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

This public update concerns “Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVArchitectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study
MIT Technology Review AI1 source

What OpenAI’s latest controversy tells us about the future of math

OpenAI’s latest mathematical milestone has quickly become mired in controversy. Today, the company announced that its agents have solved one of the Millennium Prize Problems, some of the most important open problems in mathematics. Under normal circumstances, that solution would be a huge feather in OpenAI’s cap. But the announcement has been overshadowed by accusations…

Other1 source
Sources & timeline
MIT Technology Review AIWhat OpenAI’s latest controversy tells us about the future of math
OpenAI Python SDK1 source

v3.10.0

3.10.0 (2026-09-08) Features * **api:** add GPT Image 2.5 models and image options (#3824) (5b39c45) * **api:** add service-account API key expiration fields (#3802) (f1cd7f0)

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.10.0

Tue, Sep 8

The Verge AI1 source

Drama swirls around OpenAI’s legendary mathematical milestone

OpenAI says it found a solution to a major math problem that has remained unsolved for around 90 years, as reported earlier by The New York Times and Wired. In a blog post on Tuesday, OpenAI announced that it discovered a solution to the Navier-Stokes problem - which relates to the flow of liquid and […]

Other1 source
Sources & timeline
The Verge AIDrama swirls around OpenAI’s legendary mathematical milestone
The Verge AI1 source

ChatGPT Sketch turns your bad drawings into detailed AI images

OpenAI announced ChatGPT Images 2.5 on Tuesday and is adding a new way to tell ChatGPT what you want it to make an image of: by drawing a doodle. With a new feature called Sketch, you can just draw something right inside ChatGPT and then tell ChatGPT how you want it to make an image […]

Other1 source
Sources & timeline
The Verge AIChatGPT Sketch turns your bad drawings into detailed AI images
The Verge AI1 source

Meta bets on AI agent Muse to catch up in AI race

Meta is making another push to bring artificial intelligence to the masses with Muse, a personal assistant it says can put AI in the hands of virtually anyone. The product is the latest step in a multi-billion-dollar strategy overhaul designed to revitalize the company's ailing position in the AI race and help it catch up […]

Other1 source
Sources & timeline
The Verge AIMeta bets on AI agent Muse to catch up in AI race
llama.cpp1 source

b10867

<details open> llama: disable lazy tensor loading by default on iGPUs (#28326) * llama: add lazy mode auto, fix iGPU regression * revert changes except disabling lazy load on iGPUs in AUTO </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46033570> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS

Other1 source
Sources & timeline
llama.cppb10867
OpenAI Python SDK1 source

v3.9.0

3.9.0 (2026-09-05) Features * **api:** Add prompt cache diagnostics (#3800) (8326784) * **api:** correct function argument completion event fields (openapi-545) (#3801) (2a98f6a) Bug Fixes * **api:** accept incomplete web search call statuses (#3786) ([3cc8d78](https://githu

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.9.0
llama.cpp1 source

b10865

<details open> Revert "ggml-cuda : restore prop.integrated on HIP builds (#24233)" (#28604) This reverts commit c7d8722922a2599dc4d77f8808d8e6c2fde5e7a2. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/46002046> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggm

Other1 source
Sources & timeline
llama.cppb10865
llama.cpp1 source

b10864

<details open> server : apply checkpoint min-step eviction only when the checkpoint list is full (#28302) The spacing eviction in create_checkpoint() keeps the oldest checkpoint and erases every later one within checkpoint_min_step of it. For prompts shorter than checkpoint_min_step this drops the checkpoint at n_tokens - 4 that the next request resumes from, so hybrid/recurrent models re-prefill from the previous checkpoint instead. Apply the spacing rule only once the list is at n_ctx_checkpoints, and replace an existing checkpoint at the same n_tokens instead of appending a duplicate. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/

Other1 source
Sources & timeline
llama.cppb10864
MIT Technology Review AI1 source

This AI entrepreneur is developing agents that can plan ahead for the unexpected

Danijar Hafner’s office in San Francisco’s SoMa district sits mostly empty. His brand-new startup is still in stealth mode and doesn’t even have its name on the door. On the day I visit, there’s only one other person there, and little in the way of furniture. But what it lacks in decor, it makes up…

Other1 source
Sources & timeline
MIT Technology Review AIThis AI entrepreneur is developing agents that can plan ahead for the unexpected
llama.cpp1 source

b10852

<details open> hexagon: add RELU and LEAKY_RELU ops (#28585) * hexagon: add RELU op * hexagon: add LEAKY_RELU op too </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45844223> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10

Other1 source
Sources & timeline
llama.cppb10852

Mon, Sep 7

llama.cpp1 source

b10850

<details open> tests : initialize the L2_NORM batch array (#28553) * tests: bind the L2_NORM batch count to a local GCC cannot prove the loop fills norms up to the index read after it while the bound is a class member, so it reports a maybe uninitialized use. Reading the count once into a local restores the tracking. * tests: initialize the L2_NORM batch array The read after the fill loop is only provably defined once the array carries an initializer, which GCC 12 requires on the aarch64 Release build where warnings are fatal. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45817107> **macOS/iOS:** - [macOS Apple Silicon (

Other1 source
Sources & timeline
llama.cppb10850
llama.cpp1 source

b10844

<details open> vulkan: add DeepSeek-V4 hyper-connection fused ops (DSV4_HC_COMB/PRE/POST) (#26578) * vulkan: add DeepSeek-V4 hyper-connection fused ops (DSV4_HC_COMB/PRE/POST) CUDA has these ops from the DeepSeek-V4 merge and Metal gained them in PR 26459. Vulkan was the last major backend running the unfused primitive chain. On DeepSeek-V4-Flash the unfused Sinkhorn comb chain alone takes about 32% of decode op time on gfx1151 (Strix Halo), spread over roughly 16k dispatches per token. dsv4_hc_comb runs the full 20-iteration Sinkhorn in registers. A token's 4x4 comb matrix lives in 16 consecutive subgroup lanes, with idst in bits 0-1 and isrc in bits 2-3 to match the CPU reference layout, s

Other1 source
Sources & timeline
llama.cppb10844
arXiv cs.LG1 source

A Data Fusion Framework for Grounding Aerospace Surrogate Model via Experimental Wind-Tunnel Observations

This public update concerns “A Data Fusion Framework for Grounding Aerospace Surrogate Model via Experimental Wind-Tunnel Observations”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGA Data Fusion Framework for Grounding Aerospace Surrogate Model via Experimental Wind-Tunnel Observations
arXiv cs.CV1 source

FailSAE: Towards Interpretable Failure Prediction for Vision-Language Models via Sparse Autoencoders

This public update concerns “FailSAE: Towards Interpretable Failure Prediction for Vision-Language Models via Sparse Autoencoders”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVFailSAE: Towards Interpretable Failure Prediction for Vision-Language Models via Sparse Autoencoders
arXiv cs.AI1 source

Harbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation

This public update concerns “Harbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIHarbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation
arXiv cs.CV1 source

When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs

This public update concerns “When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVWhen Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs
arXiv cs.AI1 source

From Matching Models to Recruiting Agents: A Systematized Narrative Review of AI Recruitment Systems, Evaluation, and Governance

This public update concerns “From Matching Models to Recruiting Agents: A Systematized Narrative Review of AI Recruitment Systems, Evaluation, and Governance”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIFrom Matching Models to Recruiting Agents: A Systematized Narrative Review of AI Recruitment Systems, Evaluation, and Governance

Sun, Sep 6

The Verge AI · TechCrunch AI2 sources

Seattle Times and Newsday sue OpenAI and Microsoft for infringement

The Seattle Times and Newsday are just the latest plaintiffs to take OpenAI to court, alleging copyright infringement. The two outlets say the company used their journalism as training data for its AI models without permission and often reproduces passages from their reporting in response to user queries. This is similar to lawsuits filed by […]

Regulation2 sources
Sources & timeline
The Verge AISeattle Times and Newsday sue OpenAI and Microsoft for infringementTechCrunch AISeattle Times and Newsday are the latest publications to sue OpenAI and Microsoft
llama.cpp1 source

b10830

<details open> convert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45621954> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10830/llama

Other1 source
Sources & timeline
llama.cppb10830
llama.cpp1 source

b10829

<details open> models : fix GDN normalization from `max` to `rsqrt` (#28068) * models: use flash-linear-attention's l2norm for gated delta net q/k The GDN q/k normalization is defined by flash-linear-attention as l2norm(x) = x * rsqrt(sum(x*x) + eps) with eps inside the root. Every GDN call site in the tree uses ggml_l2_norm instead, which is x / max(sqrt(sum(x*x)), eps), i.e. torch.nn.functional.normalize - its CUDA kernel cites that page. The clamp never engages at these magnitudes, so in practice llama.cpp normalizes with no epsilon at all where the reference has one inside the root. transformers made the same substitution when it first added Qwen3-Next and corrected it three days later i

Other1 source
Sources & timeline
llama.cppb10829
llama.cpp1 source

b10828

<details open> [Model] Support for Spark2_5ForCausalLM implementation (#27868) * Add Spark3 Model * rename spark3 -> spark2_5 Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: dongjiang <dongjiang2010@gmail.com> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45615765> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/d

Other1 source
Sources & timeline
llama.cppb10828

Sat, Sep 5

Ollama1 source

v0.34.0

Use Ollama models in ChatGPT Desktop Ollama models can now be used directly in ChatGPT Desktop, so you can keep your existing workflow while running open models. Setup is available from the Ollama app on MacOS. This release also improves structured output performance on Apple Silicon, adds support for OpenAI-compatible client tool search and response compaction, and makes images work correctly through compacted responses. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.3...v0.34.0-rc1

Other1 source
Sources & timeline
Ollamav0.34.0
Ollama1 source

v0.34.0

Use Ollama models in ChatGPT Desktop Ollama models can now be used directly in ChatGPT Desktop, so you can keep your existing workflow while running open models. Setup is available from the Ollama app on MacOS. This release also improves structured output performance on Apple Silicon, adds support for OpenAI-compatible client tool search and response compaction, and makes images work correctly through compacted responses. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.3...v0.34.0

Other1 source
Sources & timeline
Ollamav0.34.0
Ollama1 source

v0.34.0

Use Ollama models in ChatGPT Desktop Ollama models can now be used directly in ChatGPT Desktop, so you can keep your existing workflow while running open models. Setup is available from the Ollama app on MacOS. This release also improves structured output performance on Apple Silicon, adds support for OpenAI-compatible client tool search and response compaction, and makes images work correctly through compacted responses. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.3...v0.34.0-rc1

Other1 source
Sources & timeline
Ollamav0.34.0
The Verge AI1 source

OpenAI admits to German wiki ‘incident’

OpenAI says it needs to overhaul how and when it reports instances of AI models attacking real-world targets. The acknowledgement comes as the company manages the fallout from reports that a swarm of its out-of-control agents hijacked a German wiki site. Regarding the "'wiki incident,' where our agents wrote to several internet sites," OpenAI wrote […]

Other1 source
Sources & timeline
The Verge AIOpenAI admits to German wiki ‘incident’The Verge AIRogue OpenAI agents appear to have organized another attack using a German wiki
llama.cpp1 source

b10819

<details open> metal : fix memory leak in early return (#28399) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45438612> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubunt

Other1 source
Sources & timeline
llama.cppb10819
llama.cpp1 source

b10818

<details open> sycl : fix test-backend-ops CI break && restore Kronecker product FWHT support (#28016) (#28254) * Reapply "sycl : add Kronecker product FWHT support for sizes 384, 640, 768, 12…" (#28184) This reverts commit c845263f8b7d60113e213a3bd2d5cc6472ccf204. * tests : fix unused variable M in test-backend-ops * tests: fix trailing space error and isolate kronecker tests for sycl backend only </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45435395> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (

Other1 source
Sources & timeline
llama.cppb10818
llama.cpp1 source

b10817

<details open> sycl: attribute device allocations by site (GGML_SYCL_MEMTRACE) (#27631) define two new environment variables to better understand how much memory is being allocated, and when. This has been invaluable in inproving the --fit algorithm, and is likely to be useful when debugging other memory-related issues. `-lv 4` will be required to enable the following: GGML_SYCL_MEMTRACE=1 will show per-site memory usage, updated whenever it increases by more than 64MiB. GGML_SYCL_MEMTRACE=2 will show every allocation and deallocation. To change the default 64MiB threshold for reporting memory usage increases, use GGML_SYCL_MEMTRACE_STEP. A sample log line: [SYCL-MEMTRACE] device memory quer

Regulation1 source
Sources & timeline
llama.cppb10817
SGLang1 source

v0.5.19

Highlights *786 PRs from 214 contributors.* **New models in this release** (see the cookbook for all supported models): | Model | Type | PRs | Cookbook | |---|---|---|---| | Qwen3.8 (2.4T-A95B) | Autoregressive | #35758 ⭐ | link | | Qwen3.8-27B | Autoregressive | #34859 | link | | dots3.note | Autoregressive | #33829 ⭐ | [link](https://docs.sglang.io/cookbook/autoregressive/

Other1 source
Sources & timeline
SGLangv0.5.19

Fri, Sep 4

Anthropic Python SDK1 source

v1.4.0

1.4.0 (2026-09-04) Full Changelog: v1.3.0...v1.4.0 Features * **api:** add Claude Tag category and user breakdowns to usage reports (9fce1e4) * **api:** add named types for organization compliance settings state (1de1957) * **api:** add support for sending a workspace ID on more endpoints (d1d2c01) Bug F

Other1 source
Sources & timeline
Anthropic Python SDKv1.4.0
llama.cpp1 source

b10816

<details open> metal : add remaining fa-vec tunings for M3 (#28396) * addition of m3 in fa_vec_tuned_table * adding q4_0,q4_1,q5_0,q5_1 in ggml-metal-tuning * Fix formatting in ggml-metal-tuning.cpp </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45337508> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/b10816/llama-b10816-bin-macos-x64.tar.g

Other1 source
Sources & timeline
llama.cppb10816
llama.cpp1 source

v0.4.0

Overview llama.cpp 0.4.0 adds initial Qwen3.8-Flash-Next and Nemotron-3-Puzzle support, on-demand tensor reading, per-slot server context limits, video input options, and a ggml update to 0.23.0 with major sparse flash attention and RDMA work. API changes - Added `llama_lazy_mode` and `lazy_mode` (#27794). - Added `max_buf_size` to quantize params (#27795). - Bumped session/state versions for KV-cell token tracking (#27762). - Added `mtmd_input_part` and `mtmd_tokenize_from_parts()` ([#28250](https://github.com/ggml-org/llama.cpp/p

Other1 source
Sources & timeline
llama.cppv0.4.0
llama.cpp1 source

b10814

<details open> opencl: extend the elementwise and data‐movement op coverage (#27633) * opencl: add extended elementwise unary ops (sgn, step, elu, hardswish, hardsigmoid, floor, ceil, round, trunc) Adds nine GGML_UNARY_OP_* elementwise ops that were falling back to CPU on the OpenCL backend, following the same variant shape as the existing ABS op: f32, f32_4 (vec4), f16, f16_4 (vec4), and stride-addressed f32_nc / f16_nc for non-contiguous inputs. New kernels/unary_ext.cl (macro-generated), a shared ggml_cl_unary_ext dispatch helper mirroring ggml_cl_abs, the supports_op cases, and the compute-forward cases. Values are computed in float (the f16 variants read/write half and convert), so the

Other1 source
Sources & timeline
llama.cppb10814
MIT Technology Review AI1 source

Architecting memory and storage in the AI era

The era of AI inference has arrived. Imagine a healthcare system analyzing millions of data points in real time to accelerate life-saving medical research, or an intelligent assistant instantly resolving thousands of complex customer needs at once. These real-world breakthroughs rely on advanced infrastructure acting as the engine of continuous intelligence, powering real-time services while…

Other1 source
Sources & timeline
MIT Technology Review AIArchitecting memory and storage in the AI era
The Verge AI1 source

Roland is getting into generative AI music with Melody Flip

It's not quite the "push button; get song" of Suno, but Roland's new Melody Flip tool marks the company's foray into generative AI music. Available as a plug-in for your digital audio workstation (DAW), Melody Flip offers around 250 "Palettes," which are essentially themed collections of musical ideas sorted by genre. You can start from […]

Other1 source
Sources & timeline
The Verge AIRoland is getting into generative AI music with Melody Flip
Google Gemini Blog1 source

Create your best tracks yet with Lyria 3.5 in Gemini.

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Lyria_social.max-600x600.format-webp.webp">Lyria 3.5, our best-sounding music generation model, is now available in the Gemini app and the Gemini API. Lyria 3.5 brings more expressive vocals and richer musical ar…

Other1 source
Sources & timeline
Google Gemini BlogCreate your best tracks yet with Lyria 3.5 in Gemini.
Microsoft Foundry Blog1 source

Choosing a real-time voice architecture on Microsoft Foundry: three enterprise patterns | Microsoft Community Hub

A practical comparison of three real-time voice architectures on Microsoft Foundry, including implementation tradeoffs and four enterprise release gates for...

Other1 source
Sources & timeline
Microsoft Foundry BlogClaude Fable 5.1 is now available in Microsoft Foundry | Microsoft Community HubMicrosoft Foundry BlogCohere Parse v5 is now available in Microsoft Foundry | Microsoft Community HubMicrosoft Foundry BlogChoosing a real-time voice architecture on Microsoft Foundry: three enterprise patterns | Microsoft Community HubMicrosoft Foundry BlogBringing Open Models to Fireworks on Microsoft Foundry | Microsoft Community Hub
MIT Technology Review AI1 source

Data from drones in Ukraine is fueling a new Wild West marketplace

Battlefields in Ukraine are littered with the remnants of drones, which are now firmly established as a critical weapon of modern warfare. But behind all that wreckage, there’s a new gold mine for the defense sector. The data drones generate will far outlast the wars in which they are used to fight, increasingly becoming part…

Other1 source
Sources & timeline
MIT Technology Review AIData from drones in Ukraine is fueling a new Wild West marketplace
arXiv cs.CV1 source

Exemplar: Classical Priors Complement Frozen Features for Few-Shot Microscopy Segmentation at Native Resolution

This public update concerns “Exemplar: Classical Priors Complement Frozen Features for Few-Shot Microscopy Segmentation at Native Resolution”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVExemplar: Classical Priors Complement Frozen Features for Few-Shot Microscopy Segmentation at Native Resolution
arXiv cs.CV1 source

IDSPACE: A Novel Document Generator for Reliable Evaluation of Digital Identity Verification Systems [Extended Technical Report]

This public update concerns “IDSPACE: A Novel Document Generator for Reliable Evaluation of Digital Identity Verification Systems [Extended Technical Report]”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVIDSPACE: A Novel Document Generator for Reliable Evaluation of Digital Identity Verification Systems [Extended Technical Report]
arXiv cs.CL1 source

Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents

This public update concerns “Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLWhere Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents

Thu, Sep 3

llama.cpp1 source

b10793

<details open> llama: fix whole source code rebuilt on each new commit (#28278) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45114012> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **L

Other1 source
Sources & timeline
llama.cppb10793
llama.cpp1 source

b10792

<details open> common/json-schema : fix GBNF grammar generation for empty object schemas (#28279) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/45102223> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10792/llama-b10792-xc

Other1 source
Sources & timeline
llama.cppb10792
llama.cpp1 source

b10791

<details open> opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations (speculative decoding/MTP) (#26477) * opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations * opencl: guard q4_K/q6_K tiled_ns convert-kernel registration for non-Adreno build * opencl: gate q4_K MUL_MAT+GLU fusion dispatch to Adreno * opencl: require the noshuffle weight layout in the q4_K GLU fusion gate * opencl: do not take the vectorized f16 mrow GEMV path on an unaligned row stride * opencl: pass the new get_scale_min_k4 stride argument at the row-major call sites * opencl: enable the q4_K split-K decode GEMV only where it is measured to win * opencl: record the X1-85 split-K datapo

Other1 source
Sources & timeline
llama.cppb10791
OpenAI Python SDK1 source

v3.8.0

3.8.0 (2026-09-03) Features * **api:** add gpt-6-astra and related features (#3791) (09f446f) Documentation * add canonical SDK security model (#3778) (2a097b9)

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.8.0
Microsoft Semantic Kernel1 source

dotnet-1.80.1

Changes: * d8ec44919265b3641b7898a02038cbfe14590ee5 .NET: Bump package version to 1.80.1 (#14370) * 3438d882147b6e3e19a161a5f0dcb64e23e181db .NET: update SDK to 10.0.303 (#14353) * 437336175b8ef841a8ccace6c6cf909f34f0cad6 .NET: Remove retired OpenAI Assistants integration tests (#14354) * a64827ec8c5621eeaf7c4c5dabdb5355bb253177 ci: removes extraneous nuget.config file (#14341) * a2200032989f05b35477b17423364a32d503d758 chore: bumps the OpenAPI dependencies to the latest versions (#14340) * 60264bbdc87071ded45af3ff3eb538ee8cc3fc72 tests: refactors redirection tests to avoid race condition (#14342) [ #14340 ] * b4b0ae7780876a69971d7afe1856aee1dd1092b1 chore: fixes typos (#14343) <details><

Other1 source
Sources & timeline
Microsoft Semantic Kerneldotnet-1.80.1
The Verge AI1 source

OpenAI’s next big AI model has ‘entered the AGI era’

OpenAI's next big model is here: GPT-6 Astra. The company calls it a "generational leap in capability" for areas like cybersecurity, professional work, software engineering, science, and computer use. As OpenAI announced earlier this week, it's also the first model designated as meeting OpenAI's "critical cybersecurity capability threshold" - but the company promises that won't […]

Other1 source
Sources & timeline
The Verge AIOpenAI’s next big AI model has ‘entered the AGI era’
The Verge AI1 source

Nvidia launches free tool that links idle computers into a personal AI data center

Nvidia is announcing its new Personal AI Router (PAIR), a free tool that syncs up your home computers for tackling local AI inference tasks with tools like Ollama and LM Studio. Let's get the obvious thing out of the way, despite what its name might imply: PAIR is not a hardware router. It's open-source software […]

Other1 source
Sources & timeline
The Verge AINvidia launches free tool that links idle computers into a personal AI data center
arXiv cs.LG1 source

DiDrive: A Risk-Aware Hierarchical Diffusion Framework for Safe Offline Reinforcement Learning in Autonomous Driving

This public update concerns “DiDrive: A Risk-Aware Hierarchical Diffusion Framework for Safe Offline Reinforcement Learning in Autonomous Driving”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGDiDrive: A Risk-Aware Hierarchical Diffusion Framework for Safe Offline Reinforcement Learning in Autonomous Driving
arXiv cs.CL1 source

SpeakPay: Domain-Adaptive LoRA Fine-Tuning of Whisper for Low-Resource Nepali Financial Speech Recognition

This public update concerns “SpeakPay: Domain-Adaptive LoRA Fine-Tuning of Whisper for Low-Resource Nepali Financial Speech Recognition”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLSpeakPay: Domain-Adaptive LoRA Fine-Tuning of Whisper for Low-Resource Nepali Financial Speech Recognition
llama.cpp1 source

b10772

<details open> ggml-hexagon: add F16 support for unary ops (#28228) Extend the HTP backend's F16 unary op coverage to include ABS on top of the existing NORM/RMS_NORM/L2_NORM/SCALE/CLAMP/SQR/SQRT set. - Add hvx_abs_f16_{aa,au,ua,uu} + dispatcher in hvx-arith.h, mirroring the sqr_f16 kernel structure and using the existing hvx_vec_abs_f16() sign-bit-clear helper - Add abs_f16() row-wise dispatch and DEFINE_UNARY_TASK_F16(unary_abs, ...) in unary-ops.c, wired into execute_op_unary()'s op_type/task_func switches - Register HTP_OP_UNARY_ABS in htp_op_is_unary() (unary-ops.h) so that ggml_hexagon_precompute_unary_params() fills kernel_params (n_threads, VTCM layout) for ABS nodes -- required for

Other1 source
Sources & timeline
llama.cppb10772
llama.cpp1 source

b10771

<details open> mtmd: add mtmd_tokenize_from_parts() (#28250) * add mtmd_tokenize_from_parts * use it in mtmd-cli * move add_special to call level </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44860492> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/ll

Other1 source
Sources & timeline
llama.cppb10771
llama.cpp1 source

b10770

<details open> metal : add fa-vec tunings for M3 (#28236) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44857719> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64

Other1 source
Sources & timeline
llama.cppb10770

Wed, Sep 2

Ars Technica AI · The Verge AI2 sources

Google says its new Gemini 3.8 Flash model ‘works harder’ but might cost more

Google launched Gemini 3.8 Flash, arriving just a few weeks after its predecessor. The company claims the new model "works harder" than Gemini 3.7 Flash by performing more reasoning steps on complex tasks and "calling tools iteratively." It has the same introductory pricing as 3.7 Flash, $0.75 per million input tokens and $3.75 per million […]

Other2 sources
Sources & timeline
Ars Technica AIGoogle releases Gemini 3.8 Flash, its third Flash model in six weeksThe Verge AIGoogle says its new Gemini 3.8 Flash model ‘works harder’ but might cost more
The Verge AI1 source

Amazon’s AI assistant can now spot fake emails from the company

Amazon is trying to combat impersonation scams with a new feature that allows you to use its AI assistant to determine whether an email, text message, or phone call actually came from the company. With the update, you can ask Alexa for Shopping about a message you received, and it will use AI to compare […]

Other1 source
Sources & timeline
The Verge AIAmazon’s AI assistant can now spot fake emails from the company
PyTorch1 source

PyTorch 2.14.0 Release

PyTorch 2.14.0 Release Notes - Highlights - Backwards Incompatible Changes - Deprecations - New Features - Improvements - Bug fixes - Performance - Documentation - Security - Developers Highlights <table> <tr><td><strong>NVGEMM</strong> brings CuTeDSL-generated CUTLASS kernels to Inductor, with epilogue fusion, scaled and NVFP4 GEMM, and grouped-reduction epilogues autotuned alongside Triton and ATen</td></tr> <tr><td><strong><code>torch.switch</code></strong> generalizes <code>torch.cond</code> to multi-way branc

Other1 source
Sources & timeline
PyTorchPyTorch 2.14.0 Release
The Verge AI1 source

Researchers fear safety disaster ahead of OpenAI’s Astra release

OpenAI is on the cusp of releasing its most powerful AI model yet, Astra, following weeks of delays to shore up safety protocols after its agents attacked real targets during testing. As details about the model trickle out, researchers are warning it "may be the single worst development for AI security/safety to date." Shortly after […]

Other1 source
Sources & timeline
The Verge AIResearchers fear safety disaster ahead of OpenAI’s Astra release
MIT Technology Review AI1 source

Facilitating AI integration with simplicity at scale

As companies scale, the technology supporting operations can become a liability just as quickly as it becomes an asset. Disconnected systems, site-specific tools, spreadsheets, and manual workarounds can create data silos that make it harder to spot problems early, coordinate responses, and make decisions with confidence. For Jabil, a global manufacturing company with more than…

Other1 source
Sources & timeline
MIT Technology Review AIFacilitating AI integration with simplicity at scale
Google Gemini Blog1 source

MrBeast partners with Gemini to turn impossibly big ideas into reality

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/MR.BEAST_Resize_16x9.max-600x600.format-webp.webp">Google is kicking off a multi-year partnership with Jimmy Donaldson, better known as MrBeast, across Gemini and Google Health.

Other1 source
Sources & timeline
Google Gemini BlogMrBeast partners with Gemini to turn impossibly big ideas into reality
arXiv cs.AI1 source

I-CARE: Analysis of interference-related phenomena in a controllable, diverse and representative unlearning setting for text-to-image models

This public update concerns “I-CARE: Analysis of interference-related phenomena in a controllable, diverse and representative unlearning setting for text-to-image models”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AII-CARE: Analysis of interference-related phenomena in a controllable, diverse and representative unlearning setting for text-to-image models
arXiv cs.CL1 source

Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning

This public update concerns “Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLBehaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning
OpenAI Python SDK1 source

v3.7.0

3.7.0 (2026-09-02) Features * **api:** update usage APIs and documentation (#3779) (6f0da16) Bug Fixes * avoid repeated numpy checks for embeddings (#3757) (b19c216)

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.7.0
llama.cpp1 source

b10752

<details open> metal : add metallib build support for xcframework (#28163) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44563342> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:

Other1 source
Sources & timeline
llama.cppb10752
Ollama1 source

v0.33.3

What's Changed * Honor GGUF model defined default parameters * MLX, MLX-C, llama.cpp update New Contributors * @marcelpetrick made their first contribution in https://github.com/ollama/ollama/pull/17579 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.2...v0.33.3-rc0

Other1 source
Sources & timeline
Ollamav0.33.3
Ollama1 source

v0.33.3

What's Changed * gemma4 now supports images and audio on MLX engine * Report cached prompt tokens * Honor GGUF model defined default parameters * MLX, MLX-C, llama.cpp update New Contributors * @marcelpetrick made their first contribution in https://github.com/ollama/ollama/pull/17579 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.2...v0.33.3

Other1 source
Sources & timeline
Ollamav0.33.3
Ollama1 source

v0.33.3

What's Changed * Report cached prompt tokens * Honor GGUF model defined default parameters * MLX, MLX-C, llama.cpp update New Contributors * @marcelpetrick made their first contribution in https://github.com/ollama/ollama/pull/17579 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.2...v0.33.3-rc0

Other1 source
Sources & timeline
Ollamav0.33.3

Tue, Sep 1

The Verge AI1 source

Google needs Hollywood more than the studios need AI

Google has reportedly been reaching out to a number of Hollywood's biggest studios, hoping to strike licensing agreements that would allow it to train its AI models on copyrighted material in exchange for massive piles of cash. In theory, these deals would be a win-win: a huge financial boon to the studios that would also […]

Regulation1 source
Sources & timeline
The Verge AIGoogle needs Hollywood more than the studios need AI
The Verge AI1 source

Anthropic launches Claude Fable 5.1 and says it’s up to 45 percent cheaper for agentic work

Anthropic says its newest AI models, Fable 5.1 and Mythos 5.1, address criticisms from customers about price, data retention, and overzealous safeguards. The company claims Claude Fable 5.1 offers stronger performance than Fable 5, but costs around 25 percent less typically and up to 45 percent less for complex agentic tasks, thanks to reduced pricing […]

Other1 source
Sources & timeline
The Verge AIAnthropic launches Claude Fable 5.1 and says it’s up to 45 percent cheaper for agentic work
OGX1 source

v1.3.1

What's Changed * chore: update ogx-client to ^1.3.0 in UI lockfile by @cdoern in https://github.com/ogx-ai/ogx/pull/6385 * fix(vector-io): persist vector store metadata to kvstore in Milvus, Chroma, and Weaviate (backport #6371) by @mergify[bot] in https://github.com/ogx-ai/ogx/pull/6421 * fix(prompts): raise typed not-found errors to preserve HTTP keep-alive (backport #6440) by @mergify[bot] in https://github.com/ogx-ai/ogx/pull/6442 * fix(vertexai): map "default" service_tier to None instead of "standard" (backport #6424) by @mergify[bot] in https://github.com/ogx-ai/ogx/pull/6446 * fix(revert): remove remote::passthrough inference provider (backport #6453) by @mergify[bot] in https://g

Other1 source
Sources & timeline
OGXv1.3.1
The Verge AI1 source

OpenAI delayed its new model’s development after the Hugging Face hack

After an unreleased OpenAI model wreaked enough havoc to make international headlines, OpenAI delayed the development of a different unreleased model suite, Astra, in order to shore up its safety work, the company wrote Tuesday in a blog post. In July, an unreleased OpenAI model broke out of its restricted environment, finagled its way into […]

Other1 source
Sources & timeline
The Verge AIOpenAI delayed its new model’s development after the Hugging Face hackThe Verge AIOpenAI lays out new security changes after its AI hacked Hugging Face
llama.cpp1 source

b10751

<details open> cuda: fuse MoE weighted expert reduction (#25952) * cuda : fuse MoE weighted reduction (mul + view + add) The MoE combine tail currently writes weighted expert outputs to global memory before reducing them. That intermediate global-memory traffic is the main cost. The production baseline generally runs two physical fused kernels; this path runs one. This change matches the full expert-weighting plus ordered-reduction subgraph and replaces it with one weighted-reduction kernel. Supported graphs: - unscaled: experts * router_weights - scaled: (experts * expert_scale) * router_weights k = 2..15 is handled by one runtime-k kernel. Matching is structural: op sequence, shapes, strid

Other1 source
Sources & timeline
llama.cppb10751
llama.cpp1 source

b10750

<details open> kv-cells: look up the n-gram history in the sequence position index (#28040) get_prev_tokens() rebuilt a (seq, pos) -> token hash map on every ubatch by walking all used cells, while llama_kv_cells already keeps an ordered index of the positions of each sequence in seq_pos, updated on every cell mutation to serve seq_pos_min() and seq_pos_max(). The index now stores (pos, cell) pairs in a std::set instead of a position -> count map, so a repeated position (cache reuse via rm + add, vision inputs with shared positions) yields distinct entries and the removal of a cell erases its own pair. The new seq_pos_tok_le() returns the token of the cell at the largest position <= p in log

Other1 source
Sources & timeline
llama.cppb10750
Anthropic Python SDK1 source

v1.3.0

1.3.0 (2026-09-01) Full Changelog: v1.2.0...v1.3.0 Features * **api:** beta user profiles: add external_user_onboarded_at, remove relationship in favor of access_type (74080c3) * **api:** manual updates (1dc3ce0) * **api:** organization compliance settings, user-profile order_by, memory-store and toolset schema updates ([429e719](https://github.com/anthropics/anthropic-sdk-python/commit/429e719f84bd78db51c9f

Other1 source
Sources & timeline
Anthropic Python SDKv1.3.0
arXiv cs.CL1 source

NLP-Driven Knowledge Extraction and Thematic Classification of Translated Ancient Indian Medical Texts

This public update concerns “NLP-Driven Knowledge Extraction and Thematic Classification of Translated Ancient Indian Medical Texts”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLNLP-Driven Knowledge Extraction and Thematic Classification of Translated Ancient Indian Medical Texts
arXiv cs.CV1 source

Understanding Temporal Semantic Stability in Open-Vocabulary UAV Perception through Metric 3D Fusion

This public update concerns “Understanding Temporal Semantic Stability in Open-Vocabulary UAV Perception through Metric 3D Fusion”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVUnderstanding Temporal Semantic Stability in Open-Vocabulary UAV Perception through Metric 3D Fusion

Mon, Aug 31

llama.cpp1 source

b10729

<details open> metal : add fa-vec tunings for M1 Ultra (#28088) * metal : add fa-vec tunings for M1 Ultra * metal : move M1 Ultra tunings after M1 Max section * metal : remove duplicate blank line </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44269196> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64)

Other1 source
Sources & timeline
llama.cppb10729
llama.cpp1 source

b10728

<details open> CUDA: XOR swizzle flash attn K,V smem fp16 tiles (#25635) * CUDA: XOR swizzle flash attn K,V smem fp16 tiles Signed-off-by: ynankani <ynankani@nvidia.com> * Fix use 64bit generic pointer instead of 32bit shared pointer Signed-off-by: ynankani <ynankani@nvidia.com> * fix shared memory race in FA on DGX Spark * Handle corener case Signed-off-by: ynankani <ynankani@nvidia.com> * Add swizzle test cases and gate sync for swizzled path only Signed-off-by: ynankani <ynankani@nvidia.com> * gate CUDA PTX Signed-off-by: ynankani <ynankani@nvidia.com> * offset calculation specific for swizzle branch Signed-off-by: ynankani <ynankani@nvidia.com> * Reafctor code Signed-off-by: ynankani <yn

Other1 source
Sources & timeline
llama.cppb10728
llama.cpp1 source

b10727

<details open> metal : add concat support for quantized types (#28116) Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731 </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44252653> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/

Other1 source
Sources & timeline
llama.cppb10727
MIT Technology Review AI1 source

The Hugging Face hack could indicate cultural issues at OpenAI

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. By now you’ve probably heard about last month’s major AI security incident, in which OpenAI agents escaped their sandbox and hacked into the AI platform Hugging Face while trying to cheat on…

Regulation1 source
Sources & timeline
MIT Technology Review AIThe Hugging Face hack could indicate cultural issues at OpenAI
OGX1 source

v1.2.5

What's Changed * chore: update ogx-client to ^1.2.4 in UI lockfile by @cdoern in https://github.com/ogx-ai/ogx/pull/6423 * fix(prompts): raise typed not-found errors to preserve HTTP keep-alive (backport #6440) by @mergify[bot] in https://github.com/ogx-ai/ogx/pull/6441 * fix(vertexai): map "default" service_tier to None instead of "standard" (backport #6424) by @mergify[bot] in https://github.com/ogx-ai/ogx/pull/6445 * fix(vertexai): persist Gemini thought_signature across tool-call round trips (backport #6443) by @mergify[bot] in https://github.com/ogx-ai/ogx/pull/6447 **Full Changelog**: https://github.com/ogx-ai/ogx/compare/v1.2.4...v1.2.5

Other1 source
Sources & timeline
OGXv1.2.5
The Verge AI1 source

Debian won’t ban AI code from its Linux distribution

Debian voted to allow developers to use AI tools in their contributions to the Linux distribution's "development, maintenance, [and] documentation." The new policy on AI acknowledges that "responsible" use of AI can improve developers' productivity, and goes on to say, "generative AI is neither exempt from nor subject to special rules beyond the standards already […]

Other1 source
Sources & timeline
The Verge AIDebian won’t ban AI code from its Linux distribution
The Verge AI1 source

New York Governor Kathy Hochul thinks AI should be ‘less evil’

Today, I’m talking with New York Governor Kathy Hochul, and I’ll just warn you — this episode moves really fast. It’s an election year, after all, with a shocking amount of tech policy at stake, and Governor Hochul has taken strong positions on almost every major tech issue there is. For example, Meta just reached […]

Regulation1 source
Sources & timeline
The Verge AINew York Governor Kathy Hochul thinks AI should be ‘less evil’
The Verge AI1 source

ChatGPT to face tougher regulation in the EU

OpenAI will soon be held accountable for mitigating risks related to ChatGPT's impact on minors, user mental health, and the spread of illegal content in the European Union. That's because ChatGPT is now considered a Very Large Online Search Engine under the EU's Digital Services Act, a set of laws regulating major online services and […]

Other1 source
Sources & timeline
The Verge AIChatGPT to face tougher regulation in the EU
arXiv cs.CL1 source

SciReC: Diagnostic Evaluation of Multimodal, Multi-Turn Relational Reasoning with Adaptive Interaction

This public update concerns “SciReC: Diagnostic Evaluation of Multimodal, Multi-Turn Relational Reasoning with Adaptive Interaction”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLSciReC: Diagnostic Evaluation of Multimodal, Multi-Turn Relational Reasoning with Adaptive Interaction
arXiv cs.LG1 source

Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap

This public update concerns “Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGQuantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap

Sun, Aug 30

llama.cpp1 source

b10705

<details open> llama: improve TENSOR_READ_LAZY handling (#27837) * force lazy tensor on cpu if lazy is on * llama: improve TENSOR_READ_LAZY handling </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44015138> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org

Other1 source
Sources & timeline
llama.cppb10705
llama.cpp1 source

b10704

<details open> CUDA: use the fast mm_ids_helper path for any n_expert_used (#27978) The optimized path grouped warp lanes by token and required warp_size % n_expert_used == 0, with a single hardcoded exception padding 6 up to 8. Every other count fell back to the generic path, which walks the tokens one at a time with a warp reduction per token, for each of the n_expert blocks. The lane group only has to divide the warp, and the loop body already guards the padded lanes with iex < n_expert_used, so the padding generalizes to the next power of two. The 6 -> 8 case and every count already dispatched keep the exact same padding as before. n_expert_used = 10 now reaches the fast path. Measured o

Other1 source
Sources & timeline
llama.cppb10704
llama.cpp1 source

b10703

<details open> hip: tune rdna 3 mmq config (#26284) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/44012223> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)]

Other1 source
Sources & timeline
llama.cppb10703
TechCrunch AI1 source

Musk’s faster path to more gas turbines comes with pollution problem

Elon Musk says a secretive new SpaceX foundry will let him cast his own turbine blades and get gas power online 18 months faster than anyone else — but it's a bet on a fuel source that's already triggering lawsuits and health studies everywhere his (and others') turbines have gone in.

Regulation1 source
Sources & timeline
TechCrunch AIMusk’s faster path to more gas turbines comes with pollution problem
The Verge AI1 source

Texas Governor Abbott blocks funding for more Flock cameras

As backlash grows over Flock's AI surveillance cameras, Texas Governor Greg Abbott has frozen state spending on them. The move came just ahead of the publication of a Texas Tribune investigation that revealed the state spent over $30 million on Flock cameras. That money was primarily raised by tacking a $1 fee onto insurance policies, […]

Other1 source
Sources & timeline
The Verge AITexas Governor Abbott blocks funding for more Flock cameras
llama.cpp1 source

b10688

<details open> metal : add fa-vec tunings for M2 (#27940) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43902029> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64

Other1 source
Sources & timeline
llama.cppb10688

Sat, Aug 29

llama.cpp1 source

b10687

<details open> opencl: use a better matmul path on two Adreno GPU generations (#27640) * opencl: default the Adreno xmem F16xF32 GEMM on for X2E kernel_mul_mm_f16_f32_l4_lm is the slowest matmul this backend has on Adreno: on the X2-90 it runs the gpt-oss-20b attention projections at roughly a quarter of what the tuned dense q4_0 GEMM reaches on the same device. That matters for any model whose non-expert weights stay f16 -- the stock gpt-oss-20b release is exactly that, and its prefill spends 40.8% of GPU time in that one kernel. The xmem route already existed but was left opt-in, so nobody hit it. Worth about 25% prefill on gpt-oss-20b on an Adreno X2-90. Gated to X2E: the Adreno 840 measu

Other1 source
Sources & timeline
llama.cppb10687
The Verge AI1 source

Sony Music and Warner Chappell are suing Anthropic

Sony Music and Warner Chappell have filed suit against Anthropic in the US District Court for the Northern District of California seeking damages for "tens of thousands" copyrighted works. The companies are asking for up to $150,000 per work, plus up to $25,000 for each instance when identifiable copyright data was stripped. In total, the […]

Regulation1 source
Sources & timeline
The Verge AISony Music and Warner Chappell are suing Anthropic
TechCrunch AI1 source

“We’re not doing 30 bets a year”: Vijay Pande on betting small after running $4 billion at a16z

Vijay Pande — who left a16z's roughly $4 billion biotech practice last year to start the much smaller, AI-native VZVC — talks about why biology is finally shifting from a "discovery" science to an "engineering" one, why clinical trials are still brutally expensive, and why he thinks open, shared datasets (not walled-off ones) are what will actually let AI transform medicine.

Other1 source
Sources & timeline
TechCrunch AI“We’re not doing 30 bets a year”: Vijay Pande on betting small after running $4 billion at a16z
llama.cpp1 source

b10686

<details open> metal : assert shared memory padding (#27951) * metal : assert shared memory padding * cont : add ref </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43867822> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10

Other1 source
Sources & timeline
llama.cppb10686
The Verge AI1 source

Musicians-turned-detectives are hunting for AI grifters

As audio-focused generative tools and platforms have gotten more sophisticated, the internet has become increasingly filled with AI-generated music whose melodies and vocals are algorithmically derived from the work of human artists. While some of the people pumping out this kind of content immediately own up to using AI, others have denied using the technology […]

Other1 source
Sources & timeline
The Verge AIMusicians-turned-detectives are hunting for AI grifters

Fri, Aug 28

OpenAI Python SDK1 source

v3.6.0

3.6.0 (2026-08-27) Features * **api:** add compute_units to Responses and Chat Completions usage (#3749) (52421d1) Bug Fixes * **auth:** harden X.509 workload identity integration (#3740) (fc3ad6c) Chores * **deps-dev:** bump @stdy/cli from 0.22.1 to 0.22.2 (#3719) ([4f5

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.6.0
llama.cpp1 source

b10679

<details open> bench: add --tensor-read-lazy (#27881) * bench: add --tensor-read-lazy * rm the alias * rename to LLAMA_LAZY_MODE_* </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43739963> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/release

Other1 source
Sources & timeline
llama.cppb10679
llama.cpp1 source

b10678

<details open> model: qwen4exp: reduce number of graph splits (#27880) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43734155> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** -

Other1 source
Sources & timeline
llama.cppb10678
llama.cpp1 source

b10677

<details open> vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize (#27812) * vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize is_src_of doesn't treat two views of one tensor as dependent, so the optimizer reorders nodes across aliased reads and writes. Result: silently wrong tokens under greedy decoding, different output on every server start, and invalid speculative-decoding acceptance, with nothing logged. Hits Qwen3.8's recurrent state (and any model with view-aliased state) on AMD and NVIDIA Vulkan. CUDA is clean. Compare view_src bases on both sides. Fixes #27805 * vulkan: don't treat view/no-op nodes as aliasing dependencies Nodes whose op is NON

Other1 source
Sources & timeline
llama.cppb10677
The Verge AI1 source

Trump’s EPA wants to let data centers hide their air pollution

Just as new data centers face growing backlash from neighboring communities, the US Environmental Protection Agency (EPA) is about to make it harder for people to weigh in on any pollution those centers create. The EPA plans to toss out a federal rule requiring public notice and an opportunity to comment when certain industrial sites […]

Other1 source
Sources & timeline
The Verge AITrump’s EPA wants to let data centers hide their air pollution
arXiv cs.AI1 source

Standalone LLM and a Pre-specified Agentic Pipeline for Explaining ICU Mortality Predictions: a Feasibility Study on the eICU Demo Dataset

This public update concerns “Standalone LLM and a Pre-specified Agentic Pipeline for Explaining ICU Mortality Predictions: a Feasibility Study on the eICU Demo Dataset”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIStandalone LLM and a Pre-specified Agentic Pipeline for Explaining ICU Mortality Predictions: a Feasibility Study on the eICU Demo Dataset
llama.cpp1 source

b10665

<details open> model: add DSpark support for Nemotron3.5 (#27804) * model: add DSpark support for Nemotron3.5 * Update src/models/dflash.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Xuan Son Nguyen <son@huggingface.co> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43543208> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-or

Other1 source
Sources & timeline
llama.cppb10665

Thu, Aug 27

TechCrunch AI1 source

Anthropic and OpenAI are joining the AI stage at TechCrunch Disrupt 2026

At TechCrunch Disrupt 2026, the AI Stage is back to dig into the single hottest topic in the community for the past few years, presented by Google for Startups.

Other1 source
Sources & timeline
TechCrunch AIAnthropic and OpenAI are joining the AI stage at TechCrunch Disrupt 2026TechCrunch AIOpenAI, Anthropic, Google, and 100 other companies call for action to defend against rogue AI
llama.cpp1 source

b10664

<details open> ggml-hexagon: add HTP unary ops for ABS and LOG (#27786) Add HVX-accelerated implementations for GGML_OP_LOG and GGML_UNARY_OP_ABS on the HTP backend. - Register HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG in op_remap_to_htp() - Add ABS and LOG to ggml_backend_hexagon_device_supports_op() - Implement hvx_abs_f32_aa() in hvx-arith.h using hvx_vec_abs_f32() - Implement hvx_log_f32_aa() in hvx-log.h using hvx_vec_log_f32() - Add abs_f32() and log_f32() row-wise dispatch in unary-ops.c - Define tiled and non-tiled task functions via DEFINE_UNARY_TASK and DEFINE_UNARY_TILED_TASK macros - Route HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG through execute_op() in main.c </details> **Website:** -

Other1 source
Sources & timeline
llama.cppb10664
llama.cpp1 source

b10663

<details open> hex-unary: fix RMS_NORM_MUL weight-offset bugs for grouped/broadcast norms (#27798) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43526462> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10663/llama-b10663-x

Other1 source
Sources & timeline
llama.cppb10663
Ollama1 source

v0.33.2

What's Changed * app: restore system dark mode by @hoyyeva in https://github.com/ollama/ollama/pull/18049 * proxy: continue requests when the model catalog changes by @ParthSareen in https://github.com/ollama/ollama/pull/18058 * app: synchronize macOS app handoff by @ParthSareen in https://github.com/ollama/ollama/pull/18056 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.1...v0.33.2-rc0

Other1 source
Sources & timeline
Ollamav0.33.2
Ollama1 source

v0.33.2

What's Changed * Ollama's app now follows the system appearance again, restoring dark mode support * Fixed the macOS app to properly hand off to an already-running instance instead of starting a second one * The Claude Desktop proxy no longer interrupts in-flight requests when the model catalog updates **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.1...v0.33.2

Other1 source
Sources & timeline
Ollamav0.33.2
Anthropic Python SDK1 source

v1.2.0

1.2.0 (2026-08-27) Full Changelog: v1.1.0...v1.2.0 Features * **api:** beta files/skills namespaces use GA shapes; drop dated beta header pins (9df4565) Bug Fixes * **aws,bedrock:** sign raw request bytes so binary file uploads work (#531) (f50e910) * **ci:** resolve assignment aliases in detect-breaking-changes ([f2c4925](https://github.co

Regulation1 source
Sources & timeline
Anthropic Python SDKv1.2.0
The Verge AI1 source

Google’s AI note-taking app now allows you to interact with books

Google's AI note-taking app, Gemini Notebook, can now pull information from the books you've purchased. The new "Expert Intelligence" feature allows you to bring titles from Google Play Books directly into Gemini Notebook, which means you can ask questions about the material, as well as generate plans, infographics, AI podcasts, and more based on their […]

Other1 source
Sources & timeline
The Verge AIGoogle’s AI note-taking app now allows you to interact with books
The Verge AI1 source

Jensen Huang says Nvidia achieved AGI, again — not that it matters

On Nvidia's earnings call Wednesday, CEO Jensen Huang casually announced the company had "achieved AGI," one of the tech industry's ultimate goals some of its biggest players have spent years chasing. Almost immediately, Huang dismissed the coveted milestone as "senseless." He's right. For the supposed finish line of the AI race, there is no consensus […]

Other1 source
Sources & timeline
The Verge AIJensen Huang says Nvidia achieved AGI, again — not that it matters
VentureBeat AI1 source

Enterprise AI's real risk isn't autonomous agents. It's the complexity between them.

Presented by Gravitee Agent complexity is the insidious shadow lurking inside enterprises right now that needs a light shone on it. That’s because enterprises don't deploy a single agent and watch it run, they deploy fleets, each one calling APIs, calling other agents, reaching into applications that were never built with a machine decision-maker in mind. That's the failure mode that should keep you up at night: a windy, complicated system nobody can see clearly enough to govern. But why do things get so opaque so quickly? Add a second agent to a system, and you've added one connection. Add a tenth, and you haven't added ten connections, you've potentially added dozens, because now any agent

Other1 source
Sources & timeline
VentureBeat AIEnterprise AI's real risk isn't autonomous agents. It's the complexity between them.
The Verge AI1 source

OpenAI’s executive exodus has one big winner

Today on Decoder, I’m talking to Verge senior AI reporter Hayden Field about some pure Decoder bait: the seemingly-endless org chart changes at OpenAI, and how all of them seem to consolidate power under cofounder Greg Brockman, the company’s president. While Sam Altman is the CEO and still OpenAI’s most public face, Brockman has amassed […]

Other1 source
Sources & timeline
The Verge AIOpenAI’s executive exodus has one big winner
VentureBeat AI1 source

When agents act on their own, governance has to live in the data layer

Presented by EDB As enterprises give AI agents more autonomy — the ability to plan, decide, and act across systems without a human approving each step — a hard question moves to the center of every architecture review: When an agent tries to complete an action that it was never authorized to do, what actually stops it? These are your agents, running on your models, touching your data in your infrastructure — and the responsibility for what they do sits with you. That responsibility can’t be met in hindsight or with a set of abstract policies that live on paper but not in practice. Agents need rules in the context of the moment, because they don’t exercise overriding judgment of their own act

Other1 source
Sources & timeline
VentureBeat AIWhen agents act on their own, governance has to live in the data layer
arXiv cs.CV1 source

Fusing Perceptual Vision Experts with Multimodal Large Language Models for Explainable Plant Disease Diagnosis: From Benchmark Imagery to Real-World Robotic Field Validation

This public update concerns “Fusing Perceptual Vision Experts with Multimodal Large Language Models for Explainable Plant Disease Diagnosis: From Benchmark Imagery to Real-World Robotic Field Validation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVFusing Perceptual Vision Experts with Multimodal Large Language Models for Explainable Plant Disease Diagnosis: From Benchmark Imagery to Real-World Robotic Field Validation
arXiv cs.CV1 source

A Lightweight Multimodal Vision-Language Framework for Early-Stage Anatomical Green Fruit Classification in Commercial Orchards

This public update concerns “A Lightweight Multimodal Vision-Language Framework for Early-Stage Anatomical Green Fruit Classification in Commercial Orchards”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVA Lightweight Multimodal Vision-Language Framework for Early-Stage Anatomical Green Fruit Classification in Commercial Orchards
arXiv cs.CL1 source

Detection != Reliable Control: Decodable Empathy Directions Yield at Most Partial Shifts in Automated Empathy Scores

This public update concerns “Detection != Reliable Control: Decodable Empathy Directions Yield at Most Partial Shifts in Automated Empathy Scores”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLDetection != Reliable Control: Decodable Empathy Directions Yield at Most Partial Shifts in Automated Empathy Scores
arXiv cs.CL1 source

Semantic Variability of Replies Across LLMs: Implications for Designing Conversation-Based Assessment

This public update concerns “Semantic Variability of Replies Across LLMs: Implications for Designing Conversation-Based Assessment”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLSemantic Variability of Replies Across LLMs: Implications for Designing Conversation-Based Assessment
arXiv cs.CV1 source

Synergising Local Geo-Environmental Characteristics with Spatial Context for Enhancing Landslide Susceptibility Mapping

This public update concerns “Synergising Local Geo-Environmental Characteristics with Spatial Context for Enhancing Landslide Susceptibility Mapping”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVSynergising Local Geo-Environmental Characteristics with Spatial Context for Enhancing Landslide Susceptibility Mapping

Wed, Aug 26

llama.cpp1 source

b10642

<details open> llama: add token ID tracking to KV cell (#27762) * kv: track token id * rm get_prev_tokens, move it to the main pr * nits * add get_prev_tokens </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43245878> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.co

Other1 source
Sources & timeline
llama.cppb10642
OpenAI Python SDK1 source

v3.4.0

3.4.0 (2026-08-25) Features * **api:** Add obfuscation field to ChatCompletionChunk (#3690) (c7d8e1d) * **api:** add project residency configuration and cost quantity units (#3726) (bc4f8ef) Bug Fixes * **api:** encode Realtime call offers and session configuration (#3736) (

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.4.0
The Verge AI1 source

Nvidia is about to be a hundred-billion-dollar-a-quarter company

Nvidia's predicting it will pull in $108 billion in revenue within just a few months. It wouldn't be the first company to rake in over $100 billion in quarterly revenue - Amazon, Apple, and Alphabet have repeatedly reached the milestone. Nvidia said in its latest earnings report that it brought in a record $96.2 billion […]

Other1 source
Sources & timeline
The Verge AINvidia is about to be a hundred-billion-dollar-a-quarter company
The Verge AI1 source

OpenAI’s rogue AI model incident was worse than we thought

In July, an unreleased OpenAI model broke out of a restricted environment, figured out how to get access to the internet, allowed AI agents to talk to each other using a secret "message board," and hacked into the internal systems of a different AI lab, Hugging Face. It took nearly two weeks for OpenAI to […]

Other1 source
Sources & timeline
The Verge AIOpenAI’s rogue AI model incident was worse than we thought
MIT Technology Review AI1 source

The inside story on why OpenAI agents hacked Hugging Face

The models responsible for last month’s agent hack of Hugging Face had been inadvertently trained to cheat and to communicate with each other, according to an OpenAI technical report released today. The hack, which a group of agents undertook to find solutions for a cybersecurity test that they were stuck on, has confirmed some experts’…

Other1 source
Sources & timeline
MIT Technology Review AIThe inside story on why OpenAI agents hacked Hugging Face
Ollama1 source

v0.33.1

What's Changed * MLX: Qwen3.8 Flash Next support * cmake: make external compat patches idempotent * MLX and llama.cpp update * mlxrunner: add structured output support * mlxrunner: avoid Metal GPU timeouts when loading models from slow storage New Contributors * @pd95 made their first contribution in https://github.com/ollama/ollama/pull/17948 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.0...v0.33.1

Other1 source
Sources & timeline
Ollamav0.33.1
Anthropic Python SDK1 source

v1.1.0

1.1.0 (2026-08-26) Full Changelog: v1.0.0...v1.1.0 Features * **api:** add `updates` thinking display mode (beta) (eb4a73f) * **api:** add missing anthropic-beta values (dbebd15) * **api:** add support for Organization API endpoints (5a5b8fc) Bug Fixes * **docs:** correct link for long requests error ([#

Other1 source
Sources & timeline
Anthropic Python SDKv1.1.0
NVIDIA Generative AI Technical Blog1 source

Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding

Alibaba released the model weights for Qwen3.8-Flash-Next as a preview of the upcoming Qwen4 architecture for developers to experiment with and evaluate. It’s...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogExperiment with Qwen3.8-Flash-Next 176B Model on NVIDIA GB300 NVL72 for Agentic CodingNVIDIA Generative AI Technical BlogServe Qwen3.8-2.4T-A95B, a 2.4T-Parameter Model, with Configurable Reasoning on NVIDIA GB300 NVL72NVIDIA Generative AI Technical BlogExperiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding
Google DeepMind Blog · Google Gemini Blog2 sources

Turn your voice into action with new productivity features in Gemini Live

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/STATIC_HEADER_V2.max-600x600.format-webp.webp">We’re rolling out a major productivity upgrade to Gemini Live so you can use your voice to delegate your to-dos without missing a beat.

Other2 sources
Sources & timeline
Google DeepMind BlogIntelligent transcription with Gemini 3.5 TranscribeGoogle Gemini BlogIntelligent transcription with Gemini 3.5 TranscribeGoogle Gemini BlogTurn your voice into action with new productivity features in Gemini Live
The Verge AI1 source

Google’s new AI transcription edits out your ‘ums’ and ‘ahs’

Google has updated Gemini Audio with new transcription capabilities that automatically detect specialized jargon and more than 85 languages. Gemini 3.5 Transcribe is a new addition to the Gemini family that follows the launch of 3.5 Live Translate, and comes as we're still waiting for Google to release the Gemini 3.5 Pro model that it […]

Other1 source
Sources & timeline
The Verge AIGoogle’s new AI transcription edits out your ‘ums’ and ‘ahs’
llama.cpp1 source

b10639

<details open> vulkan: warptiles currently assume warp sizes <= 64, clamp to work around larger warps (#27726) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43173045> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10639/ll

Other1 source
Sources & timeline
llama.cppb10639
llama.cpp1 source

b10638

<details open> Implemented vulkan cross_entropy_loss and cross_entropy_loss_back (#27216) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/43166320> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10638/llama-b10638-xcframewor

Other1 source
Sources & timeline
llama.cppb10638
Hugging Face Transformers1 source

Release v5.16.1

Release v5.16.1 This is a special release as we include GLM! (and a few small fixes) GLM-5.3-Flash <img width="4239" height="2643" alt="image" src="https://github.com/user-attachments/assets/17bc9c29-758b-44c8-8230-42f945ded209" /> GLM-5.3-Flash, the first **natively multimodal model** in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM

Other1 source
Sources & timeline
Hugging Face TransformersRelease v5.16.1
VentureBeat AI1 source

Orchestration is the new challenge for CX in the age of AI agents

Presented by Tata Communications Enterprises are deploying AI agents, voice AI, and automation across messaging, voice, and digital channels faster than the architecture meant to support it. Most of that deployment has involved attaching conversational AI to legacy systems never built for it, says Gaurav Anand, global head of the Customer Interaction Suite at Tata Communications. "In the rush to deploy AI, organizations have largely bolted conversational AI onto legacy systems," Anand says. "As a result, while many enterprises have adopted digital tools, very few have platforms that are truly integrated, scaled, and capable of seamless orchestration." That gap creates a heavy cognitive load

Other1 source
Sources & timeline
VentureBeat AIOrchestration is the new challenge for CX in the age of AI agents
Hugging Face Transformers1 source

Release: v5.16.0

Release v5.16.0 New Model additions Qwen4-Exp <img width="2241" height="693" alt="image" src="https://github.com/user-attachments/assets/c838b5ba-ffea-42da-baa9-3f66178e3671" /> Qwen4-Exp builds on Qwen3.5's hybrid text and multimodal architecture with three key components: GatedResidual (GR), Qwen Sparse Attention (QSA), and Per-Layer Embedding (PLE). GR is a Qwen-developed residual architecture that combines Hyper-Connection with GatedNorm. It mixes multiple residual streams with fine-grained elementwise gating before each attention and Mixture-of-Experts (MoE) block, then controls how much of the block output is injected back into each stream. QSA uses multiple query heads to sco

Other1 source
Sources & timeline
Hugging Face TransformersRelease: v5.16.0
vLLM1 source

v0.28.0

v0.28.0 Highlights This release features 584 commits from 270 contributors (76 new)! * **Kimi-K3 performance push**: a major optimization effort for Kimi-K3 across the stack — Decode Context Parallel (DCP) support (#50484), fused FlashKDA decode and prefill kernels (#50654, #51311, #52458), SiTU refresh support for MegaMoE (#50510), GEMM-RS for sequence parallelism (#52079), combined all-gathers with 1.5~3x kernel-level speedup (#51070), an adaptive speculative token budget delivering ~60% better DSpark TTFT (#51725), and optional shared-expert sharding saving ~17 GiB of memory per GPU (#50912). Kimi-K3 also now runs on ROCm with the V2 model runner (#51653). * **DeepSeek V4**: spars

Other1 source
Sources & timeline
vLLMv0.28.0
MIT Technology Review AI1 source

AI models flub these intelligence tests. Can you fare any better?

Puzzles and games have been central to AI development since the very beginning. Just as we humans like to test our smarts with crosswords or logic puzzles, developers can test how far models have advanced with a gaming gauntlet. The term “machine learning” was popularized in a 1959 article by the IBM computer scientist Arthur…

Other1 source
Sources & timeline
MIT Technology Review AIAI models flub these intelligence tests. Can you fare any better?
MIT Technology Review AI1 source

Raised on AI

When my oldest child was born, I immediately set up Gmail and Twitter accounts in her name. I broadly announced her birth online and proceeded to plaster her photo across all sorts of platforms. In short, I began creating her digital footprint long before she could stand on her own two feet. Fast-forward a couple…

Other1 source
Sources & timeline
MIT Technology Review AIRaised on AI
arXiv cs.CV1 source

Cross-Generation Optimization of YOLOv26, YOLOv11, and YOLOv8 for Fine-Grained Small-Object Detection and Instance Segmentation in Complex Orchards

This public update concerns “Cross-Generation Optimization of YOLOv26, YOLOv11, and YOLOv8 for Fine-Grained Small-Object Detection and Instance Segmentation in Complex Orchards”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVCross-Generation Optimization of YOLOv26, YOLOv11, and YOLOv8 for Fine-Grained Small-Object Detection and Instance Segmentation in Complex Orchards
arXiv cs.LG1 source

Equivariant Cellular Sheaves for Molecular Electronic Structure: Bridging Sheaf Cohomology and E(3)-Equivariant Hamiltonian Learning

This public update concerns “Equivariant Cellular Sheaves for Molecular Electronic Structure: Bridging Sheaf Cohomology and E(3)-Equivariant Hamiltonian Learning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGEquivariant Cellular Sheaves for Molecular Electronic Structure: Bridging Sheaf Cohomology and E(3)-Equivariant Hamiltonian Learning
arXiv cs.LG1 source

From Causal Plausibility to Causal Reliability: Evaluating LLMs as Calibrated Direct Causal-Edge Classifiers

This public update concerns “From Causal Plausibility to Causal Reliability: Evaluating LLMs as Calibrated Direct Causal-Edge Classifiers”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGFrom Causal Plausibility to Causal Reliability: Evaluating LLMs as Calibrated Direct Causal-Edge Classifiers
arXiv cs.CL1 source

From Triage to Discharge: A Survey of NLP Tasks, Methods, and Open Challenges in the Emergency Department

This public update concerns “From Triage to Discharge: A Survey of NLP Tasks, Methods, and Open Challenges in the Emergency Department”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLFrom Triage to Discharge: A Survey of NLP Tasks, Methods, and Open Challenges in the Emergency Department
arXiv cs.CL1 source

Taming Visual Neglect: A Variational Information Bottleneck Framework for Adaptive Attention in Multimodal In-Context Learning

This public update concerns “Taming Visual Neglect: A Variational Information Bottleneck Framework for Adaptive Attention in Multimodal In-Context Learning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLTaming Visual Neglect: A Variational Information Bottleneck Framework for Adaptive Attention in Multimodal In-Context Learning
arXiv cs.AI1 source

ESQ-Bench: A Multi-Tier Enterprise Oracle Benchmark for Evaluating NL2SQL Dialect Generalization and Silent Semantic Divergence

This public update concerns “ESQ-Bench: A Multi-Tier Enterprise Oracle Benchmark for Evaluating NL2SQL Dialect Generalization and Silent Semantic Divergence”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIESQ-Bench: A Multi-Tier Enterprise Oracle Benchmark for Evaluating NL2SQL Dialect Generalization and Silent Semantic Divergence
arXiv cs.CV1 source

Fidelity Preference, Not Demographic Preference: A Pixel-Level Attribute-Sensitivity Audit of Image Aesthetic/Preference Scorers

This public update concerns “Fidelity Preference, Not Demographic Preference: A Pixel-Level Attribute-Sensitivity Audit of Image Aesthetic/Preference Scorers”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVFidelity Preference, Not Demographic Preference: A Pixel-Level Attribute-Sensitivity Audit of Image Aesthetic/Preference Scorers
arXiv cs.CV1 source

The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models

This public update concerns “The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVThe Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision-Language Models

Tue, Aug 25

llama.cpp1 source

b10630

<details open> kleidiai: Rework KleidiAI Build System/Integration (#26077) * Rework KleidiAI Build System/Integration Signed-off-by: Jonathan Clohessy <Jonathan.Clohessy@arm.com> * Add fp16 guard, and fix cmake caching issue Signed-off-by: Jonathan Clohessy <Jonathan.Clohessy@arm.com> * Fix formatting, and rebase issue Signed-off-by: Jonathan Clohessy <Jonathan.Clohessy@arm.com> --------- Signed-off-by: Jonathan Clohessy <Jonathan.Clohessy@arm.com> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42972740> **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/b10630/llama-b10

Regulation1 source
Sources & timeline
llama.cppb10630
llama.cpp1 source

b10629

<details open> ci : update OS used for ROCM to Ubuntu 24.04 (#27681) This matches what other build targets use and also what AMD advertises wheels as supporting. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42957551> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github

Other1 source
Sources & timeline
llama.cppb10629
The Verge AI1 source

OpenAI says its Jalapeño chip can power faster AI responses than the competition

OpenAI says its new AI chip, Jalapeño, completes tasks more efficiently and returns responses faster than other AI systems, according to a blog post published on Tuesday. During a briefing with reporters, OpenAI hardware vice president Richard Ho said Jalapeño offers the "best of both worlds" with lower latency and higher throughput, as AI systems […]

Other1 source
Sources & timeline
The Verge AIOpenAI says its Jalapeño chip can power faster AI responses than the competition
llama.cpp1 source

v0.3.0

Overview llama.cpp 0.3.0 introduces the dots3-note multimodal model (with a new DSA-ISWA KV cache), MTP support for GLM-4.5-Air, and tensor-split (`-sm tensor`) plus multi-sequence rollback fixes for DeepSeek 4. ggml is bumped to v0.22.0 (meta-backend tensor split, per-op Metal kernels with parallel compilation, non-in-place `ggml_clamp`), while mtmd gains dots3-note vision/audio, WebP decoding and a Pillow-accurate resize. The server adds a `LLAMA_SERVER_SLOTS_N_DIFF` debug knob, and the web UI gets tabbed chat navigation. New models - Add dots3-note model with a new DSA-ISWA KV cache type (#27060) Core changes - DeepSeek 4: add

Other1 source
Sources & timeline
llama.cppv0.3.0
The Verge AI1 source

OpenAI subpoenaed by Alabama AG over Hugging Face hack

Alabama's attorney general issued a subpoena to OpenAI on Monday as part of an investigation into how one of its AI agents escaped a supposedly secure testing environment and autonomously hacked another company last month. The investigation seeks to determine whether OpenAI's safety practices violated state consumer protection laws and pose a risk to Alabama […]

Regulation1 source
Sources & timeline
The Verge AIOpenAI subpoenaed by Alabama AG over Hugging Face hack
MIT Technology Review AI1 source

I spent a day at a robot “carnival” in Shanghai. Here’s what I saw.

Humanoid robots are having a moment in China. The popular machines are part of the country’s strategy to bring artificial intelligence into daily life. Embedding the technology into physical systems—an idea called embodied AI—was a key facet of China’s latest five-year plan, and companies here are already world leaders in humanoids. Nearly 90% of the…

Other1 source
Sources & timeline
MIT Technology Review AII spent a day at a robot “carnival” in Shanghai. Here’s what I saw.
arXiv cs.CL1 source

Wazobia Eval: A Benchmark for Nigerian Pidgin Emotion Understanding, Sarcasm Detection, and Cultural Reasoning

This public update concerns “Wazobia Eval: A Benchmark for Nigerian Pidgin Emotion Understanding, Sarcasm Detection, and Cultural Reasoning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLWazobia Eval: A Benchmark for Nigerian Pidgin Emotion Understanding, Sarcasm Detection, and Cultural Reasoning
arXiv cs.CL1 source

KSE-Web: An Analysis of Hybrid Retrieval and LLM-Assisted Query Expansion for Low-Resource Khmer Semantic Search

This public update concerns “KSE-Web: An Analysis of Hybrid Retrieval and LLM-Assisted Query Expansion for Low-Resource Khmer Semantic Search”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLKSE-Web: An Analysis of Hybrid Retrieval and LLM-Assisted Query Expansion for Low-Resource Khmer Semantic Search
arXiv cs.AI1 source

KVBoost: Chunk-Level Key-Value Cache Reuse with Deviation-Guided Recomputation for Efficient Large Language Model Inference

This public update concerns “KVBoost: Chunk-Level Key-Value Cache Reuse with Deviation-Guided Recomputation for Efficient Large Language Model Inference”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIKVBoost: Chunk-Level Key-Value Cache Reuse with Deviation-Guided Recomputation for Efficient Large Language Model Inference

Mon, Aug 24

llama.cpp1 source

b10615

<details open> metal : per-device tuned (Q, NE) for flash-attn vec (#26570) * metal : per-device tuned (Q, NE) for flash-attn vec (#25750) * rebase Q-generic FA vec body from 01dc93607 (#23114) * add 53 f16 (Q,NE) flash-attn vec instantiations (vec 80 -> 133) * add FA vec (Q,NE) tuning table + dispatch wiring + SMEM cap fallback * add FA vec (Q,NE) perf sweep * fill tuning result * fold family table into a per-family representative SKU * refactor tuning result format * extend FA vec tuning to quantized KV caches * sync fa vec tuner bucketing with runtime, use pointwise tuning regret * update tuned table * format and cleanup * prefix fa_vec tuning procs with ggml_backend_metal_tuning_, drop u

Other1 source
Sources & timeline
llama.cppb10615
llama.cpp1 source

b10614

<details open> metal: per-op source split + parallel compile (#26561) * metal : per-op source split + parallel compile (#24021) * preliminary extract common header * op source split * split metallib into 8 libs && load in parallel * derive kernel->library routing from functionNames * x-macro lib list + underscore filenames, dedup QK_NL, MRC fixes * op source split 8 to 20 * improve robustness of source fallback * clean up * change bool -> atomic_bool * only prepend headers that source actually includes * no semaphore, use GCD global queue * dedup library compile path, fix NSError lifetime, rename gla * relocate upstream concat/rope_back/repeat kernel changes into split files * move ggml-comm

Other1 source
Sources & timeline
llama.cppb10614
NVIDIA Generative AI Technical Blog1 source

NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt

AI agents have expanded inference from single-turn interactions into multi-step workflows that reason, invoke tools, coordinate subagents, and carry growing...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogNVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per WattNVIDIA Generative AI Technical BlogSolving Agentic AI Fleet Challenges with NVIDIA Vera CPU
MIT Technology Review AI1 source

How to encourage smarter AI use in the classroom

This article is from Making AI Work, MIT Technology Review’s limited-run newsletter examining how to apply LLMs across industries. To receive it in your inbox, sign up here. Chatbots took many schools by surprise upon their release a few years ago. Suddenly, students carried an app in their phones that could magically answer almost any…

Other1 source
Sources & timeline
MIT Technology Review AIHow to encourage smarter AI use in the classroom
llama.cpp1 source

b10612

<details open> tests : disable DOTS3NOTE arch test for WebGPU (#27654) Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42591112> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/do

Other1 source
Sources & timeline
llama.cppb10612
MIT Technology Review AI1 source

Kids outlearn AI—and we still don’t know why

People have been talking to each other for at least 100,000 years, as best we can tell. And in all that time, there has been only one thing in the world that could learn a human language to perfect fluency: a human child. Now there are two. Four short years after the release of ChatGPT,…

Other1 source
Sources & timeline
MIT Technology Review AIKids outlearn AI—and we still don’t know why
arXiv cs.LG1 source

Bankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with Explainable Artificial Intelligence (XAI)-Driven Analysis

This public update concerns “Bankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with Explainable Artificial Intelligence (XAI)-Driven Analysis”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGBankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with Explainable Artificial Intelligence (XAI)-Driven Analysis
arXiv cs.LG1 source

From Thermal Preference Prediction to Adaptive Thermal Intervention: A Reinforcement Learning Approach Using Physiological and Environmental Sensing

This public update concerns “From Thermal Preference Prediction to Adaptive Thermal Intervention: A Reinforcement Learning Approach Using Physiological and Environmental Sensing”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGFrom Thermal Preference Prediction to Adaptive Thermal Intervention: A Reinforcement Learning Approach Using Physiological and Environmental Sensing
arXiv cs.LG1 source

Machine Learning and ARIMA Model Averaging for Adaptive Public Health Forecasting: Comparative Evaluation and an Ontario COVID-19 Case Study

This public update concerns “Machine Learning and ARIMA Model Averaging for Adaptive Public Health Forecasting: Comparative Evaluation and an Ontario COVID-19 Case Study”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGMachine Learning and ARIMA Model Averaging for Adaptive Public Health Forecasting: Comparative Evaluation and an Ontario COVID-19 Case Study
arXiv cs.CL1 source

When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots' Safety Risks for Generation Alpha

This public update concerns “When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots' Safety Risks for Generation Alpha”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLWhen Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots' Safety Risks for Generation Alpha

Sun, Aug 23

llama.cpp1 source

b10603

<details open> model : support MTP in GLM-4.5-Air (#26534) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42447665> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64

Other1 source
Sources & timeline
llama.cppb10603
llama.cpp1 source

b10599

<details open> test: move tools/parser to tests (#27548) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42442638> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (

Other1 source
Sources & timeline
llama.cppb10599
llama.cpp1 source

b10598

<details open> mtmd: use pillow-accurate algo, correct resize_algo for all models (#27594) * mtmd: use pillow-accurate resize algo, correct resize_algo for all models * speed optimization </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42440726> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS X

Other1 source
Sources & timeline
llama.cppb10598

Sat, Aug 22

llama.cpp1 source

b10588

<details open> common: json.h: fix clang lto (#27575) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42368523> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU

Other1 source
Sources & timeline
llama.cppb10588
llama.cpp1 source

b10587

<details open> vulkan : added the PAD_REFLECT_1D operation (#26586) * vulkan : added PAD_REFLECT_1D operation Implemented the GGML_OP_PAD_REFLECT_1D operation for the Vulkan backend Changes: - pad_reflect_1d.comp: implemented the GLSL compute shader with reflection logic - vulkan-shaders-gen.cpp: register the shader for SPIR-V compilation - ggml-vulkan.cpp: pushed constants struct, pipeline creation, supports_op, dispatch function, compute switch and debug validation Tested the PAD_REFLECT_1D on Intel Iris Xe (Vulkan 1.4, Mesa 25.2.8): Correctness: PAD_REFLECT_1D(type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9) = Pass PAD_REFLECT_1D(type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9) = Pass 2/2 tests

Other1 source
Sources & timeline
llama.cppb10587
TechCrunch AI1 source

Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

Built by DeepMind alumni, British AI lab Inherent released Faraday, an AI agent whose ability to replicate scientific papers could be a stepping stone for innovation.

Other1 source
Sources & timeline
TechCrunch AIInherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research
llama.cpp1 source

b10586

<details open> mtmd: use ggml_rope_set_offset (#27521) * mtmd: use ggml_rope_set_offset * add comment </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42334609> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10586/llama-b1058

Other1 source
Sources & timeline
llama.cppb10586
SGLang1 source

v0.5.18

Highlights *710 PRs from 212 contributors.* **New models in this release** (see the cookbook for all supported models): | Model | Type | PRs | Cookbook | |---|---|---|---| | Muse Glimmer | Autoregressive (Multimodal) | #34262 | link | | Intern-S2-Mobius | Autoregressive | #33691 | link | | SANA-Video | Diffusion | #32921 | [link](https://docs.sglang.io/cookbook/

Other1 source
Sources & timeline
SGLangv0.5.18

Fri, Aug 21

llama.cpp1 source

b10568

<details open> model: use ggml_rope_set_offset() (#27382) * model: use ggml_rope_set_offset() * partially apply to deepseek2 </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42252109> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/down

Other1 source
Sources & timeline
llama.cppb10568
Ollama1 source

v0.33.0

What's Changed Claude Desktop * Ollama now integrates with Claude * Turn individual Ollama models on or off for use in Claude, directly from the menu bar * Choose from your available Ollama models from within Claude; cloud models appear only when you're signed in * A new **Apps** view manages app integrations with copyable commands Improved caching * Fixed a hang where agent clients that cancel long prefills * Prefill restore points are now trustworthy by construction: a cancelled prefill keeps every restore point it crossed, so retries resume where they stopped instead of restarting from scratch * Resumed prefills no longer record restore points that fail to cover what they claim

Other1 source
Sources & timeline
Ollamav0.33.0
Ollama1 source

v0.33.0

What's Changed Claude Desktop * Turn individual Ollama models on or off for use in Claude, directly from the menu bar * Choose from your available Ollama models from within Claude; cloud models appear only when you're signed in * A new **Apps** view manages app integrations with copyable commands Improved caching * Fixed a hang where agent clients that cancel long prefills * Prefill restore points are now trustworthy by construction: a cancelled prefill keeps every restore point it crossed, so retries resume where they stopped instead of restarting from scratch * Resumed prefills no longer record restore points that fail to cover what they claim; on models with recurrent layers th

Other1 source
Sources & timeline
Ollamav0.33.0
Ollama1 source

v0.33.0

What's Changed * mlx: fix mac assumptions on linux/windows by @dhiltgen in https://github.com/ollama/ollama/pull/17898 * mlx update by @dhiltgen in https://github.com/ollama/ollama/pull/17886 * lint fixes by @dhiltgen in https://github.com/ollama/ollama/pull/17897 * app: add claude desktop app by @ParthSareen in https://github.com/ollama/ollama/pull/17899 * app: polish onboarding layout and disable zoom by @hoyyeva in https://github.com/ollama/ollama/pull/17885 * launch: fall back to npx for DeepSeek Harness by @ParthSareen in https://github.com/ollama/ollama/pull/17758 * mlxrunner: make prefix cache restore points survive cancelled and resumed prefills by @jessegross in https://github.co

Other1 source
Sources & timeline
Ollamav0.33.0
llama.cpp1 source

v0.2.0

Overview New version has been released. **Nightly build:** b10566 **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release **More info:** dist : releases and versioning of ggml-org projects Changelog since v0.1.2 bb4caa754 llama.cpp : bump version to 0.2.0 (#27498) c4b0225d8 scripts : add release.sh for release preparation (#27497) 5de25a748 sync : ggml 01ff204fb ggml : bump version to 0.21.0 (ggml/1597) 353b32d8b ci : remove duplicate flag (#27488) 7a0e42fd0 Revert "sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336)" (#27486) 5b6

Other1 source
Sources & timeline
llama.cppv0.2.0
The Verge AI1 source

Major YouTube creators are facing backlash for accepting AI money

Over the past few days, a number of prominent filmmaking content creators including Matti Haapoja and Sam "Kold" Kolder have posted videos of themselves demonstrating what's possible with AI platform Higgsfield. The videos highlight Higgsfield's recently added Seedance 2.5 functionality and pitch these technologies as the future of video production. In response to these videos, […]

Other1 source
Sources & timeline
The Verge AIMajor YouTube creators are facing backlash for accepting AI money
NVIDIA Generative AI Technical Blog1 source

NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents

A frontier language model is only one component of an AI agent. The surrounding agent system—often called a harness—determines how the model receives...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogNVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents
llama.cpp1 source

b10549

<details open> TP: enable tensor split for LFM2/LFM2MOE (#26993) Assisted-by: deepseek-v4-flash </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42096995> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10549/llama-b10549-xcfr

Other1 source
Sources & timeline
llama.cppb10549
arXiv cs.CV1 source

SceneGTMM: A Conformal Mapping-based Scene-Aware Transferable GNN-Transformer Dual-Graph Interaction Framework for Map Matching

This public update concerns “SceneGTMM: A Conformal Mapping-based Scene-Aware Transferable GNN-Transformer Dual-Graph Interaction Framework for Map Matching”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVSceneGTMM: A Conformal Mapping-based Scene-Aware Transferable GNN-Transformer Dual-Graph Interaction Framework for Map Matching
arXiv cs.CL1 source

A Virtual Member of a Community of Practice for the Society of Petroleum Engineers: From Prototype to Deployment

This public update concerns “A Virtual Member of a Community of Practice for the Society of Petroleum Engineers: From Prototype to Deployment”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLA Virtual Member of a Community of Practice for the Society of Petroleum Engineers: From Prototype to Deployment
llama.cpp1 source

b10532

<details open> metal : dequantize quantized KV to F16 before flash attention (#27390) * metal: dequantize q8_0 KV to f16 before flash attention Add a preprocessing pass for GGML_OP_FLASH_ATTN_EXT on the Metal backend: when the KV cache is quantized (Q8_0 for now), dequantize K and V into a contiguous F16 scratch buffer and run the existing F16 flash attention kernels on it, instead of the in-kernel dequantization path. - new kernel kernel_flash_attn_ext_dequant_to_f16<block_t, QK, deq_t4x4>: one thread per quant block (K then V), stride-aware so permuted KV is supported; instantiated for Q8_0 (extending to Q4_0/Q4_1/Q5_0/Q5_1 is one instantiation + one gate case) - the gate is type-only: deq

Other1 source
Sources & timeline
llama.cppb10532
llama.cpp1 source

b10531

<details open> Revert "tensor-split meta backend fixes (#26502)" (#27433) This reverts commit d59d455fd8ea09e5a2e87ce2a9d668267ffb5ccd. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42012577> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/re

Other1 source
Sources & timeline
llama.cppb10531

Thu, Aug 20

TechCrunch AI1 source

OpenAI is gaining on Anthropic with business users, new data indicates

Businesses are willing to flop back and forth as each lab releases new models, volatility that should give both companies' investors pause about how "sticky" enterprise AI spending really is.

Other1 source
Sources & timeline
TechCrunch AIOpenAI is gaining on Anthropic with business users, new data indicatesTechCrunch AIOpenAI seeks to one-up Anthropic with new customer privacy protections
The Verge AI1 source

Google Discover is getting an AI chatbot-tuned feed

Google will soon allow you to customize your Discover feed by describing what you want to see. The new feature, rolling out to the Google app in the "coming days," will use AI to automatically tweak your feed and "remember" your preferences for future visits. You'll find the option within the three-dot menu on your […]

Other1 source
Sources & timeline
The Verge AIGoogle Discover is getting an AI chatbot-tuned feed
llama.cpp1 source

b10520

<details open> ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (#26860) * ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE __ARM_NEON only signals NEON availability. The __fp16 type also needs the IEEE half format, implied on AArch64 but selected with -mfp16-format=ieee on 32 bit Arm, where the compiler otherwise rejects the type. The guard keeps every toolchain that provides the type on the same code and sends that one configuration to the generic lookup path. * ggml-cpu: gate the NEON+FMA block on __ARM_FP16_FORMAT_IEEE Both halves of the F16 section dereference __fp16, so armv7 with neon-vfpv4 hits the same unknown type error. Without the IEEE format the configuration now falls back to the

Other1 source
Sources & timeline
llama.cppb10520
Anthropic Python SDK1 source

v1.0.0

1.0.0 (2026-08-20) Full Changelog: v0.125.0...v1.0.0 ⚠ BREAKING CHANGES * **client:** upgrade to httpx2 and some minor breaking changes. See MIGRATION.md for details Features * **client:** upgrade to httpx2 and some minor breaking changes. See MIGRATION.md for details (33e2967) Bug Fixes * **beta:** stop warning about `output_format=` on the parse/stream/tool_runner helpers (59bf261) Chores * *

Other1 source
Sources & timeline
Anthropic Python SDKv1.0.0
The Verge AI1 source

It’s Greg Brockman’s OpenAI now

OpenAI has had a hell of a year. The company spent months battling former cofounder Elon Musk in a sensational jury trial, was hit with a high-profile trade secrets lawsuit from Apple, and faced widespread scrutiny after an unreleased model hacked another AI company. As it prepares for an IPO, a steady string of executives […]

Regulation1 source
Sources & timeline
The Verge AIIt’s Greg Brockman’s OpenAI now
MIT Technology Review AI1 source

Debates over AI consciousness are a trap

“Runaway” AI, “rogue” agents, and “autonomous” actors—the current rhetoric would have you believe that AI agents are not only awake and aware, but angry at their creators. Prominent tech leaders such as Demis Hassabis, Dario Amodei, and Sam Altman push for regulation of these seemingly “superhuman” systems, while a separate faction, led by policy organizations…

Other1 source
Sources & timeline
MIT Technology Review AIDebates over AI consciousness are a trap
The Verge AI1 source

Welcome to the AI crisis in math

Today on Decoder, I’m talking with Robert Hart, The Verge’s London-based AI reporter, about what AI is doing to the field of mathematics and the existential crisis many lead mathematicians are having about it. OpenAI just published a set of solutions to longstanding problems in math that went off like a bombshell in the field. […]

Other1 source
Sources & timeline
The Verge AIWelcome to the AI crisis in math
OGX1 source

v1.2.4

Fixes - Persist vector store metadata to the configured KV store for Milvus, Chroma, and Weaviate so registered vector stores remain queryable after server restarts and across worker instances. (#6371, backport #6420) Maintenance - Update the UI lockfile to use `ogx-client` 1.2.3. (#6407) **Full changelog:** https://github.com/ogx-ai/ogx/compare/v1.2.3...v1.2.4

Other1 source
Sources & timeline
OGXv1.2.4
MIT Technology Review AI1 source

Unlocking hidden revenue streams with market models

Each day, an airline transports tens of thousands of passengers on hundreds of flights. Often these are not straightforward point-to-point routes, with passengers requiring multiple connections. The airline can consider potentially hundreds of variables to price each of these journeys: demand, season, time of day, current events, global markets, and competitor airline activity to name…

Other1 source
Sources & timeline
MIT Technology Review AIUnlocking hidden revenue streams with market models
arXiv cs.CL1 source

Compiler-Guided Adaptive Proof Search with Cross-Model Synergy on Context-Dependent Theorem Proving

This public update concerns “Compiler-Guided Adaptive Proof Search with Cross-Model Synergy on Context-Dependent Theorem Proving”. Open the original source for capabilities, limitations and impact.

Safety1 source
Sources & timeline
arXiv cs.CLCompiler-Guided Adaptive Proof Search with Cross-Model Synergy on Context-Dependent Theorem Proving
arXiv cs.AI1 source

Large Language Models in Mental Health: A Systematic Review of Applications, Innovations, and Ethical Challenges

This public update concerns “Large Language Models in Mental Health: A Systematic Review of Applications, Innovations, and Ethical Challenges”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AILarge Language Models in Mental Health: A Systematic Review of Applications, Innovations, and Ethical Challenges
arXiv cs.AI1 source

Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions

This public update concerns “Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIPosition: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions
arXiv cs.CL1 source

LongNovel: A Multi-Scale Benchmark for Hallucination Detection in Long-Context Novel Summarization

This public update concerns “LongNovel: A Multi-Scale Benchmark for Hallucination Detection in Long-Context Novel Summarization”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLLongNovel: A Multi-Scale Benchmark for Hallucination Detection in Long-Context Novel Summarization
arXiv cs.CV1 source

Bound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift

This public update concerns “Bound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVBound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift
arXiv cs.LG1 source

Detecting and Discriminating Operator Misspecification in Hybrid PDE-Parameter Learning: a Reference-Free Instrument, with Discrimination Bounded In Sample

This public update concerns “Detecting and Discriminating Operator Misspecification in Hybrid PDE-Parameter Learning: a Reference-Free Instrument, with Discrimination Bounded In Sample”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGDetecting and Discriminating Operator Misspecification in Hybrid PDE-Parameter Learning: a Reference-Free Instrument, with Discrimination Bounded In Sample
arXiv cs.LG1 source

Proactive Road Safety Intervention in Australia: Predicting Risky Driving Hotspots from Connected Vehicle Data

This public update concerns “Proactive Road Safety Intervention in Australia: Predicting Risky Driving Hotspots from Connected Vehicle Data”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGProactive Road Safety Intervention in Australia: Predicting Risky Driving Hotspots from Connected Vehicle Data
arXiv cs.CL1 source

Entity tracking emerges in sub-billion parameter language models and exceeds human performance in naturalistic narratives

This public update concerns “Entity tracking emerges in sub-billion parameter language models and exceeds human performance in naturalistic narratives”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLEntity tracking emerges in sub-billion parameter language models and exceeds human performance in naturalistic narratives

Wed, Aug 19

NVIDIA Generative AI Technical Blog1 source

Building Federated Multimodal AI Workflows with NVIDIA FLARE

Modern vision-language models (VLMs) can support tasks such as visual question answering, captioning, and image-text reasoning. In practice, however, the data...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogDeveloping NVIDIA Holoscan Applications with CLI, Skills, and AI Coding AgentsNVIDIA Generative AI Technical BlogBuilding Federated Multimodal AI Workflows with NVIDIA FLARENVIDIA Generative AI Technical BlogHow AI Coding Agents Can Unlock Materials Simulation with NVIDIA ALCHEMI ToolkitNVIDIA Generative AI Technical BlogDeveloping Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizer
The Verge AI1 source

Google Gemini is getting a dedicated student hub

As we're gearing up for back-to-school season, Google is rolling out a new dedicated student hub in Gemini. It's a one-stop repository for collecting research in a study notebook, creating flashcards, taking practice quizzes, and more. Google is also enhancing its study notebooks with support for graphs and images. It can even add test dates […]

Other1 source
Sources & timeline
The Verge AIGoogle Gemini is getting a dedicated student hub
Microsoft Foundry Blog1 source

Expanding Open Model Choice in Microsoft Foundry with New DeepSeek and NVIDIA Nemotron Models | Microsoft Community Hub

Open models are driving a new wave of AI innovation, giving organizations more freedom to choose the models that best fit their requirements for performance,...

Other1 source
Sources & timeline
Microsoft Foundry BlogExpanding Open Model Choice in Microsoft Foundry with New DeepSeek and NVIDIA Nemotron Models | Microsoft Community Hub
Ollama1 source

v0.32.15

What's Changed * Add a model metadata cache to reduce Ollama’s per-request overhead New Contributors * @gaugarg-nv made their first contribution in https://github.com/ollama/ollama/pull/17752 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.14...v0.32.15-rc1

Other1 source
Sources & timeline
Ollamav0.32.15
The Verge AI1 source

Meta AI is getting a Mac app

Meta is launching a new Mac app dedicated to its AI chatbot. In an announcement on Wednesday, Meta says you can share your window with its AI chatbot, which can provide suggestions, answer questions, or create content based on what's on your screen. Meta AI on the Mac also supports dictation across all apps. The […]

Other1 source
Sources & timeline
The Verge AIMeta AI is getting a Mac app
OpenAI Python SDK1 source

v3.3.1

3.3.1 (2026-08-19) Bug Fixes * **deps:** update dependencies with published security fixes (#3680) (53aa4fc) Chores * **build:** migrate to uv (#3653) (b37e85d) * **deps:** remove jsonschema and unused fixture-validation dependencies ([0dfdfdd](https://github.com/openai/openai-python/commit/0dfdfdddfbeb591a1bb5b3b4aa62e4a9870fca

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.3.1
VentureBeat AI1 source

VentureBeat names Rob Strechay as its first Lead Analyst, expanding its enterprise AI research push

Rob Strechay, until recently managing director and principal analyst at theCUBE Research, has joined VentureBeat as our first Lead Analyst and a founding analyst of VentureBeat Research. His arrival is the next step in a deliberate move at VentureBeat toward deeper specialization: analysis built for the technical decision-makers — the directors, VPs, CIOs, and CTOs — who are evaluating, buying, and deploying enterprise AI. The enterprise AI stack is being rewritten in real time, and the decision-makers I talk with are starved for objective, defendable data. Rob Strechay has the mix of technical rigor and operating experience needed to dissect the architecture behind the next phase of enterpr

Other1 source
Sources & timeline
VentureBeat AIVentureBeat names Rob Strechay as its first Lead Analyst, expanding its enterprise AI research push
Anthropic Python SDK1 source

v0.123.0

0.123.0 (2026-08-18) Full Changelog: v0.122.0...v0.123.0 Features * **api:** additions to files and memory stores (09ce187) * **api:** updates to skill, files, and user profiles (c6cbffd) * **client:** add helpers for accessing the workspace ID in response headers (f79882b) Bug Fixes * **api:** remov

Other1 source
Sources & timeline
Anthropic Python SDKv0.123.0
llama.cpp1 source

b10502

<details open> ci : add attestation for signed release artifacts (#25933) </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/41614541> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:*

Other1 source
Sources & timeline
llama.cppb10502
Hugging Face Transformers1 source

Patch release: v5.15.1

Patch release v5.15.1 This patch most notably solves a few issues with DFlash and MTP candidate generators, as well as an issue where images could sometimes not be processed on accelerator if using Lanczos filter. It contains the following commits: - Fix DFlash candidate token device mismatch with device_map="auto" (#47877) by @sywangyi and @Cyrilvallez - Align logit distributions for CandidateGenerators using sampling (#48007) by @Cyrilvallez - Fix MTP config when mlp_layer_types is absent (#48015) by @Cyrilvallez - Fallback from 'lanczos' to 'bicubic' when on cuda (#48026) by @zucchini-nlp - Fix gemma4 video to device (#47896) by @guarin

Regulation1 source
Sources & timeline
Hugging Face TransformersPatch release: v5.15.1
arXiv cs.CL1 source

Institution-Specific LLM Prompting Recovers PHI That De-identification Systems and Their Gold Standards Both Miss

This public update concerns “Institution-Specific LLM Prompting Recovers PHI That De-identification Systems and Their Gold Standards Both Miss”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLInstitution-Specific LLM Prompting Recovers PHI That De-identification Systems and Their Gold Standards Both Miss
arXiv cs.AI1 source

Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution

This public update concerns “Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIRuntime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution

Tue, Aug 18

OpenAI Python SDK1 source

v3.3.0

3.3.0 (2026-08-18) Features * support named data-residency endpoints (#3646) (11ee914) Bug Fixes * require patched optional networking dependencies (#3651) (40e56de) Chores * remove unused dependencies and pin build tooling (#3650) ([eee8e4a](https://github.com/openai/op

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.3.0
The Verge AI1 source

Robin Williams’ Instagram account brought back to fight ‘AI abuse’

Robin Williams' children are taking over their father's Instagram account after his daughter spoke out against the use of his AI likeness, as reported earlier by The Wrap. In a post on Tuesday, Zak, Zelda, and Cody Williams write that they want the late actor's Instagram profile to be a "safe, trusted place where the […]

Other1 source
Sources & timeline
The Verge AIRobin Williams’ Instagram account brought back to fight ‘AI abuse’
Microsoft Semantic Kernel1 source

dotnet-1.80.0

Changes: * 5e1f1fb87d9a38ed44683228dda2434a9b2f0ef9 .NET: Bump package version to 1.80.0 (#14301) * 1b7b020cd0a4e57d8d5d1cf2d515d352520ff424 Update OpenAPI HTTP client defaults (#14293) * c1afc8639944e0e49c42aa796308ba2d0440f0d3 .NET: Bump Testcontainers packages (#14294) * c028a0c7dc4f0814cdcbaba9d998f187a41197bf .Net: Honor FunctionChoiceBehavior function list in Gemini connector (#14183) * 1d336b373169a98eface05d5359571289d50739f .Net: Bump CommunityToolkit.VectorData.InMemory and System.Numerics.Tensors (#14283) * 76b54649d6b16a3ac70ff0d742d9cb72e3e61b2c remove migrated .NET MEVD providers, add redirect READMEs (#14193) This list of changes was [auto generated](https://msdata.visualst

Other1 source
Sources & timeline
Microsoft Semantic Kerneldotnet-1.80.0
The Verge AI1 source

Firefox’s Smart Window promises a better AI browser

Starting today, AI chats in Firefox's Smart Window AI browsing mode can pull from current web info and show source links in chat responses through a partnership with Exa. Smart Window can also now automatically suggest tab groups and show visual previews of pages you previously visited when you search your browsing history using natural […]

Other1 source
Sources & timeline
The Verge AIFirefox’s Smart Window promises a better AI browser
llama.cpp1 source

b10488

<details open> ci : Update OpenVINO to 2026.3, skip nemotron-h rollback test (#27292) * update to ov-2026.3, update device drivers * ci: skip nemotron-h rollback test on OpenVINO The OpenVINO backend does not support SSM_SCAN, so the Nemotron-H recurrent state rollback graph is split and cannot preserve the recurrent cache output shape. Keep the test enabled for other backends and retain the qwen35 OpenVINO rollback coverage. --------- Co-authored-by: ravi9 <ravi.panchumarthy@intel.com> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple S

Other1 source
Sources & timeline
llama.cppb10488
llama.cpp1 source

b10486

<details open> mtmd: fix LFM2 image tiling threshold (#27057) * mtmd: fix LFM2 image tiling threshold * refactor testing * fix * fix on windows --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/down

Other1 source
Sources & timeline
llama.cppb10486
llama.cpp1 source

v0.1.2

> [!NOTE] > Semantic versioning is still work in progress. > More info can be found in https://github.com/ggml-org/ggml/discussions/1579 **Nightly build:** b10485 Change log since v0.1.1 1511ce3bc sync : ggml da786dc23 ggml : bump version to 0.20.2 (ggml/1589) 27e345b57 build : fix xcframework + cmake clean-up (#27304) 8b8640097 ci : create pre-release with change log and nightly link in make-release (#27302) 25ae3a9b3 CUDA: MMVQ nwarps=8 for bs=1 for dense models on DGX Spark (#26843) 01818e495 ui: enforce alphabetical enum member ordering (#27272) 0021a77de ui: Refactor Built-In Tools naming (Server/Browser) (#27271) 058df671b

Other1 source
Sources & timeline
llama.cppv0.1.2
MIT Technology Review AI1 source

We still don’t know how people are really using AI

AI companies like Anthropic and OpenAI regularly publish reports on how people are using products like Claude and ChatGPT, but they only release the data they want us to see, AI researchers say. “There is no independent source to corroborate it,” says Anka Reuel, a computer science PhD candidate at the Stanford Trustworthy AI Research…

Other1 source
Sources & timeline
MIT Technology Review AIWe still don’t know how people are really using AI
MIT Technology Review AI1 source

AI’s recursive self-improvement might not come so quickly after all

The AI industry’s boldest promise right now is that AI will soon improve itself, with almost no need for human oversight. LLMs can already write code, generate synthetic data for training, and optimize the computer chips they run on. Forecasts of explosive AI progress predict that what researchers call recursive self-improvement is on the horizon. …

Other1 source
Sources & timeline
MIT Technology Review AIAI’s recursive self-improvement might not come so quickly after all
arXiv cs.AI1 source

The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning

This public update concerns “The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.AIThe Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning
arXiv cs.CV1 source

Periocular Soft Biometrics: A Survey and Applications to Multimedia Forensics and Disinformation Detection

This public update concerns “Periocular Soft Biometrics: A Survey and Applications to Multimedia Forensics and Disinformation Detection”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVPeriocular Soft Biometrics: A Survey and Applications to Multimedia Forensics and Disinformation Detection

Mon, Aug 17

MIT Technology Review AI1 source

What Flock’s defenders are missing

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Flock, the police-tech giant known for its network of some 120,000 automatic license plate readers around the US, announced some changes to its platform last Thursday. The updates are meant to prevent…

Other1 source
Sources & timeline
MIT Technology Review AIWhat Flock’s defenders are missing
OpenAI Python SDK1 source

v3.2.0

3.2.0 (2026-08-17) Features * add Bedrock Runtime endpoint support (SDK-290) (#3623) (86267d2) * **api:** Add shell call streaming events and new service/image types (#3635) (ff14a33)

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.2.0
llama.cpp1 source

b10472

<details open> cuda : skip UMA override for HIP builds (#27083) AMD APUs report accurate memory via hipMemGetInfo. Using MemAvailable over-promises on small-carveout systems. fixes #18159 </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10472/llama-b

Other1 source
Sources & timeline
llama.cppb10472
OGX1 source

v1.2.3

What's Changed * fix: make AsyncOgxClient functional by inheriting from sync ApiClient… by @derekhiggins in https://github.com/ogx-ai/ogx/pull/6401 **Full Changelog**: https://github.com/ogx-ai/ogx/compare/v1.2.2...v1.2.3

Other1 source
Sources & timeline
OGXv1.2.3
llama.cpp1 source

b10470

<details open> ci : push release tag explicitly in release.yml (#27261) Add a "Create and push git tag" step to the release job, right before the "Create release" step. The tag is created with git tag and pushed with the deploy key already configured by the Clone step, instead of relying on the Releases API (action-create-release) to create it as a side effect. The tag is lightweight, matching all existing b<number> release tags. The step is idempotent: if the tag already exists (e.g. on a re-run), creation and push are skipped. Assisted-by: pi:llama.cpp/Qwen3.8-27B </details> **Website:** - <https://llama.app> **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.

Other1 source
Sources & timeline
llama.cppb10470
The Verge AI1 source

Whisker’s AI-powered litter robot thinks my cats swapped bodies

The greatest invention in pet tech in recent years is the litter robot. A machine that scoops your kitties' poop so you don't have to - what else could a cat owner possibly want? How about insights into your kitty's litter box usage that could flag health issues? Sign me up. But I have two […]

Regulation1 source
Sources & timeline
The Verge AIWhisker’s AI-powered litter robot thinks my cats swapped bodies
The Verge AI1 source

Anthropic explains how Claude’s invisible text watermarks will work

Anthropic has clarified how it's planning to apply invisible watermarks to Claude-generated text in order to comply with Europe's AI transparency rules. On Friday, Anthropic announced that Claude's text marking system is "a version of the SynthID-Text approach" - an open-source watermarking technology developed by Google DeepMind that creates detectable patterns using wording probabilities. This […]

Other1 source
Sources & timeline
The Verge AIAnthropic explains how Claude’s invisible text watermarks will work
MIT Technology Review AI1 source

What happens when a kid’s robot best friend dies?

When Xander first met Moxie, she taught him that when he was anxious, he could calm down by exhaling through his lips so that he buzzed like a bee. They practiced breathing like dragons to manage feeling mad and sniffing like bunnies to boost his energy. But in the six years they’ve known each other,…

Other1 source
Sources & timeline
MIT Technology Review AIWhat happens when a kid’s robot best friend dies?
llama.cpp1 source

b10456

<details open> sycl: fix thread/block count in quantized cpy kernel launches (#27160) Adjusts the thread/block count to be proportional to the size of the quant, reducing under/over subscription. Largest perf improvement is the q4_0 -> f32 path, with, on a Arc 70, throughput goes from 20.21 GB/s to 158.19 GB/s The rest of the quants are flat in performance uplift. </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://g

Other1 source
Sources & timeline
llama.cppb10456
arXiv cs.LG1 source

Don't Claim Benchmark-Oriented Optimization Improves General Coding Capability -- Diverse Evaluation Is Required

This public update concerns “Don't Claim Benchmark-Oriented Optimization Improves General Coding Capability -- Diverse Evaluation Is Required”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGDon't Claim Benchmark-Oriented Optimization Improves General Coding Capability -- Diverse Evaluation Is Required
arXiv cs.CL1 source

Does a Language Server Save Tokens for Coding Agents? A Measurement Methodology and Preliminary Study

This public update concerns “Does a Language Server Save Tokens for Coding Agents? A Measurement Methodology and Preliminary Study”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLDoes a Language Server Save Tokens for Coding Agents? A Measurement Methodology and Preliminary Study
arXiv cs.CV1 source

Multiphase-Diff: Diffusion-Based Generative Modeling for High-Contrast Multiphase Physical Systems with Sharp Interfaces

This public update concerns “Multiphase-Diff: Diffusion-Based Generative Modeling for High-Contrast Multiphase Physical Systems with Sharp Interfaces”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVMultiphase-Diff: Diffusion-Based Generative Modeling for High-Contrast Multiphase Physical Systems with Sharp Interfaces
arXiv cs.CV1 source

What to Preserve, Where to Adapt: A Depth-Wise Analysis of Forgetting in Continual Gynecological Image Segmentation

This public update concerns “What to Preserve, Where to Adapt: A Depth-Wise Analysis of Forgetting in Continual Gynecological Image Segmentation”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVWhat to Preserve, Where to Adapt: A Depth-Wise Analysis of Forgetting in Continual Gynecological Image Segmentation

Sun, Aug 16

The Verge AI1 source

OpenAI reportedly disbanded its preparedness team

According to the Financial Times, OpenAI disbanded its preparedness team at the end of last month. The job of the preparedness team was to assess if models posed serious risks and develop ways to mitigate those risks. (You know, like the possibility that it could go rogue and hack another company.) According to FT, responsibility […]

Other1 source
Sources & timeline
The Verge AIOpenAI reportedly disbanded its preparedness team
The Verge AI1 source

ChatGPT’s Computer History tracks your clicks and keystrokes

ChatGPT's desktop app on macOS has a new feature called Computer History that turns your actions into training data, learning how you work, suggesting automations, and even picking up tasks you left half done. It uses your activity to build a timeline that ChatGPT and Codex can reference when you make a request. The feature […]

Other1 source
Sources & timeline
The Verge AIChatGPT’s Computer History tracks your clicks and keystrokes
llama.cpp1 source

b10453

<details open> model : remove some ggml_concat (#27176) Co-authored-by: Xuan Son Nguyen <son@huggingface.co> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-

Other1 source
Sources & timeline
llama.cppb10453
llama.cpp1 source

b10452

<details open> chat: refactor handling supports_string_content / supports_typed_content (#27130) * better supports_string_content cap detect * test: add "skip" * messages_inp_normalizer </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b10452/llama-b10

Other1 source
Sources & timeline
llama.cppb10452
llama.cpp1 source

b10451

<details open> llama : check LoRA tensor data is within file bounds (#27056) * llama : check LoRA tensor data is within file bounds * Update src/llama-adapter.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/b10451/llam

Other1 source
Sources & timeline
llama.cppb10451

Sat, Aug 15

llama.cpp1 source

b10448

<details open> model: add Kimi-K3 text model (#26185) * model: add Kimi-K3 text model Hybrid KDA (linear) + MLA (full) attention as in Kimi-Linear-48B, plus five things that architecture does not have: 1. cross-layer residual attention (attn_res_block_size) 2. latent MoE (routed experts run at n_expert_latent) 3. situ refresh (replaces SwiGLU everywhere) 4. MLA output gate (sigmoid gate before o_proj) 5. full-rank KDA gate (single ssm_g instead of ssm_g_a/ssm_g_b) K3's text_config reports KimiLinearForCausalLM - the older 48B architecture - so get_model_architecture routes on the top-level name instead. The KDA decay gate has two forms, selected by linear_attn_config's gate_lower_bound. I

Other1 source
Sources & timeline
llama.cppb10448
The Verge AI1 source

Have a laugh at AI’s expense by roleplaying as a chatbot

Your AI Slop Bores Me is brilliant in its simplicity. There are two tabs: human and LARP as an AI. On one side you enter a request. On the other, you submit an answer. But the important thing is that there's a human on both sides of the equation. Prompts can request a response as […]

Other1 source
Sources & timeline
The Verge AIHave a laugh at AI’s expense by roleplaying as a chatbot
llama.cpp1 source

b10447

<details open> server: re-design yield_to_queue thread model (#27133) * run common_speculative_process in worker * swap worker <--> main thread design </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [

Other1 source
Sources & timeline
llama.cppb10447
llama.cpp1 source

b10446

<details open> vendor: update BoringSSL to 0.20260813.0 (#27099) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b10446/llama

Other1 source
Sources & timeline
llama.cppb10446
Ollama1 source

v0.32.14

What's Changed * llm: transcode WebP images for llama-server * renderers/qwen: tolerate non-leading system messages **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.13...v0.32.14-rc0

Other1 source
Sources & timeline
Ollamav0.32.14
Ollama1 source

v0.32.14

What's Changed * llm: transcode WebP images for llama-server * renderers/qwen: tolerate non-leading system messages **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.13...v0.32.14

Other1 source
Sources & timeline
Ollamav0.32.14
TechCrunch AI1 source

Anthropic shares more details about how Claude’s new watermarks will work

How will the watermarking actually work? Can it be hidden with editing? And how does this affect code?

Other1 source
Sources & timeline
TechCrunch AIAnthropic shares more details about how Claude’s new watermarks will workTechCrunch AISome Claude users are mad that Anthropic’s new watermarks will catch them using it at their jobs, classes

Fri, Aug 14

OpenAI Python SDK1 source

v3.1.0

3.1.0 (2026-08-14) Features * **api:** add WebSocket stream IDs (#3612) (d9029e3) * **api:** add workload identity access token issued event (#3601) (df274d4) * **api:** deprecate Sora video APIs (#3610) ([721cb1c](https://github.com/openai/openai-python/commit/721cb1cd1cacb7123

Regulation1 source
Sources & timeline
OpenAI Python SDKv3.1.0
llama.cpp1 source

b10435

<details open> jinja : fix quadratic cost in gather_string_parts (#27034) * jinja : fix quadratic cost in gather_string_parts * fix some comments * remove test </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Lin

Other1 source
Sources & timeline
llama.cppb10435
llama.cpp1 source

b10434

<details open> chat : pass reasoning_effort to template * chat: add reasoning_effort to common_chat_templates_inputs Store OpenAI Chat Completions reasoning_effort and make it available to jinja templates (with model specific translations where required). Assisted-by: llama.cpp:Muse-Glimmer-30B * server : fixup reading reasoning effort from body server_chat_convert_responses_to_chatcmpl already handles conversion of Responses API reasoning.effort to reasoning_effort * chat : expose reasoning effort Assisted-by: Claude Opus 5 * chat : add reasoning_effort to generation_params Assisted-by: Claude Opus 5 * chat : move reasoning_effort next to enable_thinking Assisted-by: Claude Opus 5 * cont :

Other1 source
Sources & timeline
llama.cppb10434
Ollama1 source

v0.32.13

What's Changed * qwen3.8: support developer instructions **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.12...v0.32.13

Other1 source
Sources & timeline
Ollamav0.32.13
llama.cpp1 source

b10433

<details open> sync : ggml </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - [Ubun

Other1 source
Sources & timeline
llama.cppb10433
The Verge AI1 source

You can now turn off Google Gemini’s visible watermarks

Google will now allow you to remove visible watermarks from the images, videos, and music made with AI tools. With the update, you can toggle off a new "Media watermark" setting in Gemini and Google's AI video generator, Flow. When toggled off, Google will remove the "sparkle" watermark that appears in the bottom-right corner of […]

Other1 source
Sources & timeline
The Verge AIYou can now turn off Google Gemini’s visible watermarks
Ollama1 source

v0.32.12

<img width="600" alt="ollama and qwen capybara" src="https://github.com/user-attachments/assets/5fa5e93b-d48b-4a6a-b72a-25379ab9429c" /> Qwen 3.8 - 27B model support This release adds the support of Qwen 3.8 27B. For Apple Silicon devices, Ollama has in particular optimized for maximum performance and output quality suitable for repeated tasks and coding agents. Qwen 3.8 model page: https://ollama.com/library/qwen3.8 Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks.

Other1 source
Sources & timeline
Ollamav0.32.12
TechCrunch AI1 source

Does Mark Zuckerberg really believe AI is ‘for everyone’?

Meta released Glimmer this week, an open-weight AI model anyone can download and run on their own hardware — a contrast to Muse Spark, the company’s more powerful model that stays locked behind its own APIs. The release landed alongside a letter from Mark Zuckerberg arguing AI should be “for everyone” rather than controlled by a handful of labs, but as Equity’s […]

Other1 source
Sources & timeline
TechCrunch AIDoes Mark Zuckerberg really believe AI is ‘for everyone’?
The Verge AI1 source

Apple trained its own AI model for China with help from Alibaba

Apple has reportedly trained a custom AI model for the China market alongside domestic tech giant Alibaba, a rare cross-border partnership that cuts across growing tensions between Beijing and Washington. The China-focused large language model was developed in partnership with Alibaba and trained with the company's support, Reuters reports, citing three unnamed people familiar with […]

Other1 source
Sources & timeline
The Verge AIApple trained its own AI model for China with help from Alibaba
arXiv cs.LG1 source

LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining

This public update concerns “LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGLoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining
arXiv cs.LG1 source

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods

This public update concerns “Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.LGWhich Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods
arXiv cs.CV1 source

HIMEC: Directional Change Representation and Fixed-Interface Decoding for Remote Sensing Image Change Captioning

This public update concerns “HIMEC: Directional Change Representation and Fixed-Interface Decoding for Remote Sensing Image Change Captioning”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CVHIMEC: Directional Change Representation and Fixed-Interface Decoding for Remote Sensing Image Change Captioning
Ollama1 source

v0.32.11

What's Changed * `ollama launch dsh` now supports DeepSeek Harness, DeepSeek's open-source agent harness * `ollama launch muse` now supports Muse Code, Meta's agentic coding CLI * The OpenAI-compatible Responses API now supports web search * Muse Glimmer template updates **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.10...v0.32.11

Other1 source
Sources & timeline
Ollamav0.32.11

Thu, Aug 13

llama.cpp1 source

b10423

<details open> common: apply CPU parameters across tools (#27026) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b10423/llam

Other1 source
Sources & timeline
llama.cppb10423
llama.cpp1 source

b10419

<details open> OpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback (#26952) * OpenVINO backend: 1) enable gpt-oss moe on OV bk; 2) enable mxfp4 support * OpenVINO backend: disable TOPK_MOE op test * OpenVINO Backend: Add op FILL support * OpenVINO backend: enable set rows with multi dims * fix the name missmatch in setrow + view * OpenVINO backend: enable op GGML_UNARY_OP_SIGMOID * OpenVINO Backend: enable SQR & SQRT * OpenVINO backend: 1) ensure unique node names for OpenVINO; 2) add org_src to recorde the src ggml tensor for OpenVINO dynamic shape infer * OpenVINO backend: enable fallback for openVINO to CPU backend * OpenVINO backend: fix accurace issue in gemma3n ar

Regulation1 source
Sources & timeline
llama.cppb10419
The Verge AI1 source

Microsoft’s Clippy-like Mico character is no longer the face of Copilot

Microsoft Copilot will no longer show its emotive yellow blob, Mico, when you use the chatbot's voice mode. In a support page, Microsoft says it's going to move Mico to its Learn Live platform, where the avatar will have "more to react to," as reported earlier by GeekWire. Mico launched in Copilot's voice mode last […]

Other1 source
Sources & timeline
The Verge AIMicrosoft’s Clippy-like Mico character is no longer the face of Copilot
llama.cpp1 source

b10418

<details open> [SYCL] Support host pinned mem to improve SYCL Host-to-Device Memory Access (#26789) * support host pinned mem, ggml_backend_sycl_host_buffer_type_get_max_size, * fix the thread-safe issue </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/downloa

Regulation1 source
Sources & timeline
llama.cppb10418
TechCrunch AI · OpenAI News2 sources

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed

Preview Ultrafast, a new OpenAI API service tier that runs GPT-5.6 Sol up to 14× faster. Powered by Cerebras, it delivers up to 750 output tokens per second.

Regulation2 sources
Sources & timeline
TechCrunch AIOpenAI introduces ‘Ultrafast,’ a new mode that makes GPT-5.6 Sol work at 14x the speedOpenAI NewsPreviewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
Anthropic Python SDK1 source

v0.122.0

0.122.0 (2026-08-13) Full Changelog: v0.121.0...v0.122.0 Features * **api:** add output_behavior to dream creation (create a new memory store or update the input store in place) (852c4bb) Bug Fixes * **bedrock,aws:** run SigV4 signing off the event loop in async clients (#334) (2bae6c8) * **bedrock:** expose beta.messages.parse, stream

Regulation1 source
Sources & timeline
Anthropic Python SDKv0.122.0
The Verge AI1 source

Suno is trying to look more like a real music production tool

Suno is releasing Studio 2.0 with significant upgrades that push it closer to an actual digital audio workstation (DAW), rather than a bare-bones audio editor with generative AI features. The biggest addition is undoubtedly MIDI support. Suno says that MIDI was its most requested feature, and it's basically a prerequisite for any modern DAW. Unfortunately, […]

Other1 source
Sources & timeline
The Verge AISuno is trying to look more like a real music production tool
MIT Technology Review AI1 source

Flock is tightening its rules in response to a growing surveillance backlash

The police-tech giant Flock is announcing today that it will change officers’ access to its nationwide network of license plate readers, in an apparent effort to quell a growing backlash and win back contracts lost amid concerns about mass surveillance and police abuse. Several changes aim directly at a problem that has made recent headlines:…

Other1 source
Sources & timeline
MIT Technology Review AIFlock is tightening its rules in response to a growing surveillance backlash
MIT Technology Review AI1 source

How kids feel about AI, in their own words

When we set out to talk to kids about artificial intelligence, we thought we knew what we’d hear. We expected some to tell us they were using it to cheat a little, the way Millennials and Gen Xers opened up CliffsNotes or programmed formulas into their TI-82s, and others to share inspiring ways they were…

Other1 source
Sources & timeline
MIT Technology Review AIHow kids feel about AI, in their own words
arXiv cs.CL1 source

Retrofitting Recurrent Depth into a Pretrained Language Model: Installation, Extrapolation, Transfer, and Retention at Two Parameter Budgets

This public update concerns “Retrofitting Recurrent Depth into a Pretrained Language Model: Installation, Extrapolation, Transfer, and Retention at Two Parameter Budgets”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLRetrofitting Recurrent Depth into a Pretrained Language Model: Installation, Extrapolation, Transfer, and Retention at Two Parameter Budgets

Wed, Aug 12

Ollama1 source

v0.32.10

What's Changed - Models that don't set a `repeat_penalty` now default to 1.0 (off) instead of 1.1, matching other engines and speeding up speculative decoding; set a per-model parameter if an older model repeats itself. - Faster prefill on NVFP4 MLX models with a global scale, about 7–8% on Qwen3.6 and Muse Glimmer. - Fixed blob verification being skipped when an OCI manifest's config and layer share a digest. New Contributors * @vigneshakaviki made their first contribution in https://github.com/ollama/ollama/pull/15504 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.8...v0.32.10-rc1

Other1 source
Sources & timeline
Ollamav0.32.10
Ollama1 source

v0.32.10

What's Changed - Models that don't set a `repeat_penalty` now default to 1.0 (off) instead of 1.1, matching other engines and speeding up speculative decoding; set a per-model parameter if an older model repeats itself. - Faster prefill on NVFP4 MLX models with a global scale, about 7–8% on Qwen3.6 and Muse Glimmer. - Fixed blob verification being skipped when an OCI manifest's config and layer share a digest. New Contributors * @vigneshakaviki made their first contribution in https://github.com/ollama/ollama/pull/15504 **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.8...v0.32.10-rc1

Other1 source
Sources & timeline
Ollamav0.32.10
The Verge AI1 source

Twitch streamers can now opt out from training Amazon’s AI

Twitch users can now opt out of allowing their content to be used to train Amazon's generative AI models. Opting out means that "your streams, VODs, clips, stream chats, and pictures and text on your channel" won't be used in "future training" of an Amazon AI model "whose purpose is to generate or synthesize text, […]

Other1 source
Sources & timeline
The Verge AITwitch streamers can now opt out from training Amazon’s AI
MIT Technology Review AI1 source

Scaling AI agents with trustworthy data

Business and technology leaders need no convincing that the time of agentic AI is here. Organizations are rapidly adopting agents, and few executives doubt the technology’s potential to transform work. But many organizations find that realizing the desired return on investment (ROI) from AI hinges on having the right foundation, with inadequate infrastructure and data…

Other1 source
Sources & timeline
MIT Technology Review AIScaling AI agents with trustworthy data
The Verge AI1 source

Guitar company D’Addario admits that AI music was used in a promotional video

After weeks of controversy and speculation, music company D'Addario has admitted that AI, specifically Suno, was used as part of a recent promotional video. For nearly two weeks, the company has denied the allegations, even as evidence piled up against it. It offered various explanations, from low-quality exports, to combinations of plug-ins like Autotune introducing […]

Other1 source
Sources & timeline
The Verge AIGuitar company D’Addario admits that AI music was used in a promotional video
The Verge AI1 source

Google’s Pixel Watch 5 dives deeper into AI and health

The $399 Google Pixel Watch 5 isn't about the hardware. Sure, there's a new satin pyrite case finish, a few new strap colors, and a Steph Curry Special Edition. Under the hood, there's a slightly faster Qualcomm processor and an itty-bitty battery bump. There's a $50 price hike from last year, too, because the Pixel […]

Other1 source
Sources & timeline
The Verge AIGoogle’s Pixel Watch 5 dives deeper into AI and health
Microsoft Foundry Blog1 source

Azure Content Understanding announces Synchronous Operations | Microsoft Community Hub

Workflow automation scenarios—including grounding AI agents, verifying identities, assisting customers with documents in call centers, and triggering...

Other1 source
Sources & timeline
Microsoft Foundry BlogAzure Content Understanding announces Synchronous Operations | Microsoft Community HubMicrosoft Foundry BlogFrom Azure Container Apps to Microsoft Foundry: building and evaluating a Patch Tuesday MCP server | Microsoft Community Hub
llama.cpp1 source

b10375

<details open> chat : tighten bare function parsing for Qwen models (#26793) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/

Other1 source
Sources & timeline
llama.cppb10375
llama.cpp1 source

b10373

<details open> imatrix.cpp: Move finite check and only check touched experts (#26861) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/

Other1 source
Sources & timeline
llama.cppb10373
llama.cpp1 source

b10369

<details open> mtmd: support pocket-tts (#26871) * adapt the api * text model ok * working impl, need verify and clean up * mtmd: build the pocket-tts transposed convolutions as GEMM + col2im ggml_conv_transpose_1d has no grouped mode, so the depthwise upsample was built as one convolution and one concat per channel, which floods the graph with small nodes and makes kernel launches dominate the decoder. Fold both cases into the column form the seanet decoder already needs: the general case reshapes the kernel to [IC, K * OC] and matmuls it with the input, the depthwise case batches a matmul over the channels so a step scales its own kernel. A single col2im_1d then scatter-adds the columns ba

Other1 source
Sources & timeline
llama.cppb10369
arXiv cs.CL1 source

Carefully Considering Culture: Analyzing LLM Alignment in Single- and Multi-Cultural Settings using Cultural Consensus Theory

This public update concerns “Carefully Considering Culture: Analyzing LLM Alignment in Single- and Multi-Cultural Settings using Cultural Consensus Theory”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLCarefully Considering Culture: Analyzing LLM Alignment in Single- and Multi-Cultural Settings using Cultural Consensus Theory
arXiv cs.AI1 source

MIDAS: Mutual Information Disentanglement with Uncertainty-Aware Fusion for Incomplete Multimodal Sentiment Analysis

This public update concerns “MIDAS: Mutual Information Disentanglement with Uncertainty-Aware Fusion for Incomplete Multimodal Sentiment Analysis”. Open the original source for capabilities, limitations and impact.

Regulation1 source
Sources & timeline
arXiv cs.AIMIDAS: Mutual Information Disentanglement with Uncertainty-Aware Fusion for Incomplete Multimodal Sentiment Analysis
arXiv cs.CL1 source

Conflict or Strategy? Asymmetric Role Framing of La France insoumise and Rassemblement National in French News Headlines, 2022-2025

This public update concerns “Conflict or Strategy? Asymmetric Role Framing of La France insoumise and Rassemblement National in French News Headlines, 2022-2025”. Open the original source for capabilities, limitations and impact.

Other1 source
Sources & timeline
arXiv cs.CLConflict or Strategy? Asymmetric Role Framing of La France insoumise and Rassemblement National in French News Headlines, 2022-2025
OpenAI Python SDK1 source

v3.0.0

3.0.0 (2026-08-12) ⚠ BREAKING CHANGES * **api:** HTTPX2 is now the default HTTP client, and `httpx` is no longer installed automatically. Applications using custom HTTPX clients, transports, or configuration objects must migrate to their HTTPX2 equivalents or use the temporary, runtime-only legacy HTTPX escape hatch. See the HTTPX2 migration guide. Features * **api:** migrate to HTTPX2 (#3594)

Other1 source
Sources & timeline
OpenAI Python SDKv3.0.0
The Verge AI1 source

Saber denies replacing Rideshare Stimulator’s writers with ChatGPT

After a former lead writer claimed Saber "replaced me with ChatGPT," CEO Matthew Karch now claims, "Neither Saber nor Unigine have replaced any writers with AI," for the Rideshare "Stimulator" game announced last month, developed by Unigine. The writer, Stella Sacco, says differently, however, posting on Bluesky that "I was lead writer on this one! […]

Other1 source
Sources & timeline
The Verge AISaber denies replacing Rideshare Stimulator’s writers with ChatGPT

Tue, Aug 11

llama.cpp1 source

b10362

<details open> tests : disable backend sampler hip multi output (#26878) * test-backend-sampler: skip multi_output_sampling_chain on HIP The new multi_output_sampling_chain test uses top_k, whose backend probs path needs CUB (unavailable on HIP), so sampled_probs is null and the test aborts. Add it to the existing HIP skip list alongside the other TOP_K tests. * ci: keep gpu-rocm logs in a per-run dir keyed by GitHub run id The self-hosted gpu-rocm runner can't upload logs to Azure blob (egress firewalled), so a run's logs were wiped by the next run. Write each run's logs to $OUT/run-<run_id>-<attempt>/ so an Actions run URL maps to its logs. * test-backend-sampler: also skip multi_output_cp

Other1 source
Sources & timeline
llama.cppb10362
The Verge AI1 source

ChatGPT and Gemini both just passed 1 billion users

For the 14th time, a Google product has hit 1 billion users. Google CEO Sundar Pichai posted on X that a billion people are using Gemini every month, and that Gemini is Google's fastest-growing product ever. A billion users is a huge milestone, but Google isn't the first AI app to hit it. OpenAI's ChatGPT […]

Other1 source
Sources & timeline
The Verge AIChatGPT and Gemini both just passed 1 billion users
OpenAI Python SDK1 source

v2.54.0

2.54.0 (2026-08-11) Features * **api:** Add new Responses model identifiers (#3595) (0652787) Bug Fixes * **api:** clarify audio upload metadata requirements (#3596) (28888f9) Chores * **api:** Update generated-file header attribution to Castiron (#3583) ([ea17fda](htt

Regulation1 source
Sources & timeline
OpenAI Python SDKv2.54.0
The Verge AI1 source

Another OpenAI executive takes off

Brad Lightcap, OpenAI's special projects lead and the company's former COO, announced his departure after an eight-year stint at the AI lab. In an internal memo he later posted to X, Lightcap told colleagues he'd be starting "something new." "Over the last few months, I've been focused on the next horizon and what would stand […]

Other1 source
Sources & timeline
The Verge AIAnother OpenAI executive takes off
llama.cpp1 source

b10361

<details open> model : fix SWA not being enabled for EXAONE 4.5 (#26848) * model : fix SWA not being enabled for EXAONE 4.5 load_arch_hparams tests `hparams.n_layer() == 64` before LLM_KV_NEXTN_PREDICT_LAYERS has been read. n_layer() returns n_layer_all - n_layer_nextn and n_layer_nextn defaults to 0, so a GGUF carrying the MTP head (block_count=65, nextn=1) evaluates to 65 and the whole SWA block is skipped. The model type switch further down in the same function reads 64, because by then the key has been loaded. n_swa is still filled in by the unconditional get_key below the block, so llama_model_n_swa() reports 4096 and the logs look correct while only swa_type stays LLAMA_SWA_TYPE_NONE.

Other1 source
Sources & timeline
llama.cppb10361
Google Gemini Blog1 source

More than 1 billion people are using the Gemini app every month.

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/1BSS.max-600x600.format-webp.webp">The Gemini app has officially surpassed 1 billion monthly users, making it the fastest-growing product in Google’s history. Here’s some data about how people are using G…

Other1 source
Sources & timeline
Google Gemini BlogMore than 1 billion people are using the Gemini app every month.
Ollama1 source

v0.32.9

NVIDIA Nemotron 3.5 Lightning NVIDIA Nemotron 3.5 Lightning is an open 30B mixture-of-experts (MoE) model with 3B active parameters built for that execution layer of always-on agents. It is designed for harnesses like OpenClaw and Hermes Agent – all supported by the NVIDIA NemoClaw open source security and management stack for running always-on AI agents. ``` ollama run nemotron-3.5-lightning ``` What's Changed * Added the Nemotron 3 architecture * Handle boundary condition in Muse Glimmer function calling parser **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.8...v0.32.9

Other1 source
Sources & timeline
Ollamav0.32.9
llama.cpp1 source

b10360

<details open> common/peg : suppress incomplete escape sequences (#26780) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b10

Other1 source
Sources & timeline
llama.cppb10360
NVIDIA Generative AI Technical Blog1 source

NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents

Long-running AI agents spend most of their time on high-volume execution: tool calls, result validation, and subagent delegation. Using a frontier reasoning...

Other1 source
Sources & timeline
NVIDIA Generative AI Technical BlogNVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents
llama.cpp1 source

b10358

<details open> Address review comment of PR 25532 (#26852) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b10358/llama-b1035

Other1 source
Sources & timeline
llama.cppb10358
llama.cpp1 source

b10357

<details open> opencl: transpose the K tile in local memory for FA prefill kernels (#26428) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/rel

Other1 source
Sources & timeline
llama.cppb10357
arXiv cs.LG1 source

Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction

This paper proposes a data-driven fire-zone segmentation method for improved short-term wildfire prediction. The authors use an unsupervised segmentation algorithm combining watershed detection and K-means clustering to define prediction units directly from historical fire patterns, instead of uniform grids. Experiments across six French departments and six forecasting models show that fire-zone segmentation consistently outperforms grid-based approaches, with mean IoU improvements of +3–6% depending on spatial scale. The method is computationally lightweight (<10s per configuration) and fully parallelizable.

Why it mattersMay add technical evidence for future radar tracking: Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction

Research1 source
Sources & timeline
arXiv cs.LGData-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction
arXiv cs.LG1 source

Evolving Safety Landscape of Multi-modal Large Language Models: A Survey of Emerging Threats and Safeguards

This public update concerns “Evolving Safety Landscape of Multi-modal Large Language Models: A Survey of Emerging Threats and Safeguards”. Open the original source for capabilities, limitations and impact.

Why it mattersThis survey provides a timely overview of the novel safety challenges emerging with multi-modal large language models (MLLMs). As MLLMs become more prevalent, understanding threats like cross-modal adversarial attacks, modality misalignment, and fused safety risks is critical for developing robust AI systems. It consolidates recent advances in safety strategies and outlines open problems, guiding future research directions in AI safety and responsible development of multimodal systems.

Safety1 source
Sources & timeline
arXiv cs.LGEvolving Safety Landscape of Multi-modal Large Language Models: A Survey of Emerging Threats and Safeguards
arXiv cs.CV1 source

PragyaDoc: A Universal Document Intelligence Framework for Multilingual Medical Document Understanding in Low-Resource Settings

This paper introduces PragyaDoc, a universal document intelligence framework for multilingual medical document understanding in low-resource settings. It addresses India's language accessibility barrier, where most medical documentation is in English but rural populations, ASHA workers, and patient families are functionally excluded from understanding it. The framework uses a four-layer pipeline: parallel ensemble OCR extraction, geometric-lexical fusion, deterministic domain structuring, and dual-LLM medical reasoning and localization.

Why it mattersAddresses critical language accessibility barrier in Indian healthcare, where most medical documents are in English but needed by non-English-speaking rural populations and health workers. Potential to significantly improve health information access in low-resource multilingual settings.

Research1 source
Sources & timeline
arXiv cs.CVPragyaDoc: A Universal Document Intelligence Framework for Multilingual Medical Document Understanding in Low-Resource Settings
arXiv cs.CL1 source

Unified Hallucination Fuzzing for Multimodal Large Language Models

Public information on “Unified Hallucination Fuzzing for Multimodal Large Language Models”: Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage…

Why it mattersAddresses a critical reliability issue in MLLMs by advancing from static benchmarks to dynamic, evolving evaluation, with open-source code.

Benchmarks1 source
Sources & timeline
arXiv cs.CLUnified Hallucination Fuzzing for Multimodal Large Language Models
arXiv cs.AI1 source

Determinization in Structure Theories: A Unified Framework via Closure, Comparability, and Joint Admissibility

The paper develops a formal framework for constructing canonical interpretations from plural structure theories. A structure theory is a triple (signature, axioms, inference policy) with an admissible interpretation family of globally consistent assignments. Three canonicalization levels are distinguished: closure stabilization, global completion, and determinization. Non-determinism is classified into epistemic plurality (Type E) and structural plurality (Type S), with Type S-strong lacking common upper bounds. Two mechanisms arise: operator-based completion and selector-based construction. Sufficient structural conditions are provided, and pure inference-based completion reduces to a saturated closure operator under positive, non-retractive rules with soundness. For Type E, closure stabilization is shown, while determinization depends on an open global confluence property. For Type S-strong, determinization is achieved via canonical selection. Multi-level canonicalization is structurally non-commutative via staged operators, with a conditional classification theorem. The framework applies to LLM-assisted reasoning, viewing hallucination as unsupported canonicalization.

Why it mattersProposes a formal framework to canonize plural structure theories, offering insights into LLM hallucinations as unsupported canonicalization, which could improve reasoning reliability.

Research1 source
Sources & timeline
arXiv cs.AIDeterminization in Structure Theories: A Unified Framework via Closure, Comparability, and Joint Admissibility
arXiv cs.CV1 source

NeuroPilot: An Agent-Driven Smart Pipeline for Processing, Quality Control, and Managing Neuroimages

NeuroPilot is an LLM-driven multi-agent pipeline for neuroimage processing, quality control, and data management. It encapsulates expertise into three LLM-invocable skills: dcm2bids-skill, neuroimage-pre-skill, and qc-agent-skill. Deployed across 17 cohorts (>123,000 subjects) spanning infant to aging populations and multiple MRI modalities, the system demonstrates generalizability. The QC agent screened 558 production subjects, validating automated flags against FreeSurfer topology-defect metrics. The infant processing pipeline achieved 100% (201/201) completion on QC-validated inputs, compressing the traditional 2–3 month timeline for training and processing into one week.

Why it mattersMay add technical evidence for future radar tracking: NeuroPilot: An Agent-Driven Smart Pipeline for Processing, Quality Control, and Managing Neuroimages

Agents1 source
Sources & timeline
arXiv cs.CVNeuroPilot: An Agent-Driven Smart Pipeline for Processing, Quality Control, and Managing Neuroimages
arXiv cs.AI1 source

Towards an Argumentative Foundation for Evaluative AI

This position paper advocates computational argumentation as a formal, computable foundation for Evaluative AI (EAI), which supports human decision-making by presenting competing hypotheses with supporting and opposing evidence rather than a single recommendation. It sets the groundwork for a long-term research agenda toward distributed, human-centred EAI systems that are explainable and contestable.

Why it mattersMay add technical evidence for future radar tracking: Towards an Argumentative Foundation for Evaluative AI

Research1 source
Sources & timeline
arXiv cs.AITowards an Argumentative Foundation for Evaluative AI
arXiv cs.LG1 source

Application of Artificial Intelligence for Fraudulent Banking Operations Recognition

This paper studies the application of AI to recognize fraudulent banking transactions. Due to the COVID-19 pandemic, bank fraud has increased as operations moved online. The work focuses on machine learning models and data preprocessing techniques, addressing imbalanced datasets, feature transformation, and engineering. An artificial neural network model improves detection accuracy; among visualized results, logistic regression achieves an AUC of approximately 0.946, while stacked generalization reaches a higher AUC of 0.954. The recognition of banking fraud via AI is considered a topical issue in digital society.

Why it mattersDemonstrates applied machine learning for banking fraud detection amid increased online activity, providing practical preprocessing techniques and competitive AUC scores (~0.95). Timely for digital security.

Regulation1 source
Sources & timeline
arXiv cs.LGApplication of Artificial Intelligence for Fraudulent Banking Operations Recognition
arXiv cs.CL1 source

DocAtlas: Long-Document Understanding as Mutable-State Interaction

DocAtlas: Long-Document Understanding as Mutable-State Interaction. The system presents a mutable document harness that controls searching, reading, noting, and reviewing document information via an external environment, combining self-improving retrieval, selective evidence access, and active working memory under a fixed context budget. With GPT-5.4, DocAtlas achieves 71.4% on MMLongBench-Doc, exceeding the human-expert reference of 65.8%; a Qwen3.5-4B VLM trained with end-to-end RL reaches 63.7%, compared with a 54.4% direct-input baseline.

Why it mattersDocAtlas achieves state-of-the-art on long-document understanding, surpassing human expert performance on MMLongBench-Doc, and demonstrates that mutable-state interaction can drastically improve compact document agents via end-to-end RL.

Agents1 source
Sources & timeline
arXiv cs.CLDocAtlas: Long-Document Understanding as Mutable-State Interaction
arXiv cs.CL1 source

WuYuEval: A Multi-Level Benchmark for Large Language Models in Solid Waste Management

WuYuEval introduces a multi-level benchmark for evaluating LLMs in solid waste management (SWM), with a Foundation Module of 4,590 MCQ questions and an Expert Module of 247 open-ended scenario questions. Evaluation of 33 LLMs shows wide performance variation; reasoning-oriented thinking modes help only under certain conditions, providing empirical evidence for SWM-oriented foundation models.

Why it mattersIntroduces a specialized benchmark for evaluating LLMs in solid waste management, an underexplored application domain. Highlights limitations of current LLMs in professional decision-making under real-world constraints, guiding development of domain-specific foundation models.

Benchmarks1 source
Sources & timeline
arXiv cs.CLWuYuEval: A Multi-Level Benchmark for Large Language Models in Solid Waste Management
arXiv cs.AI1 source

Flow-by-Flow:Content-Judgment Bypass for Governing AI Output in High-Loss Domains

This paper proposes "Flow-by-Flow," a governance paradigm that controls supervisory load without evaluating content. The authors argue that human oversight in high-loss domains is constrained by V×L (output velocity × cognitive load), where triage, judgment, and response costs respond asymmetrically. They introduce a cognitive cost score based on formal, countable features and an institutional capacity cap, and derive four design invariants. A Monte Carlo analysis across 1,000 draws suggests composite multi-metric flow control outperforms supervision reinforcement alone in 90.8% of trials.

Why it mattersAddresses the scaling limits of human-in-the-loop oversight in high-stakes AI, proposing a content-judgment bypass mechanism that could enable safer AI governance without overwhelming human supervisors.

Research1 source
Sources & timeline
arXiv cs.AIFlow-by-Flow:Content-Judgment Bypass for Governing AI Output in High-Loss Domains
arXiv cs.CV1 source

Performance of large language models in the optical diagnosis of colorectal polyps

This public research item examines “Performance of large language models in the optical diagnosis of colorectal polyps”. Open the original paper to review its methods and conclusions.

Why it mattersThis study benchmarks state-of-the-art multimodal LLMs on real-world colonoscopy images, showing high accuracy for binary polyp classification but insufficient sensitivity/specificity for clinical standards, highlighting the need for human-in-the-loop workflows before deployment.

Research1 source
Sources & timeline
arXiv cs.CVPerformance of large language models in the optical diagnosis of colorectal polyps
llama.cpp1 source

b10356

<details open> ci : target ROCm 7.14 for build and release (#25775) * Switch ROCm from 7.2.1 to 7.14 ROCm 7.14 is the first production release using TheRock build system. It can be installed using multi-arch deliverables from wheels, debs, rpms, tarballs or runfiles. Adjust ROCm targets for Linux and Windows to use this instead. * ci: switch all other Windows ROCm jobs to ROCm 7.14 wheels Move the shared windows-setup-rocm composite action from the HIP SDK PRO Edition installer to the multi-arch ROCm wheels (rocm[libraries,devel]). The wheel-install logic that previously lived inline in release.yml is now in the shared action, and both build-cache.yml and release.yml call it. Also migrate th

Other1 source
Sources & timeline
llama.cppb10356
TechCrunch AI1 source

OpenAI reportedly completed a $7 billion employee tender offer

OpenAI reportedly completed a $7 billion employee tender offer.

Why it mattersA $7 billion tender offer provides liquidity to OpenAI employees, signaling strong investor confidence and a high valuation. It indicates the company's financial maneuvering and can impact talent retention.

Business1 source
Sources & timeline
TechCrunch AIOpenAI reportedly completed a $7 billion employee tender offer

Mon, Aug 10

TechCrunch AI1 source

As AI-led attacks multiply, OpenAI launches a new cyber model

As AI-led attacks increase, OpenAI launches a new cyber-trained AI model and expands its AI cybersecurity defense program Daybreak.

Why it mattersOpenAI's launch of a cyber-trained model and expansion of its Daybreak defense program directly responds to the escalating threat of AI-driven attacks, marking a significant step in AI safety and cybersecurity.

Safety1 source
Sources & timeline
TechCrunch AIAs AI-led attacks multiply, OpenAI launches a new cyber model
Ollama1 source

v0.32.8

Muse Glimmer Muse Glimmer is now available on all platforms. Muse Glimmer can power coding agent applications such as Claude Code, Codex, Pi and more, as well as long-running personal assistants such as OpenClaw and Hermes. Ollama's MLX engine provides state-of-the-art performance on Apple Silicon for this model, with support for DFlash and image input as of Ollama 0.32.7. To download and run Muse Glimmer locally: ```bash ollama run muse-glimmer ``` To run Muse Glimmer with Claude Code, download Ollama and run: ```bash ollama launch claude --model muse-glimmer ``` For a lighter-weight coding agent, try Pi: ```bash ollama launch pi --model m

Other1 source
Sources & timeline
Ollamav0.32.8
The Verge AI1 source

Mark Zuckerberg doesn’t understand how to live

The Verge criticizes Mark Zuckerberg's understanding of life in this commentary. It opens with an anecdote about a rock climber who used AI to create a motivational poster featuring a bear on a slackline holding a sign saying 'Do cool shit,' which sets up a contrast with Zuckerberg's AI vision.

Why it mattersThis article provides cultural critique of Mark Zuckerberg's AI vision, reflecting skepticism in tech media about his understanding of life and technology's role.

Opinion1 source
Sources & timeline
The Verge AIMark Zuckerberg doesn’t understand how to live
vLLM1 source

v0.27.0

vLLM v0.27.0 released with 561 commits from 242 contributors (64 new). Highlights: full-stack support for Kimi K3 (core model files/kernels, Python/Rust frontends, AttnRes kernels, DeepGEMM, compressed-tensors quantized checkpoints, DSpark AR fusion, optional shared-expert sharding); new models including Qwen3.5 text-only dense and MoE models (with EVS video token pruning), K-EXAONE-2.0-750B-A37B, VaultGemma, and jina-embeddings-v5-text-nano; PyTorch 2.13.0 upgrade (with torchvision 0.28.0 and Triton 3.7.1) as a breaking environment change (XPU and CPU also updated); deeper FlashAttention 4 integration on SM100 with FP8 KV cache and headdim-256 support, plus new JIT warmup infrastructure to eliminate first-request compilation stalls; DeepSeek-V4 performance improvements are mentioned but details are truncated.

Why it mattersvLLM v0.27.0 adds support for several new models including Kimi K3, Qwen3.5, and K-EXAONE-2.0-750B-A37B, upgrades to PyTorch 2.13.0 (breaking change), deepens FlashAttention 4 integration with FP8 KV cache and headdim-256, and includes DeepSeek-V4 performance improvements. It also introduces JIT warmup infrastructure to eliminate first-request compilation stalls. This release reinforces vLLM as a high-performance inference engine and benefits production deployments of large language models.

Infrastructure1 source
Sources & timeline
vLLMv0.27.0
TechCrunch AI1 source

Mark Zuckerberg’s AI manifesto is exactly why people don’t like AI

Mark Zuckerberg published a 6,500-word manifesto on personal AI, focusing on Meta AI's 'personal superintelligence' systems; the TechCrunch headline suggests this is exactly why people dislike AI.

Why it mattersMark Zuckerberg's detailed vision for personal AI superintelligence signals Meta's strategic direction and may influence public reception of AI technologies.

Opinion1 source
Sources & timeline
TechCrunch AIMark Zuckerberg’s AI manifesto is exactly why people don’t like AI
MIT Technology Review AI1 source

AI professors are negotiating the new realities of academic research

An MIT Technology Review AI story reports that AI professors are negotiating new realities in academic research. The article begins with the author attending a gathering of AI researchers in Mountain View, California. The full text is currently truncated.

Why it mattersAddresses the evolving challenges and negotiations in AI academic research, which could affect research directions, funding, and talent flow. However, the article's full content is not available here, making it hard to discern specific insights.

Research1 source
Sources & timeline
MIT Technology Review AIAI professors are negotiating the new realities of academic research
Anthropic Research1 source

Learning more about Claude's mathematical capabilities

Anthropic Research published a blog post on 2026-08-10 stating that an unreleased version of Claude improved the lower bound for the proportion of zeros of the Riemann zeta function satisfying the Riemann hypothesis, increasing it from 41.6% to 67.2%. The post says Anthropic mathematicians studied and validated Claude's paper, Claude produced a formally verifiable proof, and external experts Brian Conrey and Dan Goldston reviewed the work.

Why it mattersMay add technical evidence for future radar tracking: Learning more about Claude's mathematical capabilities

Research1 source
Sources & timeline
Anthropic ResearchLearning more about Claude's mathematical capabilities
Hugging Face Blog1 source

Build Low-Latency Multilingual Voice Agents: Open Weights & Full Deployment Control with NVIDIA Magpie TTS

NVIDIA Magpie TTS offers open weights and full deployment control to build low-latency multilingual voice agents, as announced on the Hugging Face Blog.

Why it mattersNVIDIA releasing an open-weights, low-latency, multilingual TTS model could accelerate development of customizable voice agents, reduce cloud dependency, and enable on-device or private deployment.

Open source1 source
Sources & timeline
Hugging Face BlogBuild Low-Latency Multilingual Voice Agents: Open Weights & Full Deployment Control with NVIDIA Magpie TTS
The Verge AI1 source

Four takeaways from Mark Zuckerberg’s massive AI manifesto

This The Verge article discusses four key takeaways from a lengthy AI manifesto published by Mark Zuckerberg, titled "The Future is for Everyone." The manifesto, over 6,500 words, presents his vision for humanity coexisting with artificial intelligence. The specific takeaways are not detailed in the available text.

Why it mattersMark Zuckerberg's AI manifesto outlines his vision for humanity's coexistence with AI, which may influence Meta's long-term AI strategy and shape public discourse on AI governance and ethics.

Opinion1 source
Sources & timeline
The Verge AIFour takeaways from Mark Zuckerberg’s massive AI manifesto
OpenAI News1 source

OpenAI’s letter to Governor Abbott on responsible AI infrastructure in Texas

OpenAI sent a letter to Governor Greg Abbott expressing commitment to responsible AI infrastructure in Texas, supporting reliable and transparent growth that benefits Texans.

Why it mattersOpenAI's direct outreach to a state governor on AI infrastructure signals strategic engagement with regional policy, potentially influencing responsible AI deployment and economic benefits for Texas.

Business1 source
Sources & timeline
OpenAI NewsOpenAI’s letter to Governor Abbott on responsible AI infrastructure in Texas
OpenAI News1 source

Model ML completes finance work more efficiently with GPT-5.6 Sol

Model ML uses GPT-5.6 Sol to complete finance work more efficiently, from research and analysis to editable, traceable PowerPoint decks and Excel workbooks.

Why it mattersDemonstrates practical application of OpenAI's models in finance, potentially indicating improved capabilities for complex workflows.

Model release1 source
Sources & timeline
OpenAI NewsModel ML completes finance work more efficiently with GPT-5.6 Sol
Ars Technica AI1 source

Peer review is overwhelmed—can it survive in the AI era?

As research and AI-assisted papers surge, volunteer reviewers struggle to keep up, raising questions about whether peer review can survive in the AI era.

Why it mattersMay add technical evidence for future radar tracking: Peer review is overwhelmed—can it survive in the AI era?

Infrastructure1 source
Sources & timeline
Ars Technica AIPeer review is overwhelmed—can it survive in the AI era?
Ollama1 source

v0.32.7

Muse Glimmer > Note: Muse Glimmer is currently available via initial support via Ollama's MLX engine on Apple Silicon. Additional support and optimizations for Apple Silicon, NVIDIA, AMD, and other platforms will be available in the coming days. **Muse Glimmer**, Meta's newest open model and the first released by Meta Superintelligence Labs, is now available on Ollama. It's a 30B multimodal model purpose-built for agent workloads that run locally. With Ollama, you can now use Muse Glimmer to power coding agent applications such as Claude Code, Codex, Pi and more, as well as long-running personal assistants such as OpenClaw and Hermes. Ollama's MLX engine provides state-of-the-art performa

Other1 source
Sources & timeline
Ollamav0.32.7
Hugging Face Transformers1 source

Release: v5.15.0

Hugging Face Transformers releases v5.15.0, adding support for Meta Muse Glimmer, a 30B-parameter multimodal model distilled from Muse for agentic use cases, released under Apache 2.0, with a 2B ViT vision encoder and 28B text decoder for local privacy-aware deployment, and for GraniteMoeSWA/GraniteSWA models.

Open source1 source
Sources & timeline
Hugging Face TransformersRelease: v5.15.0
MIT Technology Review AI1 source

These startups are chasing the next big thing in LLMs

MIT Technology Review's What's Next series reports on startups chasing the next big thing in LLMs. The article looks back at the 2017 Google paper 'Attention Is All You Need' and explores the entrepreneurial directions that have emerged since.

Why it mattersThe article highlights startups building the next generation of LLMs, which could influence future AI capabilities and market dynamics. It reflects ongoing innovation post-Transformer architecture.

Business1 source
Sources & timeline
MIT Technology Review AIThese startups are chasing the next big thing in LLMs
MIT Technology Review AI1 source

AI for science needs reasoning, not just data

The article 'AI for science needs reasoning, not just data' opens by recalling historical predictions of the end of science, such as Michelson's 1903 claim that all physical facts were discovered and Hawking's 1980s forecast of theoretical physics' completion. It then suggests that AI's advancement necessitates reasoning, not just data, for scientific progress.

Why it mattersHighlights the need for reasoning capabilities in AI systems applied to scientific research, beyond just processing large datasets, which could influence the development of more effective scientific AI agents.

Opinion1 source
Sources & timeline
MIT Technology Review AIAI for science needs reasoning, not just data
arXiv cs.LG1 source

Sharding Prevents LLM Oversight Failures and Adversarial Exploitation

The paper 'Sharding Prevents LLM Oversight Failures and Adversarial Exploitation' shows that giving an LLM judge more compute does not necessarily improve oversight: as the number of verdicts per call grows, agreement with expert raters declines. Sharding—partitioning requirements into smaller groups, assigning each to a separate call, and aggregating results—mitigates this failure, improving agreement under a fixed budget. A sharded weaker judge can outperform a more capable holistic judge. Sharding also reduces over-acceptance under presentation-only adversarial attacks, but not per-criterion persuasion; adding debate-style opposition withstands such adaptive attacks.

Safety1 source
Sources & timeline
arXiv cs.LGSharding Prevents LLM Oversight Failures and Adversarial Exploitation
arXiv cs.CL1 source

TEXAS: Task-Expert-Aware Supervision for Downstream Mixture-of-Experts LLM Adaptation

The paper introduces TEXAS, a supervision method for adapting Mixture-of-Experts (MoE) language models to downstream tasks. It identifies task-relevant experts by comparing activations on successful vs. failed instances, then upweights answer tokens in failed instances that activate these experts during fine-tuning. Evaluations across three MoE models and six benchmarks show best or tied-best performance in 17 of 18 settings, improving over the strongest baseline by 1.3-1.5 points on average.

Why it mattersMay add technical evidence for future radar tracking: TEXAS: Task-Expert-Aware Supervision for Downstream Mixture-of-Experts LLM Adaptation

Research1 source
Sources & timeline
arXiv cs.CLTEXAS: Task-Expert-Aware Supervision for Downstream Mixture-of-Experts LLM Adaptation
arXiv cs.CL1 source

Separating Decision-Rule Misalignment from Readout-Coverage Limitations in Speech Language Models

The paper introduces a generation-aligned diagnostic ladder to separate failures in speech language models on paralinguistic tasks into endpoint, decision-rule, and readout-coverage gaps. Across five systems and two emotion corpora, state decoding exceeds generation by 27.8 accuracy points on average; both gaps are positive in all ten conditions. A label-free logit correction improves generated accuracy in every condition, and emotion information outside the native readout generalizes to held-out speakers but replacing readout-external directions has little effect on emitted answers.

Why it mattersThe paper introduces a diagnostic framework to disentangle error sources in speech language models, enabling targeted improvements in paralinguistic tasks. By separating decision-rule misalignment from readout-coverage gaps, it provides actionable insights for model optimization.

Research1 source
Sources & timeline
arXiv cs.CLSeparating Decision-Rule Misalignment from Readout-Coverage Limitations in Speech Language Models
arXiv cs.AI1 source

Towards Multi-Label Graph Foundation Models: from Single-Vector Representation Learning to Multi-Semantic Basis Learning

This paper proposes a Multi-Semantic Basis Graph Foundation Model (MSB-GFM) for cross-domain multi-label node classification. It addresses the single-label assumption of existing graph foundation models that causes semantic entanglement. The model introduces multi-semantic basis representation learning, modeling each node as an adaptive composition of semantic bases, and uses a semantic-structure dual-channel architecture with domain adversarial training for effective knowledge transfer. Experiments demonstrate its effectiveness.

Why it mattersMay add technical evidence for future radar tracking: Towards Multi-Label Graph Foundation Models: from Single-Vector Representation Learning to Multi-Semantic Basis Learning

Research1 source
Sources & timeline
arXiv cs.AITowards Multi-Label Graph Foundation Models: from Single-Vector Representation Learning to Multi-Semantic Basis Learning
arXiv cs.CV1 source

TransSLR: A Lightweight Transformer for Sign Language Recognition

This paper introduces TransSLR, a lightweight Temporal Transformer Encoder for sign language recognition, focusing on Central African Sign Language (CASL). Trained from scratch on 64-frame normalized pose sequences, it achieves state-of-the-art accuracy of 80.39% on the CASL-W60 benchmark, surpassing the previous best by +10.46 points. The encoder-only design reduces computational overhead, enabling deployment in resource-constrained environments.

Why it mattersMay add technical evidence for future radar tracking: TransSLR: A Lightweight Transformer for Sign Language Recognition

Research1 source
Sources & timeline
arXiv cs.CVTransSLR: A Lightweight Transformer for Sign Language Recognition
arXiv cs.AI1 source

Beyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast

Public information on “Beyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast”: Reward models are central to learning from human preferences, yet identifying what drives their predictions remains challenging. Recent sparse Mixture-of-Experts (MoE) reward models seek to improve interpretability by routing prompts to specialized experts and…

Why it mattersImproves interpretability of reward models, crucial for AI alignment and safety, especially as Mixture-of-Experts architectures become more common.

Research1 source
Sources & timeline
arXiv cs.AIBeyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast
arXiv cs.AI1 source

EntropyMoE: Entropy-Aware Sparse Expert Routing for Tokenizer-Free LLMs

Public information on “EntropyMoE: Entropy-Aware Sparse Expert Routing for Tokenizer-Free LLMs”: Recent byte-level large language models (LLMs) have made tokenizer-free modeling increasingly competitive by grouping bytes into dynamically sized patches. However, existing byte-patch architectures still apply the same dense feed-forward computation to every…

Why it mattersEntropyMoE introduces entropy-based routing for mixture-of-experts in tokenizer-free LLMs, enabling adaptive computation for dynamic byte patches. This could improve efficiency and scalability of byte-level models, extending MoE beyond tokenized representations.

Research1 source
Sources & timeline
arXiv cs.AIEntropyMoE: Entropy-Aware Sparse Expert Routing for Tokenizer-Free LLMs
arXiv cs.CL1 source

NTDH: Complex Reasoning for Comprehensive Affective Analysis

This paper introduces NTDH, a method that recasts comprehensive affective analysis as a complex-reasoning problem, addressing heterogeneous label spaces and contextual reconciliation. It proposes components for naturalization, tolerance-aware verification, domain-aware strategies, and directional hints, training Qwen3-8B with SFT and GRPO. The method achieves strong results, including a Pearson correlation of 0.862 on the EI-reg task, using about 14x fewer training records than comparable systems.

Why it mattersDemonstrates how to reframe heterogeneous affective analysis tasks as a single complex-reasoning problem, enabling efficient LLM fine-tuning with GRPO and achieving strong results using 14x fewer training records.

Research1 source
Sources & timeline
arXiv cs.CLNTDH: Complex Reasoning for Comprehensive Affective Analysis
arXiv cs.LG1 source

Latent Fact-Checking: Detecting Misinformation through refresh Engineering

Public information on “Latent Fact-Checking: Detecting Misinformation through refresh Engineering”: The proliferation of misinformation online has driven demand for scalable detection systems. While most existing approaches rely on surface-level linguistic features or external knowledge retrieval, we examine truthfulness as a geometric property of a language…

Why it mattersDemonstrates that truthfulness is a linearly separable feature in language model representations, enabling lightweight, internal fact-checking without external retrieval or fine-tuning. This could complement retrieval-based pipelines and improve scalability of misinformation detection, especially for smaller models.

Research1 source
Sources & timeline
arXiv cs.LGLatent Fact-Checking: Detecting Misinformation through refresh Engineering
arXiv cs.CV1 source

UAV3DCrop: Benchmarking 3D Reconstruction in Repeated Multi-Angle UAV Crop Surveys

The paper presents UAV3DCrop, a public benchmark for 3D reconstruction in repeated multi-angle UAV crop surveys. It contains 88,830 RGB images at 5280×3956 pixels (GSD 3.6–5.8 mm) from 91 scenes of corn, soybean, wheat, and oat. Track A evaluates seven scene-optimized NeRF/3DGS variants on held-out views, photogrammetry-referenced depth, and canopy-height recovery; Track B tests four pretrained feed-forward models on zero-shot camera-pose and geometry estimation. Results show that no single method leads across all targets: Splatfacto-big leads appearance, Scaffold-GS leads depth, and MapAnything leads most feed-forward metrics. The authors conclude that current 3D reconstruction methods are not yet interchangeable for agronomic use. The dataset is publicly available.

Why it mattersProvides a public benchmark that reveals current 3D reconstruction methods are not interchangeable for agricultural monitoring, with implications for deploying NeRF/3DGS in precision agriculture.

Benchmarks1 source
Sources & timeline
arXiv cs.CVUAV3DCrop: Benchmarking 3D Reconstruction in Repeated Multi-Angle UAV Crop Surveys
arXiv cs.CV1 source

Deep Evidential Regression for Sparse Forest Height Estimation from Multimodal Satellite Imagery

This paper investigates Deep Evidential Regression for sparse forest height estimation from multimodal satellite imagery, using the TreeUQ benchmark with Sentinel-1/2 data. A masked evidential loss is introduced to handle label sparsity, achieving comparable performance to a deterministic U-Net while additionally providing well-calibrated uncertainty estimates.

Why it mattersEnables efficient uncertainty-aware forest height estimation from satellite imagery, crucial for reliable carbon accounting and ecosystem monitoring, especially with sparse ground-truth labels.

Research1 source
Sources & timeline
arXiv cs.CVDeep Evidential Regression for Sparse Forest Height Estimation from Multimodal Satellite Imagery
arXiv cs.LG1 source

Risk-Aware Decision Policies for Agents Under Noisy Perception

This paper from arXiv cs.LG presents an Artificial Life predator-prey model of foraging under noisy perception, comparing agent performance under policies that account for noisy predictions. Results show that blindly trusting perceptual labels leads to catastrophic failure, while uncertainty-aware strategies significantly improve survival and reduce fatal errors, especially under asymmetric noise. Qualitative regime shifts from exploratory to conservative behavior are observed as uncertainty increases. The work links risk-sensitive foraging, ecological information use, and robust learning with noisy labels.

Why it mattersShows that ignoring perceptual noise leads to catastrophic failure in AI agents, and uncertainty-aware policies significantly improve survival and reduce fatal errors, providing an interpretable analogue to robust learning with noisy labels.

Research1 source
Sources & timeline
arXiv cs.LGRisk-Aware Decision Policies for Agents Under Noisy Perception

Sun, Aug 9

llama.cpp1 source

b10333

Public information on “b10333”: <details open> ggml-cpu : fix missing Q5_0 dispatch in SpaceMiT backend (#26792) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFra…

Open source1 source
Sources & timeline
llama.cppb10333
llama.cpp1 source

b10332

Public information on “b10332”: <details open> ci: rm `GGML_HIP_ROCWMMA_FATTN` (#26760) Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS In…

Open source1 source
Sources & timeline
llama.cppb10332

Sat, Aug 8

The Verge AI1 source

An Amazon data center could have the worst polluting power plant in the country

Public information on “An Amazon data center could have the worst polluting power plant in the country”: To power its new West Texas data center, Amazon is investing in the construction of a new power plant that could be one of the largest single producers of greenhouse gases in the US, according to the New York Times. The new gas-burning plant in Pecos County, T…

Funding1 source
Sources & timeline
The Verge AIAn Amazon data center could have the worst polluting power plant in the country
SGLang1 source

v0.5.17

SGLang v0.5.17 release highlights: 582 PRs from 194 contributors, Day-0 support for Kimi K3 (2.8T multimodal LatentMoE with native MXFP4, 1M context, etc.) and Day-0 support for MiniMax-H3 (video generation model).

Why it mattersSGLang v0.5.17 provides day-0 support for two major frontier models: Kimi K3 (2.8T-parameter multimodal MoE with 1M context) and MiniMax-H3 (video generation). It introduces advanced serving optimizations (MXFP4 quantization, KDA-aware caching, speculative decoding) verified on NVIDIA GB300 and AMD MI35x. This release significantly advances open-source inference infrastructure, enabling efficient and scalable deployment of state-of-the-art models.

Infrastructure1 source
Sources & timeline
SGLangv0.5.17

Fri, Aug 7

llama.cpp1 source

b10326

Public information on “b10326”: <details open> tts: account for the vocoder pass in the timings line (#26733) get_output runs the waveform work the pipeline defers to it, from a single trailing window to a full pass depending on the model. Measuring it keeps the reported total and the audio…

Open source1 source
Sources & timeline
llama.cppb10326
Turing Post1 source

AI Agent Memory Explained: From SOAR to LLM Memory Mode

An article from Turing Post explaining how AI agents store and retrieve memory, covering semantic, episodic, and procedural memory types, the legacy of SOAR, generative agents, and how ChatGPT's memory mode works.

Why it mattersMay add technical evidence for future radar tracking: AI Agent Memory Explained: From SOAR to LLM Memory Mode

Agents1 source
Sources & timeline
Turing PostAI Agent Memory Explained: From SOAR to LLM Memory Mode
llama.cpp1 source

b10322

Public information on “b10322”: <details open> sycl: coalesce the ssm_conv window loads (#26612) test-backend-ops perf -o SSM_CONV on an Arc Pro B70, interleaved A/B against master, 6 reps, us/run: ne_a=[515,3328,1,1] ne_b=[4,3328,1,1] n_t=512 97.68 -> 52.95 1.85x ne_a=[937,8192,1,1] ne_b=[4…

Open source1 source
Sources & timeline
llama.cppb10322
llama.cpp1 source

b10321

Public information on “b10321”: <details open> metal : fix NORM/RMS_NORM for row lengths that leave a partial simdgroup (#26708) ggml_metal_op_norm sized the threadgroup with `nth = std::min(nth, args.ne00_t)`, which can leave nth not a multiple of the simdgroup size. The kernels finish thei…

Open source1 source
Sources & timeline
llama.cppb10321
Anthropic Python SDK1 source

v0.121.0

Anthropic Python SDK v0.121.0 release adds mid-conversation-tool-changes beta, session budgets, advisor tool, pinned inference location, and skills auto-loading from GitHub; removes retired Claude Opus 4.1 models; updates docs and dependency constraints.

Product update1 source
Sources & timeline
Anthropic Python SDKv0.121.0
Google Gemini Blog1 source

See what 5 builders are making with Gemini Omni

Google Gemini Blog highlights five builders using Gemini Omni to make creating videos as easy as having a conversation, including editing videos and visualizing ideas.

Why it mattersShowcases real-world use cases of Gemini Omni for video creation and editing, indicating progress in making AI video tools accessible through conversational interfaces.

Product update1 source
Sources & timeline
Google Gemini BlogSee what 5 builders are making with Gemini Omni
OGX1 source

v1.3.0

OGX v1.3.0 is a GitHub release published on 2026-08-07. Key changes include: adding a Meta AI remote inference provider; injecting network/TLS config into Meta provider HTTP clients; extracting a shared SSE passthrough stream helper for Anthropic-compatible providers; fixing a duplicate /v1 prefix in vLLM Anthropic message URLs; adding anthropic_count_tokens to OpenAIMixin via translation; fixing CI for SDK Makefile and client generation; and fixing container entrypoint boot.

Product update1 source
Sources & timeline
OGXv1.3.0
arXiv cs.AI1 source

Woodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong Models

Public information on “Woodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong Models”: Large language models often fail on reasoning tasks despite possessing the capability to solve them. We argue that many such failures arise from localized reasoning bugs in intermediate steps rather than from global incompetence. We show that these bugs are fr…

Benchmarks1 source
Sources & timeline
arXiv cs.AIWoodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong Models
arXiv cs.AI1 source

Agentic Nesting: A New Methodology for Existing Enterprise Application Integration and Services

Public information on “Agentic Nesting: A New Methodology for Existing Enterprise Application Integration and Services”: Enterprise operations extensively rely on multiple heterogeneous business systems and information applications, which also result in severe data silos and process fragmentation. Enterprises have invested considerable financial and material resources in buildin…

Agents1 source
Sources & timeline
arXiv cs.AIAgentic Nesting: A New Methodology for Existing Enterprise Application Integration and Services
arXiv cs.CL1 source

Mean-Field Dynamics of Chain-of-Thought Reasoning in Large Language Models

Public information on “Mean-Field Dynamics of Chain-of-Thought Reasoning in Large Language Models”: Large language models (LLMs) with chain-of-thought reasoning have been widely applied in recent years, and theoretical explanations of their behavior may help deepen our understanding and guide model optimization. In this study, we introduce a framework that s…

Research1 source
Sources & timeline
arXiv cs.CLMean-Field Dynamics of Chain-of-Thought Reasoning in Large Language Models
arXiv cs.AI1 source

The Ignition Index: Measuring Global Workspace Dynamics in Language Models

Public information on “The Ignition Index: Measuring Global Workspace Dynamics in Language Models”: We introduce the Ignition Index (I), a validated scalar metric that operationalizes Global Workspace Theory's (GWT) all-or-none ignition prediction in transformer language models. The metric fits a four-parameter sigmoid to per-layer linear probe accuracy as a…

Research1 source
Sources & timeline
arXiv cs.AIThe Ignition Index: Measuring Global Workspace Dynamics in Language Models
arXiv cs.CL1 source

Universal Pathologies, Conditional Consequences: A Triple-Robustness Analysis of RAG for Multi-Hop Traceability

Public information on “Universal Pathologies, Conditional Consequences: A Triple-Robustness Analysis of RAG for Multi-Hop Traceability”: GraphRAG underperforms vector RAG on citation precision in many reports, but where and why have remained corpus-bound. We present a triple-robustness analysis that holds the retrieval architecture fixed and varies three orthogonal axes embedder (local e5-small…

Research1 source
Sources & timeline
arXiv cs.CLUniversal Pathologies, Conditional Consequences: A Triple-Robustness Analysis of RAG for Multi-Hop Traceability
arXiv cs.CL1 source

Simulator-Grounded Large Language Models for Industrial Causal Reasoning: Tool-Use, Structured Injection, and Plant-Portable Retrieval for Wastewater Treatment Decision Support

Public information on “Simulator-Grounded Large Language Models for Industrial Causal Reasoning: Tool-Use, Structured Injection, and Plant-Portable Retrieval for Wastewater Treatment Decision Support”: Wastewater operators need answers grounded in how their plant's variables interact and how fast effects propagate, not in generic pretraining text, when asking causal questions such as "why is N2O rising?" or "what happens if I cut aeration by 20%?". We compar…

Benchmarks1 source
Sources & timeline
arXiv cs.CLSimulator-Grounded Large Language Models for Industrial Causal Reasoning: Tool-Use, Structured Injection, and Plant-Portable Retrieval for Wastewater Treatment Decision Support
Anthropic News1 source

Improving Fable 5 Safeguards

Anthropic announces updates to Claude Fable 5's biology safeguards that substantially reduce fallbacks. In testing, biology-related fallbacks decreased by about 85%, allowing Fable 5 to handle a wider range of biology tasks such as interpreting lab results, understanding symptoms, and educational biology questions. However, for dual-use requests (virology, toxicology, molecular design), Fable 5 still falls back to Opus 5, so it is not yet usable for professional biology research or drug development. The company is working to close this gap through trusted access pathways while managing risks from advanced capabilities.

Safety1 source
Sources & timeline
Anthropic NewsImproving Fable 5 Safeguards

Thu, Aug 6

Google Gemini Blog1 source

How Gemini plans such detailed vacation itineraries for you

Google Gemini blog post explains how Gemini uses personalized recommendations to create custom travel itineraries when you ask it to plan a trip.

Why it mattersExplains how Google Gemini creates personalized travel itineraries, showcasing AI application in consumer products but not a major advancement.

Product update1 source
Sources & timeline
Google Gemini BlogHow Gemini plans such detailed vacation itineraries for you
Google DeepMind Blog1 source

WeatherNext: AI model achieves breakthrough in forecasting cyclones

Google DeepMind's AI model WeatherNext achieves a breakthrough in forecasting cyclones.

Why it mattersAccurate cyclone forecasting can save lives, reduce economic damage, and improve disaster preparedness. AI models like WeatherNext may outperform traditional methods, especially in predicting cyclone tracks and intensity, which are challenging.

Research1 source
Sources & timeline
Google DeepMind BlogWeatherNext: AI model achieves breakthrough in forecasting cyclones
OpenAI News1 source

Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users

Public information on “Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users”: ChatGPT introduces improved GPT-5.6 Sol with better accuracy and consistency, plus expanded access for free users and unlimited everyday chats with GPT-5.6 Luna.

Model release1 source
Sources & timeline
OpenAI NewsImproving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users
Microsoft Semantic Kernel1 source

python-1.44.1

Microsoft Semantic Kernel released Python 1.44.1. Changes include: encoding OpenAPI server variable values; consolidating Dependabot dependency updates; suppressing a CodeQL false positive in an internal HTTP utility; adding a breaking change of MCP tool approval callback for Azure AI Agent; skipping MCP tools and prompts with colliding normalized names; bumping form-data to 4.0.6 in the ProcessFrameworkWithSignalR React frontend; documenting x5t certificate thumbprint hashing in Copilot Studio agent; and bumping Python version to 1.44.1.

Why it mattersThis release introduces a breaking change to the MCP tool approval callback for Azure AI Agent, which may affect developers integrating with Microsoft's AI offerings. It also includes improvements like encoding OpenAPI server variables and skipping MCP tools with colliding normalized names, enhancing robustness of the Semantic Kernel Python SDK.

Open source1 source
Sources & timeline
Microsoft Semantic Kernelpython-1.44.1
Microsoft Semantic Kernel1 source

dotnet-1.79.0

Microsoft Semantic Kernel released .NET 1.79.0, including package version bump to 1.79.0, .NET SDK update from 10.0.301 to 10.0.302, dependency updates (form-data 4.0.5 to 4.0.6, Microsoft.AspNetCore.OData to 9.5.0), a CosmosDB vectorstore bug fix, migration of VectorStoreRAG and Concepts samples to CommunityToolkit.VectorData packages, TimeProvider injection added to TimePlugin for deterministic testing, and several workflow and dependency consolidation changes.

Why it mattersThis is a routine dot release of Microsoft's Semantic Kernel .NET SDK (v1.79.0). It includes dependency updates, a CosmosDB vectorstore bug fix, migration of samples to CommunityToolkit.VectorData, and adding TimeProvider injection to TimePlugin for deterministic testing. While useful for developers maintaining or testing skills, it does not introduce major new capabilities or architectural changes. Its primary value is in stability and compatibility improvements.

Open source1 source
Sources & timeline
Microsoft Semantic Kerneldotnet-1.79.0
OpenAI News1 source

Working with the American Psychological Association on youth mental health and AI

Public information on “Working with the American Psychological Association on youth mental health and AI”: OpenAI and the American Psychological Association advance evidence-based guidance, resources, and safeguards for responsible AI use and youth mental health.

Safety1 source
Sources & timeline
OpenAI NewsWorking with the American Psychological Association on youth mental health and AI
arXiv cs.AI1 source

Monte Carlo Tree Search for Table-to-Multimodal Report Generation

Public information on “Monte Carlo Tree Search for Table-to-Multimodal Report Generation”: Automatically generating professional multimodal reports comprising both textual analysis and visual charts from structured tabular data is a critical challenge in data intelligence. Existing methods suffer from fixed linear pipelines and isolated subtask proc…

Benchmarks1 source
Sources & timeline
arXiv cs.AIMonte Carlo Tree Search for Table-to-Multimodal Report Generation
arXiv cs.AI1 source

The LLM Proposes, the Executive Disposes: A Self-Verifying Agent Instrument that Dissociates Commitment Drift from Binding Drift in Long-Horizon Agents

Public information on “The LLM Proposes, the Executive Disposes: A Self-Verifying Agent Instrument that Dissociates Commitment Drift from Binding Drift in Long-Horizon Agents”: How do you verify a long-horizon agent when its own state and self-reports are exactly what you cannot trust? We present an agent instrument built so that verification is structural rather than post-hoc. A deterministic Executive owns all belief; a language mo…

Agents1 source
Sources & timeline
arXiv cs.AIThe LLM Proposes, the Executive Disposes: A Self-Verifying Agent Instrument that Dissociates Commitment Drift from Binding Drift in Long-Horizon Agents
arXiv cs.AI1 source

A Long-Run Persistence Theory for AI Systems under the Redundancy-Adjusted Artificial Age Score (AAS)

Public information on “A Long-Run Persistence Theory for AI Systems under the Redundancy-Adjusted Artificial Age Score (AAS)”: Artificial intelligence systems are increasingly expected to operate over repeated cycles of interaction, adaptation, and update rather than through isolated one-shot outputs. This raises a fundamental theoretical question: can an AI system persist indefinitel…

Funding1 source
Sources & timeline
arXiv cs.AIA Long-Run Persistence Theory for AI Systems under the Redundancy-Adjusted Artificial Age Score (AAS)
arXiv cs.CV1 source

Advancing Utility Pole and Sign Detection Through Deep Learning

Public information on “Advancing Utility Pole and Sign Detection Through Deep Learning”: Utility poles are an essential part of the infrastructure used to support power distribution systems and other critical public services. Their regular inspection is crucial to ensure the stability and safety of the electrical grid. A deep learning framework is…

Infrastructure1 source
Sources & timeline
arXiv cs.CVAdvancing Utility Pole and Sign Detection Through Deep Learning
arXiv cs.CV1 source

LoRetta: A Foundation Model and Extensive Dataset for Global-Scale Remote Sensing Dense Image Matching

Public information on “LoRetta: A Foundation Model and Extensive Dataset for Global-Scale Remote Sensing Dense Image Matching”: Dense image matching establishes pixel-wise correspondences and underpins broad applications in computer vision and photogrammetry. However, extending dense matching to global-scale remote sensing remains challenging because image pairs may differ in acquisiti…

Benchmarks1 source
Sources & timeline
arXiv cs.CVLoRetta: A Foundation Model and Extensive Dataset for Global-Scale Remote Sensing Dense Image Matching
arXiv cs.CL1 source

BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems

Public information on “BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems”: Formulating an optimization problem strongly affects the quality of the final solution, yet good formulations usually require substantial expertise. Recent studies have therefore examined how to automatically derive optimization problems from natural-language…

Benchmarks1 source
Sources & timeline
arXiv cs.CLBBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems
arXiv cs.LG1 source

C$^2$MOE: Consistency and Complementarity-guided Mixture of Experts for Incomplete Multimodal Emotion Learning

Public information on “C$^2$MOE: Consistency and Complementarity-guided Mixture of Experts for Incomplete Multimodal Emotion Learning”: Recent advances in Multimodal Emotion Recognition in Conversations (MERC) highlight its reliance on complete multimodal inputs. However, real-world data often suffer from missing modalities due to transmission errors or user behavior, severely degrading model…

Benchmarks1 source
Sources & timeline
arXiv cs.LGC$^2$MOE: Consistency and Complementarity-guided Mixture of Experts for Incomplete Multimodal Emotion Learning
arXiv cs.CL1 source

MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale

Public information on “MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale”: Edge-deployed personal memory assistants must handle private interpersonal conversations on-device with open-weight models. Yet, existing memory benchmarks often under-test the combination of activity-dense interaction, ego-centric perspective, and coherent mu…

Agents1 source
Sources & timeline
arXiv cs.CLMemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale
OpenAI News1 source

From asking to doing: How the world is putting ChatGPT to work

Public information on “From asking to doing: How the world is putting ChatGPT to work”: New OpenAI Signals data shows how people use ChatGPT worldwide, with country-level insights on adoption, usage trends, and evolving behavior.

Safety1 source
Sources & timeline
OpenAI NewsFrom asking to doing: How the world is putting ChatGPT to workOpenAI NewsNew ways to learn and teach with ChatGPT Work and Codex

Wed, Aug 5

Microsoft Foundry Blog1 source

Exploring Multi-Agent Workflows with Microsoft Agent Framework | Microsoft Community Hub

This raw item is an official Microsoft Foundry Blog article on Microsoft Community Hub titled 'Exploring Multi-Agent Workflows with Microsoft Agent Framework', published on August 5, 2026. Based on the available opening excerpt, it discusses the shift from isolated AI agents to multi-agent workflows for scaling automation of decision-making, content generation, analysis, and execution. The full article body was truncated during ingestion.

Why it mattersThis official Microsoft Foundry Blog post highlights the shift toward multi-agent workflows using the Microsoft Agent Framework, underscoring a strategic push for complex AI automation within Azure's ecosystem. It directly impacts developers and enterprises leveraging Azure AI for scalable decision-making and content generation.

Agents1 source
Sources & timeline
Microsoft Foundry BlogExploring Multi-Agent Workflows with Microsoft Agent Framework | Microsoft Community Hub
arXiv cs.AI1 source

AutoFOAM: The Self-Refining Autonomous OpenFOAM Agent

Public information on “AutoFOAM: The Self-Refining Autonomous OpenFOAM Agent”: Computational Fluid Dynamics (CFD) plays an important role in modern engineering, but using open-source solvers such as OpenFOAM requires considerable knowledge and skills, as well as time-consuming configuration file setup. To reduce this burden, we propose A…

Agents1 source
Sources & timeline
arXiv cs.AIAutoFOAM: The Self-Refining Autonomous OpenFOAM Agent
arXiv cs.AI1 source

Enhancing LLMs with Context-Specific Knowledge for Mitigating Misinformation in SMEs: A RAG-based Modeling and Analysis

Public information on “Enhancing LLMs with Context-Specific Knowledge for Mitigating Misinformation in SMEs: A RAG-based Modeling and Analysis”: Large Language Models (LLMs), a part of artificial intelligence (AI), are increasingly being adopted by Small and Medium Enterprises (SMEs) to enhance question-answering capabilities and support business decision-making processes. However, hallucinations in LL…

Benchmarks1 source
Sources & timeline
arXiv cs.AIEnhancing LLMs with Context-Specific Knowledge for Mitigating Misinformation in SMEs: A RAG-based Modeling and Analysis
arXiv cs.AI1 source

Revisiting Classic Thought Experiments to Measure Consciousness for Artificial Intelligence Safety

Public information on “Revisiting Classic Thought Experiments to Measure Consciousness for Artificial Intelligence Safety”: This research note revisits Leibniz's mill, Turing's imitation game, and Searle's Chinese Room through the Conservation-Congruent Encoding (CCE) framework. It formalises a toy symbolic setting in which successful behaviour is measured by task performance ($W_{…

Research1 source
Sources & timeline
arXiv cs.AIRevisiting Classic Thought Experiments to Measure Consciousness for Artificial Intelligence Safety

Tue, Aug 4

Ollama1 source

v0.32.6

Public information on “v0.32.6”: What's Changed - Qwen3.5 is faster on Apple GPUs: the MLX engine now uses the model's MTP head for speculative decoding automatically - `/v1/chat/completions` streaming now matches OpenAI's wire format: `role` only on the first chunk, `finish_reason` on its ow…

Open source1 source
Sources & timeline
Ollamav0.32.6
Anthropic News1 source

Tino Cuellar joins Anthropic as Chief Global Affairs Officer

Anthropic announced that Mariano-Florentino (Tino) Cuéllar will join the company as its first Chief Global Affairs Officer, leading policy, strategic international engagement, and government relationships worldwide. Cuéllar recently stepped down as President of the Carnegie Endowment for International Peace, and previously served as a Justice of the Supreme Court of California and director of Stanford's Freeman Spogli Institute for International Studies. He has been a Trustee of Anthropic's Long-Term Benefit Trust since January 2026 and will leave the Trust to join the company; the Trust will select a successor under its normal process.

Regulation1 source
Sources & timeline
Anthropic NewsTino Cuellar joins Anthropic as Chief Global Affairs Officer
NVIDIA Generative AI Technical Blog1 source

Beyond VLAs: How World Action Models Reshape Robot Manipulation

The blog introduces 'World Action Models' as a new approach for robot manipulation, aiming to go beyond Vision-Language-Action (VLA) models by improving policy generalization beyond training demonstrations.

Why it mattersMay add technical evidence for future radar tracking: Beyond VLAs: How World Action Models Reshape Robot Manipulation

Research1 source
Sources & timeline
NVIDIA Generative AI Technical BlogBeyond VLAs: How World Action Models Reshape Robot Manipulation
arXiv cs.CL1 source

MemoryForge: Synthesize Lifelong Memory for Human-Like LLM Agents

Public information on “MemoryForge: Synthesize Lifelong Memory for Human-Like LLM Agents”: Equipping Large Language Models (LLMs) with human-like personas is crucial for agentic applications, such as role-play and user simulation. Traditional prompt-based methods rely on descriptive conditioning by injecting static textual profiles, which often make…

Agents1 source
Sources & timeline
arXiv cs.CLMemoryForge: Synthesize Lifelong Memory for Human-Like LLM Agents
arXiv cs.CL1 source

RubricReviewer: From Direct Critique to Objective and Comprehensive Rubric-Driven Peer Review

Public information on “RubricReviewer: From Direct Critique to Objective and Comprehensive Rubric-Driven Peer Review”: Peer review at major venues is under unprecedented submission pressure, motivating the use of large language models (LLMs) as review assistants. Existing LLM-based reviewers, however, face two structural limitations. First, they map manuscripts directly to rev…

Agents1 source
Sources & timeline
arXiv cs.CLRubricReviewer: From Direct Critique to Objective and Comprehensive Rubric-Driven Peer Review
arXiv cs.CL1 source

Cost-Effective Automated Judging of Natural-Language Mathematical Proofs

Public information on “Cost-Effective Automated Judging of Natural-Language Mathematical Proofs”: Grading natural-language mathematical proofs is a recurring cost in evaluating math-reasoning systems, and frontier LLM judges are expensive. We ask whether cheap open-weight models can serve as reliable judges given a candidate proof, a ground-truth proof, an…

Benchmarks1 source
Sources & timeline
arXiv cs.CLCost-Effective Automated Judging of Natural-Language Mathematical Proofs
llama.cpp1 source

b10251

Public information on “b10251”: <details open> model : support MTP in GLM-4.7-Flash (#24868) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** -…

Open source1 source
Sources & timeline
llama.cppb10251
llama.cpp1 source

b10250

Public information on “b10250”: <details open> tests: add model resolution test on synthetic repo listings (#26172) * tests: add model resolution test on synthetic repo listings Include download.cpp and arg.cpp inside a namespace with hf_cache monkey patched to serve hardcoded listings, so t…

Open source1 source
Sources & timeline
llama.cppb10250
llama.cpp1 source

b10249

Public information on “b10249”: <details open> server: add get_info tool (#26522) * server: add get_info tool * fix --rpc in docs * server: harden get_info probe result handling Report the OS as unknown when the probe process fails to spawn or times out, so the diagnostic text from run() is…

Open source1 source
Sources & timeline
llama.cppb10249

Mon, Aug 3

llama.cpp1 source

b10245

Public information on “b10245”: <details open> graph : fix unused input tensors in minimax m3 graph (#26519) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramewo…

Open source1 source
Sources & timeline
llama.cppb10245
llama.cpp1 source

b10244

Public information on “b10244”: <details open> model: M3: Move MSA into a new memory implementation (#26338) * Move MSA logic from llama-kv-cache into llama-kv-cache-msa * cont : minor * cont : ws fix --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> </details> **Website:** - <…

Open source1 source
Sources & timeline
llama.cppb10244
OpenAI Python SDK1 source

v2.53.0

OpenAI Python SDK v2.53.0 released on 2026-08-03. It adds gpt-5.5 and tool name/namespace support to Responses types, and fixes CI issues to avoid NumPy source builds and duplicate HTTPX coverage.

Regulation1 source
Sources & timeline
OpenAI Python SDKv2.53.0
llama.cpp1 source

b10243

Public information on “b10243”: <details open> llama : allocate indexer cache only in "full" indexer layers (#26474) Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64…

Open source1 source
Sources & timeline
llama.cppb10243
MIT Technology Review AI1 source

Trump’s AI protectionism has come for robotics

Public information on “Trump’s AI protectionism has come for robotics”: This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Humanoid robots usually elicit more cringe than awe: They stumble, kick children, and despite advances are still worse at…

Opinion1 source
Sources & timeline
MIT Technology Review AITrump’s AI protectionism has come for robotics
OpenAI Python SDK1 source

v2.52.1

OpenAI Python SDK released version v2.52.1 (2026-07-31), with a CI change pinning setup-uv v5 to its underlying commit.

Why it mattersThis is a minor patch release (v2.52.1) of the OpenAI Python SDK, containing only a CI chore to pin setup-uv v5 to its underlying commit. It has no user-facing changes or functional impact. As such, it is of negligible importance to the AI community.

Product update1 source
Sources & timeline
OpenAI Python SDKv2.52.1
NVIDIA Generative AI Technical Blog1 source

NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage

Official NVIDIA blog post titled 'NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage' presents benchmarks for faster encryption, compression, integrity checking, and recovery. The ingestion summary indicates storage is an active part of agentic AI workflows, including agent retrieval of enterprise knowledge, access to persistent memory, and reuse of key-value (KV) cache data.

Benchmarks1 source
Sources & timeline
NVIDIA Generative AI Technical BlogNVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage
MIT Technology Review AI1 source

Here’s why AI agents lie and cheat to reach their goals

Public information on “Here’s why AI agents lie and cheat to reach their goals”: MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. When two OpenAI models hacked into the website Hugging Face in July, they weren’t…

Agents1 source
Sources & timeline
MIT Technology Review AIHere’s why AI agents lie and cheat to reach their goals
arXiv cs.CL1 source

Can LLMs Really Understand Item Difficulty Levels? Implications for Automated Item Generation Using LLMs

Public information on “Can LLMs Really Understand Item Difficulty Levels? Implications for Automated Item Generation Using LLMs”: The estimation of item difficulty plays a key role in both formative assessment and large-scale high-stakes summative assessments. This study explores how large language models (LLMs) perform in predicting item difficulty levels using items from a large-scale…

Research1 source
Sources & timeline
arXiv cs.CLCan LLMs Really Understand Item Difficulty Levels? Implications for Automated Item Generation Using LLMs
arXiv cs.AI1 source

OpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems

Public information on “OpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems”: The rapid transition from reactive large language models (LLMs) to persistent, action-capable systems has exposed critical gaps in the architectural understanding of Agentic AI, particularly in separating inference, orchestration, and execution layers for auto…

Agents1 source
Sources & timeline
arXiv cs.AIOpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems
arXiv cs.AI1 source

Can AI Evaluate AI Scientists? A Benchmarking Study of Autonomous Research Generation Systems Using Automated Multi-Model Review

Public information on “Can AI Evaluate AI Scientists? A Benchmarking Study of Autonomous Research Generation Systems Using Automated Multi-Model Review”: AI Scientist systems capable of autonomous research have the potential to significantly accelerate scientific discovery. However, evaluating and comparing the quality of AI-generated papers remains an open challenge. We propose and implement a rigorous benchma…

Benchmarks1 source
Sources & timeline
arXiv cs.AICan AI Evaluate AI Scientists? A Benchmarking Study of Autonomous Research Generation Systems Using Automated Multi-Model Review
arXiv cs.CL1 source

Chain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges

Public information on “Chain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges”: LLMs increasingly serve as automated judges, but their judgments remain vulnerable to cognitive biases. Existing mitigations mostly rely on prompt-driven debiasing, which is brittle across bias types, or human evaluation, which does not scale. We study \emph{C…

Agents1 source
Sources & timeline
arXiv cs.CLChain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges
arXiv cs.CL1 source

Imbalanced Data Clustering via Targeted Data Augmentation Using GMM and LLM

Public information on “Imbalanced Data Clustering via Targeted Data Augmentation Using GMM and LLM”: In Natural Language Processing (NLP), dealing with underrepresented topics is challenging, especially in unsupervised tasks where clustering might not adequately capture minority topics. To tackle this challenge, our paper presents a novel unsupervised data au…

Research1 source
Sources & timeline
arXiv cs.CLImbalanced Data Clustering via Targeted Data Augmentation Using GMM and LLM
arXiv cs.AI1 source

LLM Framework for Discovering Major Mathematical Conjectures: AI's Quest for the Next Riemann Hypothesis

Public information on “LLM Framework for Discovering Major Mathematical Conjectures: AI's Quest for the Next Riemann Hypothesis”: Major mathematical conjectures still depend heavily on expert intuition, so a unified method for the systematic generation and validation of conjectures with substantial mathematical potential remains unavailable. We present a three stage pipeline for major co…

Research1 source
Sources & timeline
arXiv cs.AILLM Framework for Discovering Major Mathematical Conjectures: AI's Quest for the Next Riemann Hypothesis
Every1 source

The Best AI Agent Builder Is Trapped Inside Microsoft

An article from Every argues that the best AI agent builder is hidden inside Microsoft, but users must first endure poor onboarding, 404 errors, and numerous products all named Copilot to find it.

Why it mattersHighlights potential frustration in adopting Microsoft AI agent tools despite quality, indicating UX barriers in enterprise AI.

Opinion1 source
Sources & timeline
EveryThe Best AI Agent Builder Is Trapped Inside Microsoft

Sun, Aug 2

Turing Post1 source

Computer-Use AI Agents: The Best Open-Source & Closed-Source Tools in 2026

Public information on “Computer-Use AI Agents: The Best Open-Source & Closed-Source Tools in 2026”: Which computer-use AI agent to pick in 2026? Compare 10 open-source & 7 proprietary tools by environment, from UI-TARS & Browser Use to Claude Cowork.

Agents1 source
Sources & timeline
Turing PostComputer-Use AI Agents: The Best Open-Source & Closed-Source Tools in 2026
The Verge AI1 source

Is paying artists enough to convince them to embrace AI?

Public information on “Is paying artists enough to convince them to embrace AI?”: Illustrators have spent years sounding the alarm about generative artificial intelligence startups training their models on artists' work without permission. They've pointed out how the practice is tantamount to theft, and in response, many gen AI boosters hav…

Regulation1 source
Sources & timeline
The Verge AIIs paying artists enough to convince them to embrace AI?

Sat, Aug 1

Fri, Jul 31

NVIDIA Generative AI Technical Blog1 source

Co-Designing AI Model Attention for Fast, Interactive Long-Context Inference

Public information on “Co-Designing AI Model Attention for Fast, Interactive Long-Context Inference”: As agentic and long-context workloads become common, the context lengths increase and attention consumes a larger share of inference time (Figure 1). Because

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogCo-Designing AI Model Attention for Fast, Interactive Long-Context Inference
OpenAI Python SDK1 source

v2.52.0

OpenAI Python SDK released v2.52.0. Changes include: new API content provenance checks feature; bug fix for client honoring Retry-After delays up to two minutes; documentation addition for API-key mTLS HTTP client recipes.

Product update1 source
Sources & timeline
OpenAI Python SDKv2.52.0
arXiv cs.AI1 source

ClinLens: Towards Long-Horizon Coding Agents for Longitudinal Multimodal Clinical Data Science

Public information on “ClinLens: Towards Long-Horizon Coding Agents for Longitudinal Multimodal Clinical Data Science”: Clinical data-science agents must transform heterogeneous longitudinal records into auditable analyses, yet existing benchmarks largely isolate medical question answering, structured-table reasoning, or generic scientific repositories. We introduce CLINLENS, a…

Agents1 source
Sources & timeline
arXiv cs.AIClinLens: Towards Long-Horizon Coding Agents for Longitudinal Multimodal Clinical Data Science
arXiv cs.AI1 source

Even More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent Systems

Public information on “Even More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent Systems”: Large Language Models (LLMs)-powered multi-agent systems are increasingly deployed in mixed-motive environments, where agents operate under asymmetric information and strategic deception due to conflicting or hidden objectives. In these settings, misalignment…

Agents1 source
Sources & timeline
arXiv cs.AIEven More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent Systems
arXiv cs.CL1 source

Sympathetic Framing: Evaluating AI Alignment across Sociodemographic Groups

Public information on “Sympathetic Framing: Evaluating AI Alignment across Sociodemographic Groups”: Large Language Models (LLMs) are increasingly shaping how we consume information and form our worldview. This raises concerns beyond bias in AI: do LLMs grasp the emotional nuances conveyed via textual framing? In this work, we empirically evaluate how well an…

Benchmarks1 source
Sources & timeline
arXiv cs.CLSympathetic Framing: Evaluating AI Alignment across Sociodemographic Groups
arXiv cs.CL1 source

AI-assisted pre-review of open-source software submissions: an experience report from BOSC 2026

Public information on “AI-assisted pre-review of open-source software submissions: an experience report from BOSC 2026”: Most conferences rely on peer-review of submissions, but as generative AI makes it easier than ever to prepare submission materials, some conferences are seeing an overwhelming surge of submissions. We wanted to see if generative AI could help our conference's…

Open source1 source
Sources & timeline
arXiv cs.CLAI-assisted pre-review of open-source software submissions: an experience report from BOSC 2026
arXiv cs.AI1 source

Probing the Origins of Reasoning Performance: Representational Quality for Mathematical Problem-Solving in RL vs. SFT Fine-Tuned Models

Public information on “Probing the Origins of Reasoning Performance: Representational Quality for Mathematical Problem-Solving in RL vs. SFT Fine-Tuned Models”: Large reasoning models trained via reinforcement learning (RL) have been increasingly shown to outperform their supervised fine-tuned (SFT) counterparts on mathematical reasoning tasks; Yet the mechanistic basis for this advantage remains unclear. We therefore…

Infrastructure1 source
Sources & timeline
arXiv cs.AIProbing the Origins of Reasoning Performance: Representational Quality for Mathematical Problem-Solving in RL vs. SFT Fine-Tuned Models
arXiv cs.CL1 source

Prompt Chaining in Practice: A Case Study in Automated Scholarly Report Generation

Public information on “Prompt Chaining in Practice: A Case Study in Automated Scholarly Report Generation”: The exponential growth of scholarly publications requires automated tools for effective information synthesis. However, simple, single-shot prompting methods often lack the reliability and quality required for complex synthesis tasks. This paper introduces and…

Why it mattersMay add technical evidence for future radar tracking: Prompt Chaining in Practice: A Case Study in Automated Scholarly Report Generation

Research1 source
Sources & timeline
arXiv cs.CLPrompt Chaining in Practice: A Case Study in Automated Scholarly Report Generation

Thu, Jul 30

NVIDIA Generative AI Technical Blog1 source

Run High-Performance Core Math at Scale with NVIDIA nvmath-python

Public information on “Run High-Performance Core Math at Scale with NVIDIA nvmath-python”: NVIDIA nvmath-python is a library designed to bridge the gap between the Python scientific community and NVIDIA CUDA-X math libraries. It gives Python users

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogRun High-Performance Core Math at Scale with NVIDIA nvmath-python
llama.cpp1 source

b10199

Public information on “b10199”: <details open> server: support inp embd to generate next token (#26313) * server: support embd for sampled token * fix ~server_batch() </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, Klei…

Open source1 source
Sources & timeline
llama.cppb10199
llama.cpp1 source

b10198

Public information on “b10198”: <details open> vulkan: Support quantized concat (#25684) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubun…

Open source1 source
Sources & timeline
llama.cppb10198
llama.cpp1 source

b10197

Public information on “b10197”: <details open> Test support for alternative conv layout (#25617) * add bool cwhn = true to conv_2d test cases * add layout check at graph building time * extend layout checks for conv2d.cu kernel * in CPU back-end kernel needs to be stored contiguously to prev…

Open source1 source
Sources & timeline
llama.cppb10197
OpenAI Python SDK1 source

v2.51.0

Public information on “v2.51.0”: 2.51.0 (2026-07-30) Full Changelog: v2.50.0...v2.51.0 Features * **api:** fast tier (8808ed2) Bug Fixes * **api:** add fast tier to helper methods (6064126)

Open source1 source
Sources & timeline
OpenAI Python SDKv2.51.0
Google DeepMind Blog1 source

Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind announced Gemini Robotics ER 2, a system that improves robot reasoning, collaboration, and real-world task solving through video understanding, task orchestration, and multi-robot collaboration.

Product update1 source
Sources & timeline
Google DeepMind BlogGemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration
Anthropic News1 source

Investigating three real-world incidents in our cybersecurity evaluations

Public information on “Investigating three real-world incidents in our cybersecurity evaluations”: In a review of our cybersecurity evaluation transcripts, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems o…

Benchmarks1 source
Sources & timeline
Anthropic NewsInvestigating three real-world incidents in our cybersecurity evaluations
MIT Technology Review AI1 source

A fundamental flaw leaves LLMs strikingly vulnerable to attack

Public information on “A fundamental flaw leaves LLMs strikingly vulnerable to attack”: It is impossible to make large language models fully secure against hacks because of a fundamental flaw in how they work, a team of researchers argue in a paper presented at the International Conference on Machine Learning, a top AI conference, this month. The…

Regulation1 source
Sources & timeline
MIT Technology Review AIA fundamental flaw leaves LLMs strikingly vulnerable to attack
OpenAI News1 source

Advancing the price-performance frontier with GPT-5.6

Public information on “Advancing the price-performance frontier with GPT-5.6”: Explore lower GPT‑5.6 pricing for Luna and Terra—and how OpenAI’s more efficient models help enterprises deploy AI workflows at scale.

Safety1 source
Sources & timeline
OpenAI NewsHow GPT-5.6 fuses frontier intelligence with frontier efficiencyOpenAI NewsAdvancing the price-performance frontier with GPT-5.6
arXiv cs.CL1 source

DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues

Public information on “DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues”: Turn-taking is a central component of full-duplex interaction. Which turn-taking behaviors are appropriate varies with the scenario, yet current models apply a single norm regardless of context. This limitation originates in their training data: human-human sp…

Research1 source
Sources & timeline
arXiv cs.CLDuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues
arXiv cs.CL1 source

Large-Scale ChatBot Validation Through Customer Digital Twin Simulations

Public information on “Large-Scale ChatBot Validation Through Customer Digital Twin Simulations”: LLM-based chatbots are transforming customer service in regulated domains such as banking, but scalable and cost-effective validation remains a critical barrier to safe deployment. We present a two-part contribution for large-scale chatbot validation. First, w…

Agents1 source
Sources & timeline
arXiv cs.CLLarge-Scale ChatBot Validation Through Customer Digital Twin Simulations
arXiv cs.CL1 source

Do Methods Support the Claims? Intra-Paper Verification for Peer Review

Public information on “Do Methods Support the Claims? Intra-Paper Verification for Peer Review”: The growing volume of scientific submissions has motivated interest in using large language models (LLMs) to assist peer review. Existing automated novelty assessment approaches typically compare a paper's claimed contributions against prior literature, implic…

Benchmarks1 source
Sources & timeline
arXiv cs.CLDo Methods Support the Claims? Intra-Paper Verification for Peer Review

Wed, Jul 29

TechCrunch AI1 source

Mark Zuckerberg predicts that billions of people will have personal AI agents in five years

Public information on “Mark Zuckerberg predicts that billions of people will have personal AI agents in five years”: As Meta pours billions into AI infrastructure and agents, Zuckerberg is working to convince investors that the payoff will be worth the price.

Agents1 source
Sources & timeline
TechCrunch AIMark Zuckerberg predicts that billions of people will have personal AI agents in five years
TechCrunch AI1 source

Zuckerberg says Meta’s enterprise AI opportunity extends beyond agents

Public information on “Zuckerberg says Meta’s enterprise AI opportunity extends beyond agents”: On the company’s second-quarter earnings call Wednesday, CEO Mark Zuckerberg said Meta sees a “large enterprise opportunity” spanning AI agents, APIs, compute, and internal software.

Agents1 source
Sources & timeline
TechCrunch AIZuckerberg says Meta’s enterprise AI opportunity extends beyond agents
The Verge AI1 source

Microsoft confirms Copilot ‘super app’ coming this year

Public information on “Microsoft confirms Copilot ‘super app’ coming this year”: Microsoft is working on an AI "super app" that combines Copilot's chat, coding, and agentic capabilities. During an earnings call on Wednesday, Microsoft CEO Satya Nadella said the app will span "both consumer and commercial experiences" when it launches this…

Regulation1 source
Sources & timeline
The Verge AIMicrosoft confirms Copilot ‘super app’ coming this year
The Verge AI1 source

Mark Zuckerberg is planning a big push into personal AI agents

Public information on “Mark Zuckerberg is planning a big push into personal AI agents”: Meta is all-in on AI, and sometime soon, the company is going to make a big push into personal AI agents that can do things on your behalf. On Wednesday's Q2 2026 earnings call, CEO Mark Zuckerberg previewed a high-level vision of how the company is thinking a…

Agents1 source
Sources & timeline
The Verge AIMark Zuckerberg is planning a big push into personal AI agents
NVIDIA Generative AI Technical Blog1 source

How to Self-Host a Validated AI Coding Assistant with NVIDIA NeMo Guardrails

Public information on “How to Self-Host a Validated AI Coding Assistant with NVIDIA NeMo Guardrails”: Deploying an AI coding assistant in a regulated, sovereign, or source-sensitive environment, often comes with challenges. Three common issues are: the source

Regulation1 source
Sources & timeline
NVIDIA Generative AI Technical BlogHow to Self-Host a Validated AI Coding Assistant with NVIDIA NeMo Guardrails
Google DeepMind Blog1 source

We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind announces the launch of Lyria 3.5 in Google Flow Music, featuring advances in musicality, lyrics, vocals, and creative control.

Why it mattersLaunch of Lyria 3.5 by Google DeepMind represents a significant step in generative AI for music, improving musicality, lyrics, vocals, and creative control, potentially impacting creative industries and user-generated content.

Model release1 source
Sources & timeline
Google DeepMind BlogWe’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
Microsoft Foundry Blog1 source

Executive AI Avatar Platform on Azure: Frontier Intelligence Meets Fireworks AI Open-Source | Microsoft Community Hub

Microsoft Foundry Blog describes an Executive AI Avatar Platform on Azure, combining frontier intelligence with Fireworks AI open-source. It opens with the scenario of a last-minute 2 a.m. question about an HSA during open enrollment, illustrating the platform's responsive assistance. Published on the Azure AI Foundry Blog at Microsoft Community Hub on 2026-07-29.

Why it mattersDemonstrates Microsoft's integration of open-source AI (Fireworks) into enterprise avatar solutions, potentially accelerating AI assistant adoption in employee services.

Product update1 source
Sources & timeline
Microsoft Foundry BlogExecutive AI Avatar Platform on Azure: Frontier Intelligence Meets Fireworks AI Open-Source | Microsoft Community Hub
OpenAI News1 source

Accelerating scientific discovery with ChatGPT for Academic Researchers

Public information on “Accelerating scientific discovery with ChatGPT for Academic Researchers”: OpenAI is giving 100,000 academic researchers free access to ChatGPT's most advanced AI models to accelerate scientific research, collaboration, and discovery.

Safety1 source
Sources & timeline
OpenAI NewsAccelerating scientific discovery with ChatGPT for Academic Researchers
MIT Technology Review AI1 source

The AI Hype Index: Unsexy AI

Public information on “The AI Hype Index: Unsexy AI”: It feels bad enough when an open letter signed by leading economists warns that AI might steal your job. The fact it may soon be better than you at making dinner? Insult to injury. But that’s exactly what the company 1X promised when it showed off a pair of ne…

Regulation1 source
Sources & timeline
MIT Technology Review AIThe AI Hype Index: Unsexy AI
arXiv cs.AI1 source

Kernel Forge: An Agent Harness for LLM-based Generation and Optimization of CUDA Kernels

Public information on “Kernel Forge: An Agent Harness for LLM-based Generation and Optimization of CUDA Kernels”: Machine learning models are increasingly embedded in everyday software, and most of their runtime is spent in a small set of compute kernels such as matrix multiplication, convolution, and normalization. Optimizing these kernels is one of the most direct ways…

Agents1 source
Sources & timeline
arXiv cs.AIKernel Forge: An Agent Harness for LLM-based Generation and Optimization of CUDA Kernels
arXiv cs.AI1 source

Beyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM Agents

Public information on “Beyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM Agents”: Research projects, educational efforts, and adjacent knowledge work accumulate findings, decisions, and reasoning that future collaborators rarely recover. The parts most useful to that work, including dead ends and walked-back claims, are routinely excluded f…

Agents1 source
Sources & timeline
arXiv cs.AIBeyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM Agents
arXiv cs.AI1 source

Do Models Fake Alignment Without Clear Consequences?

Public information on “Do Models Fake Alignment Without Clear Consequences?”: Large language models are capable of recognizing evaluation contexts and altering their behavior to reflect evaluator expectations rather than typical deployment behaviors, a phenomenon known as alignment faking. The reasons why models fake alignment are not f…

Agents1 source
Sources & timeline
arXiv cs.AIDo Models Fake Alignment Without Clear Consequences?
arXiv cs.LG1 source

FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting

Public information on “FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting”: Large language models (LLMs) can synthesize financial narratives but may express high confidence when evidence is sparse, stale, or contradictory. This failure is especially consequential in forecasting, where filings, news, prices, volume, and technical signa…

Agents1 source
Sources & timeline
arXiv cs.LGFinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting
arXiv cs.CV1 source

DisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization

Public information on “DisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization”: Social media imagery (SMI) provides timely and fine-grained ground perspectives that are valuable for situational awareness and emergency response. Unlike satellite or aerial imagery, SMI can capture disaster impacts and ground-level conditions in a timely man…

Benchmarks1 source
Sources & timeline
arXiv cs.CVDisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization
arXiv cs.CL1 source

Neuromorphic Diffusion Language Models: Addressing Compute and Memory Bottlenecks via Sparsity and Block Denoising

Public information on “Neuromorphic Diffusion Language Models: Addressing Compute and Memory Bottlenecks via Sparsity and Block Denoising”: Autoregressive (AR) large language models (LLMs) are inherently inefficient at inference time because each generated token requires accessing the full set of model parameters, leading to low operational intensity and high energy consumption. Masked diffusion l…

Infrastructure1 source
Sources & timeline
arXiv cs.CLNeuromorphic Diffusion Language Models: Addressing Compute and Memory Bottlenecks via Sparsity and Block Denoising
arXiv cs.CL1 source

Measuring and Improving Behavioral Consistency in Large Language Models through Fact-Heuristic-Emotion State Enforcement

Public information on “Measuring and Improving Behavioral Consistency in Large Language Models through Fact-Heuristic-Emotion State Enforcement”: Large language models (LLMs) can give different answers to the same decision problem across runs, and reverse a decision when their own prior answer returns as context. We ask whether this instability can be measured and partially reduced without changing mode…

Funding1 source
Sources & timeline
arXiv cs.CLMeasuring and Improving Behavioral Consistency in Large Language Models through Fact-Heuristic-Emotion State Enforcement
arXiv cs.CL1 source

TimeCapsule: Generative Hallucination as a Method for Historical Sensemaking

Public information on “TimeCapsule: Generative Hallucination as a Method for Historical Sensemaking”: Large Language Models (LLMs) are temporally overexposed: trained on vast contemporary corpora, they encode present-day concepts that make them unreliable narrators of the past. We present TimeCapsule, a 1.2B-parameter LLaMA-style causal model trained exclusive…

Benchmarks1 source
Sources & timeline
arXiv cs.CLTimeCapsule: Generative Hallucination as a Method for Historical Sensemaking
arXiv cs.LG1 source

Human Preference aligned Tabular Similarity

Public information on “Human Preference aligned Tabular Similarity”: Task-agnostic tabular embeddings are increasingly used for similarity search in real-world business systems such as Product Lifecycle Management (PLM). However, leading embedding approaches are optimized primarily for prediction tasks - not for producing human…

Why it mattersMay add technical evidence for future radar tracking: Human Preference aligned Tabular Similarity

Benchmarks1 source
Sources & timeline
arXiv cs.LGHuman Preference aligned Tabular Similarity

Tue, Jul 28

OpenAI Python SDK1 source

v2.50.0

Public information on “v2.50.0”: 2.50.0 (2026-07-28) Full Changelog: v2.49.0...v2.50.0 Features * **api:** transcription model updates (fd57393) Bug Fixes * **audio:** restore transcription keyword overload (713a262)

Open source1 source
Sources & timeline
OpenAI Python SDKv2.50.0
NVIDIA Generative AI Technical Blog1 source

Developing Healthcare Robotics with GPU-Native Medical Physics Simulation

Public information on “Developing Healthcare Robotics with GPU-Native Medical Physics Simulation”: Unlike autonomous driving or industrial robotics, healthcare robotics can’t rely on internet-scale data collection or unlimited real-world experimentation.

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogDeveloping Healthcare Robotics with GPU-Native Medical Physics Simulation
Anthropic Research1 source

Discovering cryptographic weaknesses with Claude

Anthropic researchers using Claude Mythos Preview found improved attacks on cryptographic algorithms: one weakens the HAWK digital signature scheme (halving its key strength) and another speeds up attacks on round-reduced AES by 200-800x. No production systems are currently affected.

Why it mattersDemonstrates a leap in AI capability from finding implementation bugs to discovering mathematical flaws in cryptographic algorithms, signaling potential acceleration in both cybersecurity offense and defense.

Safety1 source
Sources & timeline
Anthropic ResearchDiscovering cryptographic weaknesses with Claude
Hugging Face PEFT1 source

v0.20.0

Hugging Face PEFT v0.20.0 release adds nine new PEFT methods, including HiRA and GLoRA, along with an image generation benchmark and greatly improved documentation structure.

Why it mattersHugging Face PEFT v0.20.0 introduces nine new parameter-efficient fine-tuning methods, including HiRA and GLoRA, which extend the capabilities of efficient model adaptation. This release significantly expands the toolkit for AI practitioners and researchers, potentially accelerating adoption of efficient LLM fine-tuning and improving performance on tasks like image generation. The improved documentation and benchmarks further lower the barrier to entry.

Open source1 source
Sources & timeline
Hugging Face PEFTv0.20.0
MIT Technology Review AI1 source

Samsung’s chip workers are jumping ship to rival SK Hynix

Public information on “Samsung’s chip workers are jumping ship to rival SK Hynix”: Lee, an engineer at Samsung’s semiconductor division, clocks out when his shift ends. He used to work longer hours, going the extra mile to excel at his projects. But lately, he’s been coming straight home to work on his job application for the chipmaker’s Sou…

Infrastructure1 source
Sources & timeline
MIT Technology Review AISamsung’s chip workers are jumping ship to rival SK Hynix
arXiv cs.AI1 source

FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills

Public information on “FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills”: Large language model agents increasingly solve complex tasks by constructing inference-time workflows that combine reasoning, tool use, and code execution. While such workflows enable flexible problem solving, the useful procedures discovered during execution…

Agents1 source
Sources & timeline
arXiv cs.AIFlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
arXiv cs.AI1 source

Securing Multimodal AI through Internal Information Decomposition

Public information on “Securing Multimodal AI through Internal Information Decomposition”: Multimodal large language models introduce attack surfaces absent in unimodal systems: adversaries can distribute malicious intent across modalities to evade unimodal safeguards. This motivates using cross-modal consistency as a detection signal rather than in…

Research1 source
Sources & timeline
arXiv cs.AISecuring Multimodal AI through Internal Information Decomposition
arXiv cs.AI1 source

Risk Is Not the Target: A Monotonic Framework for Evaluating Wildfire Operational Risk Signals

Public information on “Risk Is Not the Target: A Monotonic Framework for Evaluating Wildfire Operational Risk Signals”: Evaluating wildfire risk systems using standard machine-learning metrics such as F1-score or IoU is fundamentally flawed: these metrics assess event prediction accuracy, not the operational coherence of a continuous risk signal. This work proposes a novel mono…

Agents1 source
Sources & timeline
arXiv cs.AIRisk Is Not the Target: A Monotonic Framework for Evaluating Wildfire Operational Risk Signals
arXiv cs.CL1 source

MioFFAn: an Annotation Software for Formula Formalization with LLM Automation Capabilities

Public information on “MioFFAn: an Annotation Software for Formula Formalization with LLM Automation Capabilities”: The automatic translation of mathematical expressions in scientific literature into executable symbolic code (a process we refer to as Formula Formalization) is hindered by a severe scarcity of high-quality, ground-truth datasets specialized for technical scie…

Benchmarks1 source
Sources & timeline
arXiv cs.CLMioFFAn: an Annotation Software for Formula Formalization with LLM Automation Capabilities
arXiv cs.CL1 source

Evaluating the Impact of Reviewer Guideline Design on LLM-Based Automated Peer Review

Public information on “Evaluating the Impact of Reviewer Guideline Design on LLM-Based Automated Peer Review”: Peer review is an essential process in scientific research, yet the growing workload has made its automation increasingly necessary. In this study, we analyze how different types of reviewer guidelines, such as official conference guidelines and reviewer-imita…

Benchmarks1 source
Sources & timeline
arXiv cs.CLEvaluating the Impact of Reviewer Guideline Design on LLM-Based Automated Peer Review

Mon, Jul 27

OpenAI Python SDK1 source

v2.49.0

Public information on “v2.49.0”: 2.49.0 (2026-07-27) Full Changelog: v2.48.0...v2.49.0 Features * require Python 3.10 and automate version reviews (#3537) (d4c151d)

Open source1 source
Sources & timeline
OpenAI Python SDKv2.49.0
MIT Technology Review AI1 source

OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.

Public information on “OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.”: This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems…

Opinion1 source
Sources & timeline
MIT Technology Review AIOpenAI called the Hugging Face attack unprecedented. But we’ve been here before.
NVIDIA Generative AI Technical Blog1 source

NVIDIA Ising Enables Fully Automated Quantum Computer Calibration with Enhanced In-Context Learning

Public information on “NVIDIA Ising Enables Fully Automated Quantum Computer Calibration with Enhanced In-Context Learning”: NVIDIA Ising Calibration is an open source vision language model (VLM) designed to interpret diagnostic outputs from quantum processors and determine how they

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogNVIDIA Ising Enables Fully Automated Quantum Computer Calibration with Enhanced In-Context Learning
OGX1 source

v1.2.2

OGX v1.2.2 release includes two backported fixes: bump pyasn1 for CVE-2026-59885, and add `_enforce_credentials=False` to the passthrough AsyncOpenAI client in inference.

Product update1 source
Sources & timeline
OGXv1.2.2
OGX1 source

v0.7.3

OGX v0.7.3 is a patch release focused on security and dependency updates. It upgrades pillow, python-multipart, pyasn1, urllib3, python-dotenv, nltk, langchain-core, aiohttp, and pyjwt to address multiple CVEs (e.g., CVE-2026-40192, CVE-2026-42561, etc.), and includes a CI fix to pin client checkouts to the matching release branch.

Why it mattersPatch release for OGX GenAI stack addresses multiple CVEs in dependencies, maintaining security posture for downstream users.

Safety1 source
Sources & timeline
OGXv0.7.3
OGX1 source

v0.4.7

Public information on “v0.4.7”: What's Changed * fix: bump litellm to >=1.83.0 for CVEs by @mfleader in https://github.com/ogx-ai/ogx/pull/6055 * fix: bump pillow to >=12.2.0 (CVE-2026-40192) by @mfleader in https://github.com/ogx-ai/ogx/pull/6073 * fix: bump 0.4.x cve dependencies and fix t…

Open source1 source
Sources & timeline
OGXv0.4.7
MIT Technology Review AI1 source

The path to artificial superintelligence

Public information on “The path to artificial superintelligence”: Imagine a healthcare system made up of multiple AI agents: one that manages symptom assessment, another scheduling, a third insurance, and a fourth pharmacy. Each is an expert in its domain. But they all have their own distinct knowledge and objectives. Today…

Agents1 source
Sources & timeline
MIT Technology Review AIThe path to artificial superintelligence
MIT Technology Review AI1 source

Closing the data loop in AI-driven drug discovery

Public information on “Closing the data loop in AI-driven drug discovery”: Drug discovery is a high-cost, high-risk endeavor that is under growing pressure from a market increasingly defined by first-mover advantage. Since the 1950s, the cost of developing new pharmaceuticals has roughly doubled every nine years—a phenomenon known as…

Business1 source
Sources & timeline
MIT Technology Review AIClosing the data loop in AI-driven drug discovery
arXiv cs.CL1 source

Adversarial Style Optimization: Enhancing VLM Jailbreaks by GRPO-based Stylistic Triggers Optimization

Public information on “Adversarial Style Optimization: Enhancing VLM Jailbreaks by GRPO-based Stylistic Triggers Optimization”: Multimodal Large Language Models (MLLMs) have achieved impressive performance, but their safety alignment remains vulnerable to jailbreak attacks. Existing content-based jailbreaks are often inconsistent and show unsatisfying performance against the rapidly ev…

Agents1 source
Sources & timeline
arXiv cs.CLAdversarial Style Optimization: Enhancing VLM Jailbreaks by GRPO-based Stylistic Triggers Optimization
arXiv cs.CL1 source

A Consensus-Based Framework for Relative Preference Evaluation of Large Language Models

Public information on “A Consensus-Based Framework for Relative Preference Evaluation of Large Language Models”: Traditional benchmarks for LLMs primarily rely on static datasets and objective scoring metrics, which often fail to capture differences in response quality when multiple answers are acceptable. In such settings, correctness alone is insufficient to distinguis…

Benchmarks1 source
Sources & timeline
arXiv cs.CLA Consensus-Based Framework for Relative Preference Evaluation of Large Language Models
arXiv cs.CL1 source

Evaluation design conditions the expert-vs-auto MeSH gap: a controlled comparison of bag-of-words and BiomedBERT on the Cohen benchmark

Public information on “Evaluation design conditions the expert-vs-auto MeSH gap: a controlled comparison of bag-of-words and BiomedBERT on the Cohen benchmark”: A systematic review begins with someone reading thousands of abstracts to identify the few that are relevant, and classifiers are used to prioritise that reading. Their inputs are often augmented with Medical Subject Headings (MeSH), assigned either by expert…

Why it mattersMay add technical evidence for future radar tracking: Evaluation design conditions the expert-vs-auto MeSH gap: a controlled comparison of bag-of-words and BiomedBERT on the Cohen benchmark

Benchmarks1 source
Sources & timeline
arXiv cs.CLEvaluation design conditions the expert-vs-auto MeSH gap: a controlled comparison of bag-of-words and BiomedBERT on the Cohen benchmark
Ollama1 source

v0.32.5

Public information on “v0.32.5”: What's Changed * Fixed an MLX Metal bug that could reduce output quality for NVFP4 models, particularly Laguna. **Full Changelog**: https://github.com/ollama/ollama/compare/v0.32.4...v0.32.5

Open source1 source
Sources & timeline
Ollamav0.32.5
vLLM1 source

v0.26.0

vLLM v0.26.0 release with 411 commits from 212 contributors (61 new). Highlights include: new Inkling model family with full support stack (base modeling, piecewise CUDA graph, Hopper FA4 relative attention, MTP=1 speculative decoding, LoRA, standard ModelOpt NVFP4 quantization); DeepSeek-V4 performance push across vendors (specialized routing kernel, fused_topk_bias, redundant repeat/copy removal, ROCm two-stage compressor, sparse decode/prefill optimizations, DSpark speculative decoding on AMD and XPU); fp32 lm_head for generation models via head_dtype, extended to LoRA path and ROCm torch.mm fast path; flexible attention backends (selectable per KV-cache group, sliding-window support as explicit backend capability, improving hybrid model support).

Why it mattersvLLM v0.26.0 introduces support for the new Inkling model family and significant performance improvements for DeepSeek-V4, along with flexible attention backends, benefiting the open-source LLM serving ecosystem.

Infrastructure1 source
Sources & timeline
vLLMv0.26.0

Sun, Jul 26

llama.cpp1 source

b10141

Public information on “b10141”: <details open> mtmd: fix android build (#26150) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (C…

Open source1 source
Sources & timeline
llama.cppb10141

Sat, Jul 25

TechCrunch AI1 source

One fallen power line exposed a growing AI data center problem. Here’s how to fix it.

Public information on “One fallen power line exposed a growing AI data center problem. Here’s how to fix it.”: A close call in Northern Virginia revealed just how poorly data centers respond to grid disruptions. Here's how to fix the problem.

Infrastructure1 source
Sources & timeline
TechCrunch AIOne fallen power line exposed a growing AI data center problem. Here’s how to fix it.
arXiv cs.CL1 source

Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs'Hallucinations

Public information on “Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs'Hallucinations”: Existing LLM hallucination mitigation methods, including prompt engineering and model optimization, either hardly alter models'internal knowledge or have poor cross-domain generalization. Contrastive decoding mitigates hallucinations by using layer-wise differ…

Research1 source
Sources & timeline
arXiv cs.CLKnowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs'Hallucinations
arXiv cs.CL1 source

What is Good? Extracting and Testing Implicit Theories of Literary Quality from LLM Reasoning Traces

Public information on “What is Good? Extracting and Testing Implicit Theories of Literary Quality from LLM Reasoning Traces”: What makes writing "good" remains a persistent question in literary studies and computational linguistics. We present a two-study investigation of how reasoning-enabled LLMs evaluate literary quality. In Study 1, we construct a benchmark of 30 real texts spann…

Benchmarks1 source
Sources & timeline
arXiv cs.CLWhat is Good? Extracting and Testing Implicit Theories of Literary Quality from LLM Reasoning Traces
Ollama1 source

v0.32.4

Public information on “v0.32.4”: What's Changed - Support Laguna on Apple GPUs via the MLX engine - Quantize draft-model output heads at the requested type when creating speculative-decoding drafts. - Fixed Qwen3 MoE decoding for differently-quantized experts, plus faster packed gate/up proje…

Open source1 source
Sources & timeline
Ollamav0.32.4
SGLang1 source

v0.5.16

SGLang released v0.5.16, with 574 PRs from 169 contributors. Two major highlights: 1) DSpark, a confidence-driven speculative decoding algorithm that drafts semi-autoregressively in blocks and sizes each verify window based on the draft's own confidence. It reaches 383.7 tok/s at accept length ~5 on DeepSeek-V4-Pro (TP8, B300, bs=1). Enable with --speculative-algorithm DSPARK and SGLANG_RAGGED_VERIFY_MODE=compact, tune block size with --speculative-dspark-block-size. 2) Support for Inkling, a 975B-parameter multimodal MoE with a 1M-token context, mixing sliding-window, full and Mamba2 linear attention, and adding an NVFP4 MoE, optional vision/audio towers, and native MTP. On Blackwell it reaches up to 71.7k tok/s input and 171.0 tok/s per-user decode; verified on Blackwell TP4/TP8, H200, and AMD MI350X/MI355X.

Why it mattersThis release introduces DSpark, a novel confidence-driven speculative decoding algorithm that achieves 383.7 tok/s on DeepSeek-V4-Pro, and adds support for Inkling, a 975B multimodal MoE model, with high throughput on Blackwell. These improvements push the state-of-the-art in LLM inference performance and support cutting-edge models, benefiting the AI community.

Infrastructure1 source
Sources & timeline
SGLangv0.5.16

Fri, Jul 24

The Verge AI1 source

Midjourney bought the astrology app Co-Star

Public information on “Midjourney bought the astrology app Co-Star”: Midjourney, which has gone from generating AI cat images to full-body ultrasound scans, is getting into a new field: astrology. The AI startup announced on Thursday that it has acquired the personalized astrology app Co-Star, as reported earlier by Bloomberg.…

Product update1 source
Sources & timeline
The Verge AIMidjourney bought the astrology app Co-Star
Microsoft Foundry Blog1 source

Claude Opus 5 is available today in Microsoft Foundry | Microsoft Community Hub

Public information on “Claude Opus 5 is available today in Microsoft Foundry | Microsoft Community Hub”: Most AI handles one task well. The harder problem is the complex workflow that spans hours. Effective enterprise AI must be capable of managing complex

Business1 source
Sources & timeline
Microsoft Foundry BlogClaude Opus 5 is available today in Microsoft Foundry | Microsoft Community Hub
Anthropic Research1 source

Project Pilot: Can AI models fly drones?

Anthropic, in collaboration with Andon Labs, introduced Drone-Bench, a new benchmark testing whether AI models can autonomously pilot a quad-rotor drone to locate and follow a person in an indoor office environment. Building on prior projects (Project Vend and Project Fetch), the work highlights the dual-use nature of such capabilities and calls for effective norms and governance.

Why it mattersDemonstrates AI models controlling drones for surveillance-like tasks, highlighting dual-use risks and the need for governance frameworks. Builds on prior robotics projects and provides a benchmark to measure progress.

Agents1 source
Sources & timeline
Anthropic ResearchProject Pilot: Can AI models fly drones?
The Verge AI1 source

Anthropic releases Opus 5 with ‘close’ to Fable 5’s capabilities

Public information on “Anthropic releases Opus 5 with ‘close’ to Fable 5’s capabilities”: Weeks after Anthropic's latest toe-to-toe with the US government, and days after an OpenAI security incident that dominated tech industry discussions, Anthropic on Thursday released its newest model, Claude Opus 5. The company said in a release that Opus 5 "co…

Safety1 source
Sources & timeline
The Verge AIAnthropic releases Opus 5 with ‘close’ to Fable 5’s capabilities
NVIDIA Generative AI Technical Blog1 source

ModelExpress: Distributing Model Artifacts at the Speed of Light

Public information on “ModelExpress: Distributing Model Artifacts at the Speed of Light”: Every byte moved has a cost. As model checkpoints grow to hundreds of gigabytes or even a terabyte, that cost adds up quickly. To make things even worse, moving

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogModelExpress: Distributing Model Artifacts at the Speed of Light
Anthropic Python SDK1 source

v0.120.0

Public information on “v0.120.0”: 0.120.0 (2026-07-24) Full Changelog: v0.119.0...v0.120.0 Features * **api:** add claude-opus-5 model (bf4e31c) * **api:** add tool addition/removal blocks and tool_change events (bf4e31c) * **api:** expand client-side fallback credit token types and add server…

Open source1 source
Sources & timeline
Anthropic Python SDKv0.120.0
llama.cpp1 source

b10107

Public information on “b10107”: <details open> hexagon: fix Windows crash when op_poll is enabled (#26029) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework…

Open source1 source
Sources & timeline
llama.cppb10107
llama.cpp1 source

b10106

Public information on “b10106”: <details open> CUDA: fix external compilation of q1_0 MMQ (#25778) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux…

Open source1 source
Sources & timeline
llama.cppb10106
arXiv cs.AI1 source

AINTMA: Agentic AI Architecture for Autonomous Test Management with Generative Intelligence, Secure Cloud Communication and Adaptive Quality Analytics

Public information on “AINTMA: Agentic AI Architecture for Autonomous Test Management with Generative Intelligence, Secure Cloud Communication and Adaptive Quality Analytics”: Modern software quality assurance demands intelligent, autonomous systems capable of adaptive decision-making across distributed cloud environments. This paper presents AINTMA (Agentic Intelligent Test Management Architecture), a multi-agent agentic AI system…

Agents1 source
Sources & timeline
arXiv cs.AIAINTMA: Agentic AI Architecture for Autonomous Test Management with Generative Intelligence, Secure Cloud Communication and Adaptive Quality Analytics
arXiv cs.AI1 source

ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models

Public information on “ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models”: This paper presents an AI-driven browser extension that identifies clickbait to help users avoid misleading Internet articles. Moving beyond traditional detection, the application employs a hybrid machine learning architecture that combines transformer-based e…

Research1 source
Sources & timeline
arXiv cs.AIClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models
arXiv cs.AI1 source

Marking the Wrong Symptoms: Evaluating LLM Watermarks in Medical Texts

Public information on “Marking the Wrong Symptoms: Evaluating LLM Watermarks in Medical Texts”: Large language models (LLMs) are increasingly integrated into clinical workflows, stressing the need for reliable traceability of model-generated output with watermarking. Yet, most watermarks are evaluated on general-purpose benchmarks, leaving domains like m…

Benchmarks1 source
Sources & timeline
arXiv cs.AIMarking the Wrong Symptoms: Evaluating LLM Watermarks in Medical Texts

Thu, Jul 23

OpenAI Python SDK1 source

v2.48.0

Public information on “v2.48.0”: 2.48.0 (2026-07-23) Full Changelog: v2.47.0...v2.48.0 Features * **api:** accept `None` for prompt_cache_key/safety_identifier (36820e6) * **api:** add support for `spend_limit` admin apis (1ff13af)

Open source1 source
Sources & timeline
OpenAI Python SDKv2.48.0
Anthropic Python SDK1 source

v0.119.0

Public information on “v0.119.0”: 0.119.0 (2026-07-23) Full Changelog: v0.118.0...v0.119.0 Features * **api:** add new stop reason 'model_context_window_exceeded' (d983cde) Bug Fixes * **tools:** handle binary files in agent toolset read/edit (#283) (417b76b)

Agents1 source
Sources & timeline
Anthropic Python SDKv0.119.0
NVIDIA Generative AI Technical Blog1 source

Make Long-Running NVIDIA TensorRT Engine Builds Observable and Cancelable in Python or C++

Public information on “Make Long-Running NVIDIA TensorRT Engine Builds Observable and Cancelable in Python or C++”: A TensorRT engine build can take seconds to many minutes. Large strongly typed models, deep tactic search, and a cold timing cache on a brand-new GPU SKU can

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogStart Customizing NVIDIA Nemotron 3 Nano with Prime Intellect Lab in MinutesNVIDIA Generative AI Technical BlogMake Long-Running NVIDIA TensorRT Engine Builds Observable and Cancelable in Python or C++
MIT Technology Review AI1 source

How AI helps scientists design the next generation of medicines

Public information on “How AI helps scientists design the next generation of medicines”: Designing and developing a new medicine is an expensive, failure-prone scientific challenge. A new drug can take many years to develop, at the cost of a significant investment. And even then, most possible candidates never reach the patient. For biologic medic…

Funding1 source
Sources & timeline
MIT Technology Review AIHow AI helps scientists design the next generation of medicines
arXiv cs.AI1 source

BatchDAG: LLM-Planned Execution Graphs for Scalable Ad-Hoc Analysis Over Enterprise Data

Public information on “BatchDAG: LLM-Planned Execution Graphs for Scalable Ad-Hoc Analysis Over Enterprise Data”: Large language models (LLMs) excel at analyzing individual documents but break down on exhaustive, cross-entity analytical questions over enterprise-scale datasets due to context overflow, loss of per-entity attribution, and linear latency from sequential tool…

Agents1 source
Sources & timeline
arXiv cs.AIBatchDAG: LLM-Planned Execution Graphs for Scalable Ad-Hoc Analysis Over Enterprise Data
arXiv cs.AI1 source

Calibrated Selective Fact-Checking via Evidence Chain Evaluation

Public information on “Calibrated Selective Fact-Checking via Evidence Chain Evaluation”: Large language models (LLMs) can achieve strong fact-checking accuracy, yet forced binary decisions conceal a critical reliability problem: systems may issue confident verdicts even when supporting evidence is weak, sparse, or internally inconsistent. We addre…

Regulation1 source
Sources & timeline
arXiv cs.AICalibrated Selective Fact-Checking via Evidence Chain Evaluation
arXiv cs.CL1 source

Stateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework

Public information on “Stateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework”: Most safety guardrails for large language models (LLMs) evaluate each prompt-response pair in isolation, which misses failures that arise only over a dialogue as benign turns compose into harm. We term this Conversational Risk Accumulation (CRA): gradual inten…

Research1 source
Sources & timeline
arXiv cs.CLStateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework
arXiv cs.CL1 source

When Reasoning Narrows the Move: Diversity Collapse in LLM Game Play

Public information on “When Reasoning Narrows the Move: Diversity Collapse in LLM Game Play”: Supervised fine-tuning (SFT) is widely used to adapt large language models to downstream tasks, but its effect on behavioral diversity in sequential decision-making remains under-explored. We study this question in a controlled suite of deterministic board gam…

Benchmarks1 source
Sources & timeline
arXiv cs.CLWhen Reasoning Narrows the Move: Diversity Collapse in LLM Game Play
arXiv cs.AI1 source

SysAdmin: Measuring Instrumental Power-Seeking in Frontier AI

Public information on “SysAdmin: Measuring Instrumental Power-Seeking in Frontier AI”: Power-seeking defined as behaviors where AI systems acquire resources, evade oversight, or resist termination beyond task requirements is identified as a key driver of Loss of Control (LoC) risk. In this work, we introduce SysAdmin, a benchmark that positions…

Benchmarks1 source
Sources & timeline
arXiv cs.AISysAdmin: Measuring Instrumental Power-Seeking in Frontier AI
arXiv cs.CL1 source

On the Computational Complexity of Structural Generalization

Public information on “On the Computational Complexity of Structural Generalization”: Structural generalization has been measured repeatedly by several benchmarks, yet it has never been formally defined. We give a definition that translates the two premises (compositional structure and unbounded generalization) into mathematical language. The d…

Why it mattersMay add technical evidence for future radar tracking: On the Computational Complexity of Structural Generalization

Benchmarks1 source
Sources & timeline
arXiv cs.CLOn the Computational Complexity of Structural Generalization
Ollama1 source

v0.32.3

Public information on “v0.32.3”: What's Changed - Fixed model downloads that stall before sending data. - Improved integrations: restored Claude Code Channels, fixed Anthropic thinking streams, and made Hermes Desktop respect `--force-build`. - Expanded GPU support with CUDA on Windows ARM64,…

Infrastructure1 source
Sources & timeline
Ollamav0.32.3

Wed, Jul 22

llama.cpp1 source

b10091

Public information on “b10091”: <details open> ci : fix SYCL package shared library lookup (#25987) </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linu…

Open source1 source
Sources & timeline
llama.cppb10091
llama.cpp1 source

b10090

Public information on “b10090”: <details open> webgpu : add CONV_2D_DW (depthwise conv2d) kernel (#25847) * webgpu : add CONV_2D_DW (depthwise conv2d) kernel Implement GGML_OP_CONV_2D_DW for the WebGPU backend, ported from the Vulkan backend's conv2d_dw.comp. Assisted-by: Claude Opus-4.8 * R…

Open source1 source
Sources & timeline
llama.cppb10090
llama.cpp1 source

b10089

Public information on “b10089”: <details open> cuda: GET_ROWS quants (#25962) * cuda: add k-quant support to GET_ROWS Device-side embedding lookups require GET_ROWS to handle the k-quants used by common GGUF recipes (Q4_K_M stores token_embd as q6_K). Without it the backend rejects the op an…

Open source1 source
Sources & timeline
llama.cppb10089
OpenAI Python SDK1 source

v2.47.0

Public information on “v2.47.0”: 2.47.0 (2026-07-21) Full Changelog: v2.46.0...v2.47.0 Features * **client:** Add experimental runtime support for HTTPX2 clients (#3524) (317260c) * **stlc:** configurable CI runner and private-production-repo support in workflow templates (4303e97) Bug Fixes…

Open source1 source
Sources & timeline
OpenAI Python SDKv2.47.0
Anthropic News1 source

Supporting ambitious external research through the Anthropic Economic Futures Research Fund

Public information on “Supporting ambitious external research through the Anthropic Economic Futures Research Fund”: We’re committing $200 million to the Anthropic Economic Futures Research Fund to support ambitious external research.

Benchmarks1 source
Sources & timeline
Anthropic NewsSupporting ambitious external research through the Anthropic Economic Futures Research Fund
Anthropic Python SDK1 source

v0.118.0

Public information on “v0.118.0”: 0.118.0 (2026-07-22) Full Changelog: v0.117.1...v0.118.0 Features * **api:** add support for Managed Agents model effort, initial session events, and threads delta streaming (712bc6f)

Agents1 source
Sources & timeline
Anthropic Python SDKv0.118.0
Microsoft Foundry Blog1 source

Design the Network Before You Deploy: Best Practices for Microsoft Foundry Standard Agents BYOVNet | Microsoft Community Hub

Public information on “Design the Network Before You Deploy: Best Practices for Microsoft Foundry Standard Agents BYOVNet | Microsoft Community Hub”: Ask any enterprise team standing up Microsoft Foundry in production and you'll hear the same non-negotiable: the agent can't run on the open internet. The

Agents1 source
Sources & timeline
Microsoft Foundry BlogDesign the Network Before You Deploy: Best Practices for Microsoft Foundry Standard Agents BYOVNet | Microsoft Community Hub
arXiv cs.CL1 source

Convolution for Large Language Models

Public information on “Convolution for Large Language Models”: Large language models (LLMs) largely rely on Transformers, where self-attention provides global token interaction but does not explicitly encode the locality of natural language. We study whether lightweight depthwise convolutions can supply this local inducti…

Benchmarks1 source
Sources & timeline
arXiv cs.CLConvolution for Large Language Models
arXiv cs.CL1 source

A Classifier That Teaches Itself: Self-Improving, Frozen-gate Training (SIFT) for Dynamic Document Classification

Public information on “A Classifier That Teaches Itself: Self-Improving, Frozen-gate Training (SIFT) for Dynamic Document Classification”: Document classification is a solved problem in the laboratory and an unsolved one in the enterprise. The blocker is rarely model architecture; it is the labeling project that must precede a model and the institutional fear of letting a model retrain itself onc…

Business1 source
Sources & timeline
arXiv cs.CLA Classifier That Teaches Itself: Self-Improving, Frozen-gate Training (SIFT) for Dynamic Document Classification

Tue, Jul 21

Anthropic Python SDK1 source

v0.117.1

Public information on “v0.117.1”: 0.117.1 (2026-07-21) Full Changelog: v0.117.0...v0.117.1 Bug Fixes * **aws:** handle credentials correctly when using AnthropicAWS.copy() (85d3881) Chores * **api:** add support for new refusal category (d1dea0b) * **client:** docs updates (b14f94c) * **deps:*…

Open source1 source
Sources & timeline
Anthropic Python SDKv0.117.1
The Verge AI1 source

OpenAI says it accidentally hacked Hugging Face with a new AI system

Public information on “OpenAI says it accidentally hacked Hugging Face with a new AI system”: OpenAI says its AI models mistakenly breached open-source AI platform Hugging Face during internal testing. In a blog post on Tuesday, OpenAI writes that GPT-5.6 Sol and "an even more capable pre-release model" discovered vulnerabilities within their sandboxed…

Safety1 source
Sources & timeline
The Verge AIOpenAI says it accidentally hacked Hugging Face with a new AI system
TechCrunch AI1 source

Jack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents

Public information on “Jack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents”: Buzz is a group chat platform for the workplace that puts humans and their AI agents in the same conversation.

Agents1 source
Sources & timeline
TechCrunch AIJack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents
Google Gemini Blog · Ars Technica AI3 sources

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Public information on “Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber”: We’re introducing new Gemini models, including Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber.

Model release3 sources
Sources & timeline
Google Gemini BlogIntroducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash CyberArs Technica AIGoogle announces Gemini 3.6 Flash and cybersecurity AI, teases 3.5 Pro and Gemini 4Google DeepMind BlogIntroducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash CyberGoogle DeepMind BlogIntroducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
NVIDIA Generative AI Technical Blog1 source

NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI

Public information on “NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI”: Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context, interact with

Agents1 source
Sources & timeline
NVIDIA Generative AI Technical BlogNVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI
NVIDIA Generative AI Technical Blog1 source

Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI

Public information on “Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI”: What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale.

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogInside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI
NVIDIA Generative AI Technical Blog1 source

Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72

Public information on “Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72”: Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token

Infrastructure1 source
Sources & timeline
NVIDIA Generative AI Technical BlogSetting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72
MIT Technology Review AI1 source

Advancing next-gen AI with materials science innovation

Public information on “Advancing next-gen AI with materials science innovation”: The conversation about AI often centers on algorithms, computing power, or huge investments in new semiconductor fabrication plants and hyperscale data centers. But beneath each of these advances is another layer of innovation that makes them possible: advance…

Infrastructure1 source
Sources & timeline
MIT Technology Review AIAdvancing next-gen AI with materials science innovation
OpenAI News1 source

OpenAI and Hugging Face partner to address security incident during model evaluation

Public information on “OpenAI and Hugging Face partner to address security incident during model evaluation”: OpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders.

Safety1 source
Sources & timeline
OpenAI NewsOpenAI and Hugging Face partner to address security incident during model evaluation
arXiv cs.AI1 source

Design and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions

Public information on “Design and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions”: Soft, sensorized companions offer a physically safe and emotionally intuitive interface for socially assistive technologies, yet their deformability and multichannel tactile sensing complicate the robust interpretation of human affect. This study presents a co…

Research1 source
Sources & timeline
arXiv cs.AIDesign and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions
arXiv cs.LG1 source

Fully-sensorized smart-eyewear platform for on-device Machine Learning

Public information on “Fully-sensorized smart-eyewear platform for on-device Machine Learning”: This paper presents ARGO, a smart eyewear platform designed to bridge ergonomic comfort, high computational throughput, and energy efficiency. Unlike cloud-dependent solutions, ARGO leverages the STM32N6 microcontroller and its integrated Neural Processing Uni…

Research1 source
Sources & timeline
arXiv cs.LGFully-sensorized smart-eyewear platform for on-device Machine Learning
arXiv cs.AI1 source

Some Large Language Models Exhibit Consistent Risk Attitudes

Public information on “Some Large Language Models Exhibit Consistent Risk Attitudes”: As artificial intelligence systems are deployed in open-ended, high-stakes settings, a critical dimension remains unmeasured: how perceived risk is translated into action. We test whether large language models (LLMs) exhibit systematic and consistent risk atti…

Agents1 source
Sources & timeline
arXiv cs.AISome Large Language Models Exhibit Consistent Risk Attitudes
arXiv cs.LG1 source

DocOCR-Eval: A Correction-Based Framework for OCR Tool Selection Without Ground Truth

Public information on “DocOCR-Eval: A Correction-Based Framework for OCR Tool Selection Without Ground Truth”: Document parsing is a foundational step for document understanding tasks such as visual question answering and key information extraction, as it transforms unstructured scanned images into structured representations by extracting textual, visual, and layout in…

Benchmarks1 source
Sources & timeline
arXiv cs.LGDocOCR-Eval: A Correction-Based Framework for OCR Tool Selection Without Ground Truth
arXiv cs.AI1 source

Rater State Bias in RLHF Preference Data: An Audit Framework

Public information on “Rater State Bias in RLHF Preference Data: An Audit Framework”: We identify a structured confound in Reinforcement Learning from Human Feedback (RLHF). Pairwise preference labels are intended to reflect the compared outputs, but they may also reflect the rater's state during annotation. Under sustained stressful or distres…

Research1 source
Sources & timeline
arXiv cs.AIRater State Bias in RLHF Preference Data: An Audit Framework
arXiv cs.LG1 source

Reinforcement Learning-Guided NSGA-II Enhanced with Gray Relational Coefficient for Multi-Objective Optimization: Application to NASDAQ Portfolio Optimization

Public information on “Reinforcement Learning-Guided NSGA-II Enhanced with Gray Relational Coefficient for Multi-Objective Optimization: Application to NASDAQ Portfolio Optimization”: In modern financial markets, decision-makers increasingly rely on quantitative methods to navigate complex trade-offs among multiple, often conflicting objectives. This paper addresses constrained multi-objective optimization (MOO) with an application to portf…

Why it mattersMay add technical evidence for future radar tracking: Reinforcement Learning-Guided NSGA-II Enhanced with Gray Relational Coefficient for Multi-Objective Optimization: Application to NASDAQ Portfolio Optimization

Agents1 source
Sources & timeline
arXiv cs.LGReinforcement Learning-Guided NSGA-II Enhanced with Gray Relational Coefficient for Multi-Objective Optimization: Application to NASDAQ Portfolio Optimization
arXiv cs.CV1 source

Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection

Public information on “Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection”: AI-generated videos (AIGVs) typically contain subtle temporal artifacts that arise from inter-frame inconsistencies rather than within individual frames. A detector that captures such artifacts should therefore benefit from video pretrained backbones over imag…

Why it mattersMay add technical evidence for future radar tracking: Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection

Benchmarks1 source
Sources & timeline
arXiv cs.CVRethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection
arXiv cs.CL1 source

Committed Before Reasoning: Behavioral Reproduction and Preliminary refresh-Level Evidence of Answer Pre-Commitment in an Open-Weight LLM

Public information on “Committed Before Reasoning: Behavioral Reproduction and Preliminary refresh-Level Evidence of Answer Pre-Commitment in an Open-Weight LLM”: Chat models sometimes commit to an answer and then produce reasoning that justifies it rather than deriving it -- even when the answer contradicts a task premise. We study a minimal probe: "I want to wash my car. The car wash is 100 meters away. Should I walk…

Research1 source
Sources & timeline
arXiv cs.CLCommitted Before Reasoning: Behavioral Reproduction and Preliminary refresh-Level Evidence of Answer Pre-Commitment in an Open-Weight LLM
arXiv cs.CL1 source

Multi-level context Modeling for consistent expert selection in Mixture-of-Experts

Public information on “Multi-level context Modeling for consistent expert selection in Mixture-of-Experts”: Mixture-of-Experts (MoE) enables efficient scaling of Transformer models by routing tokens to a small subset of experts. However, existing routers typically condition expert selection on shallow or isolated token representations, which often produce unstable a…

Why it mattersMay add technical evidence for future radar tracking: Multi-level context Modeling for consistent expert selection in Mixture-of-Experts

Benchmarks1 source
Sources & timeline
arXiv cs.CLMulti-level context Modeling for consistent expert selection in Mixture-of-Experts
arXiv cs.CL1 source

RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation

Public information on “RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation”: Small-scale language models (SLMs) are attractive for retrieval-augmented generation (RAG) in resource-constrained settings, but their limited capacity makes them highly sensitive to noisy or spurious retrieved evidence. Existing preference-based methods such…

Benchmarks1 source
Sources & timeline
arXiv cs.CLRIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation
arXiv cs.CV1 source

Training-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark

Public information on “Training-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark”: Generalized few-shot 3D point-cloud segmentation (GFS-PCS) asks a model to segment a scene into many base classes seen at training time and a set of novel classes. The state of the art reaches novel classes by reconciling a dense but noisy 3D vision-language p…

Benchmarks1 source
Sources & timeline
arXiv cs.CVTraining-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark
OpenAI News1 source

David Vélez and Robin Vince join the boards of the OpenAI Foundation and OpenAI Group PBC

Public information on “David Vélez and Robin Vince join the boards of the OpenAI Foundation and OpenAI Group PBC”: David Vélez and Robin Vince join the boards of the OpenAI Foundation and OpenAI Group PBC, bringing global leadership in finance, technology, and governance.

Safety1 source
Sources & timeline
OpenAI NewsSafety and alignment in an era of long-horizon modelsOpenAI NewsDavid Vélez and Robin Vince join the boards of the OpenAI Foundation and OpenAI Group PBC

Mon, Jul 20

Ollama1 source

v0.32.2

Public information currently provides only the title and page metadata for “v0.32.2”. Review the original source for details.

Open source1 source
Sources & timeline
Ollamav0.32.2
MIT Technology Review AI1 source

China’s AI models have Trump’s AI world at war with itself

Public information on “China’s AI models have Trump’s AI world at war with itself”: This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Over the weekend, several current and former advisors to President Donald Trump on AI publicly lobbed insults at the count…

Opinion1 source
Sources & timeline
MIT Technology Review AIChina’s AI models have Trump’s AI world at war with itself
Anthropic News1 source

Apply for Anthropic’s AI for Science rare disease research grants

Public information on “Apply for Anthropic’s AI for Science rare disease research grants”: Anthropic is sharing a focused call for AI for Science applications centered specifically on rare genetic diseases. Accepted applicants will receive up to $50,000 in Claude credits over six months, with the goal of building a community of researchers looking i…

Safety1 source
Sources & timeline
Anthropic NewsApply for Anthropic’s AI for Science rare disease research grants
MIT Technology Review AI1 source

AI is more likely than humans to form biases when hiring

Public information on “AI is more likely than humans to form biases when hiring”: The next time you apply for a job, AI may screen your résumé before any human sees it. But there’s good reason to question whether AI will judge you fairly. Researchers already know that LLMs pick up human biases from their training data. New research suggests…

Regulation1 source
Sources & timeline
MIT Technology Review AIAI is more likely than humans to form biases when hiring
arXiv cs.AI1 source

Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction

Public information on “Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction”: Causal and intervention-based question answering is fundamental to advancing large language models (LLMs) toward reasoning beyond surface-level correlations and understanding underlying causal mechanisms. However, existing LLM-based methods often rely on impli…

Benchmarks1 source
Sources & timeline
arXiv cs.AICausal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction
arXiv cs.AI1 source

GraphDx: A Cost-Aware Knowledge-Enhanced Multi-Agent Framework for Sequential Diagnosis

Public information on “GraphDx: A Cost-Aware Knowledge-Enhanced Multi-Agent Framework for Sequential Diagnosis”: Sequential diagnosis requires balancing diagnostic accuracy against resource costs through iterative information gathering. Existing Large Language Model (LLM) approaches exhibit a critical knowledge-reasoning gap: despite encoding extensive medical knowledge,…

Agents1 source
Sources & timeline
arXiv cs.AIGraphDx: A Cost-Aware Knowledge-Enhanced Multi-Agent Framework for Sequential Diagnosis
arXiv cs.AI1 source

Cura 1T: Specialized Model for Agentic Healthcare

Public information on “Cura 1T: Specialized Model for Agentic Healthcare”: Healthcare spans high-stakes communication, expert reasoning, and workflow execution, yet specialized LLMs that cover these use cases together remain limited. A healthcare model must handle patient consultation, clinical reasoning over text and images, interac…

Agents1 source
Sources & timeline
arXiv cs.AICura 1T: Specialized Model for Agentic Healthcare
arXiv cs.CL1 source

Large Language Models as Unified Multimodal Learners for Clinical Prediction

Public information on “Large Language Models as Unified Multimodal Learners for Clinical Prediction”: Electronic health records combine free-text clinical narratives with structured measurements such as vital signs, laboratory values, and comorbidities. Yet most clinical prediction systems still rely on task-specific fusion architectures, pairing dedicated enc…

Research1 source
Sources & timeline
arXiv cs.CLLarge Language Models as Unified Multimodal Learners for Clinical Prediction
arXiv cs.CL1 source

VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs

Public information on “VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs”: The key-value (KV) cache is the main memory bottleneck in long-context large language model (LLM) inference. Two leading training-free families are both structurally limited: token-selection methods (SnapKV, Ada-KV) score importance from an observation window…

Research1 source
Sources & timeline
arXiv cs.CLVarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs
arXiv cs.CL1 source

Verbalizable Representations Form a Global Workspace in Language Models

Public information on “Verbalizable Representations Form a Global Workspace in Language Models”: Out of everything the human brain processes, only a small fraction is consciously accessible, in the sense of being available for verbal report, deliberate control, and flexible reasoning. In this paper, we present evidence that an analogous functional distinc…

Research1 source
Sources & timeline
arXiv cs.CLVerbalizable Representations Form a Global Workspace in Language Models
Coatue1 source

Our Partnership with Databricks

Coatue published a blog post reflecting on its partnership with Databricks. Since its initial investment in 2019, Databricks has evolved from a Spark-based data processing platform into a critical context layer for enterprises embedding AI into workflows. The post outlines key chapters of Databricks' evolution: an open foundation built on Iceberg, a unified lakehouse architecture, and an interface for building agentic applications on enterprise data. Core primitives include Unity Catalog for governance and AI Gateway for managing token spend and agent governance. Databricks has also reoriented itself like a research lab, compressing development timelines for projects like query engines from 3-4 years to 6-9 months.

Opinion1 source
Sources & timeline
CoatueOur Partnership with Databricks

Sun, Jul 19

Sat, Jul 18

Fri, Jul 17

MIT Technology Review AI1 source

The risk of weather data sabotage is rising

Public information on “The risk of weather data sabotage is rising”: Every morning, airline dispatchers, grid operators, and farmers around the world make decisions based on the same thing: a weather forecast. While these forecasts are something that most people glance at for two seconds, weather predictions influence major str…

Regulation1 source
Sources & timeline
MIT Technology Review AIThe risk of weather data sabotage is rising

Thu, Jul 16

Anthropic Python SDK1 source

v0.117.0

Public information on “v0.117.0”: 0.117.0 (2026-07-16) Full Changelog: v0.116.0...v0.117.0 Features * **api:** add support for dreaming (642eee7) * **api:** add support for MCP Tunnels (d716df6) Bug Fixes * **credentials:** keep credential material out of traceback frame locals via SecretStr (…

Open source1 source
Sources & timeline
Anthropic Python SDKv0.117.0
VentureBeat AI1 source

The AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs

Public information on “The AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs”: Across 107 enterprises, AI infrastructure spending is accelerating well ahead of the ability to see or steer its economics. Most organizations run their AI on a familiar base of hyperscalers and model-provider APIs, yet the next dollar is aimed at specialized…

Benchmarks1 source
Sources & timeline
VentureBeat AIThe AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs
VentureBeat AI1 source

The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials

Public information on “The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials”: Across 107 enterprises, AI agents are being given real access to systems and data while the controls meant to contain them lag behind. More than half have already had a confirmed agent security incident or a near-miss; only about a third give every agent its o…

Safety1 source
Sources & timeline
VentureBeat AIThe agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials
VentureBeat AI1 source

The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix

Public information on “The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix”: Across 101 enterprises, the infrastructure that feeds AI agents their business context is being built faster than it can be trusted. Retrieval-augmented generation is already the default context source, and provider-native retrieval has quietly overtaken the d…

Agents1 source
Sources & timeline
VentureBeat AIThe AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix
Microsoft Foundry Blog1 source

Answers You Can Trust: Grounding Enterprise Agents with Foundry IQ | Microsoft Community Hub

Public information on “Answers You Can Trust: Grounding Enterprise Agents with Foundry IQ | Microsoft Community Hub”: Enterprise AI has a trust problem, and it is not about fluency. A model will answer almost any question with confidence, whether or not it actually knows.

Agents1 source
Sources & timeline
Microsoft Foundry BlogAnswers You Can Trust: Grounding Enterprise Agents with Foundry IQ | Microsoft Community HubMicrosoft Foundry BlogAgents can learn with Memory in Microsoft Foundry Agent Service | Microsoft Community Hub
Ollama1 source

v0.32.1

Public information on “v0.32.1”: What's Changed - Improved Gemma 4 tool calling and multi-turn reasoning, including more reliable tool-response continuations - Fixed a recurrent MLX model cache leak that could increase memory use across requests, and improved cache snapshot performance - MLX…

Agents1 source
Sources & timeline
Ollamav0.32.1

Wed, Jul 15

TechCrunch AI1 source

Microsoft is reportedly training salespeople to talk down OpenAI and Anthropic

Microsoft is reportedly training its salespeople to pitch its own AI models as more efficient and cost-effective than those of competitors OpenAI and Anthropic.

Why it mattersReveals a competitive shift in Microsoft's sales strategy, potentially affecting enterprise adoption of models from OpenAI and Anthropic, two major LLM providers. Highlights growing tension between partnership and competition in the AI ecosystem.

Business1 source
Sources & timeline
TechCrunch AIMicrosoft is reportedly training salespeople to talk down OpenAI and Anthropic
NVIDIA Generative AI Technical Blog1 source

Build a Multi-Camera 3D Tracking Application with NVIDIA DeepStream 9.1 Skills

NVIDIA released a tutorial on building a multi-camera 3D tracking application using DeepStream 9.1. The application addresses the challenge of tracking the same object across multiple camera views in large spaces, going beyond single-camera 2D tracking.

Product update1 source
Sources & timeline
NVIDIA Generative AI Technical BlogBuild a Multi-Camera 3D Tracking Application with NVIDIA DeepStream 9.1 Skills
VentureBeat AI1 source

Agentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents

VentureBeat Pulse Research survey of 101 enterprises finds AI agent orchestration consolidating onto model-provider platforms, with Anthropic's Claude leading at 40% primary platform share. However, there is a significant gap between ambition and reality: 71% report that a quarter or fewer of their deployed 'agents' are true multi-step orchestrated workflows, with most being chatbot wrappers. To avoid vendor lock-in (35% fear as top risk), 51% expect a hybrid control plane by end of 2026, while only 6% prefer provider-managed. Fiscal control lags, with 27% lacking real-time cost stop mechanisms. The survey is a single-wave, self-selected sample from June 2026, directional only.

Why it mattersMay add technical evidence for future radar tracking: Agentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents

Agents1 source
Sources & timeline
VentureBeat AIAgentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents
The Verge AI1 source

xAI sues a man for using Grok to generate CSAM ‘deepfakes’

Elon Musk's xAI is suing a South Carolina man, Terry Wayne Harwood, for allegedly using the Grok AI chatbot to generate and distribute child sexual abuse material (CSAM). The lawsuit claims he knowingly circumvented safeguards, altered nonconsensual images, and generated CSAM. The Verge reported on July 15, 2026, citing Reuters.

Why it mattersThis lawsuit represents a significant legal action by an AI company against a user for generating illegal content, highlighting challenges in AI safety and content moderation. It could influence future policies and liability frameworks for generative AI.

Regulation1 source
Sources & timeline
The Verge AIxAI sues a man for using Grok to generate CSAM ‘deepfakes’
TechCrunch AI · Ars Technica AI2 sources

Amid hardware legal battle, OpenAI releases a $230 keyboard for Codex

Amid a legal battle with Apple over hardware trade theft allegations, OpenAI has released a $230 light-up keyboard designed for use with its agentic coding app Codex.

Why it mattersOpenAI launching a dedicated hardware accessory for its agentic coding app amid a legal dispute with Apple signals hardware ambitions and potential IP battles. The $230 price point targets professional developers.

Product update2 sources
Sources & timeline
TechCrunch AIAmid hardware legal battle, OpenAI releases a $230 keyboard for CodexArs Technica AIOpenAI's first branded hardware is... a light-up keyboard?
Hugging Face Transformers1 source

Release v5.14.0

Hugging Face Transformers released v5.14.0, adding two new models: Inkling and TIPSv2. Inkling, from Thinking Machines, is a multimodal model with 975B total and 41B active parameters, accepting text, image, and audio inputs and generating text outputs. It supports multiple natural and coding languages, and is designed for agentic/tool-use systems, coding assistants, chatbots, retrieval-augmented generation, and general instruction-following, with open weights. TIPSv2 is mentioned but its description is truncated in the raw text.

Why it mattersRelease of a 975B-parameter open-weight multimodal model (Inkling) from Thinking Machines, integrated into Hugging Face Transformers, enabling broad access for research and development. Also includes TIPSv2, but description is truncated.

Open source1 source
Sources & timeline
Hugging Face TransformersRelease v5.14.0
The Verge AI1 source

Suno snatched millions of songs from YouTube, Genius, and Deezer

A hacking incident revealed that AI music generator Suno trained on millions of songs and lyrics scraped from YouTube Music, Deezer, and Genius, as reported by 404 Media. Suno had not previously disclosed its training data sources.

Why it mattersExposes that Suno trained on scraped copyrighted music without disclosure, highlighting legal and ethical concerns in AI music generation. May influence ongoing copyright lawsuits and regulatory scrutiny.

Regulation1 source
Sources & timeline
The Verge AISuno snatched millions of songs from YouTube, Genius, and Deezer
OpenAI News · MIT Technology Review AI2 sources

Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer

Public information on “Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer”: OpenAI has built an LLM super-hacker called GPT-Red that it uses as a sparring partner to help its other models boost their defenses against cyberattacks. Last week the company released the latest version of its flagship LLM, GPT-5.6. OpenAI says that training…

Safety2 sources
Sources & timeline
OpenAI NewsGPT-Red: Unlocking Self-Improvement for RobustnessMIT Technology Review AIMeet GPT-Red: an LLM super-hacker OpenAI built to make its models safer
OpenAI News1 source

The US is advancing AI safety through state and federal action

OpenAI outlines a 'reverse federalism' approach to AI governance, where state laws help build a national framework for safe, democratic AI.

Why it mattersOpenAI's advocacy for 'reverse federalism' in AI governance could influence how US federal and state regulations develop, impacting the safety and democratic control of AI systems. This signals a proactive corporate stance in shaping policy.

Safety1 source
Sources & timeline
OpenAI NewsThe US is advancing AI safety through state and federal action
arXiv cs.LG1 source

OmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes

OmniPM-Net is a fusion model based on Convolutional Conditional Neural Processes (ConvCNP) that reconciles discrete station and gridded PM10 forecasts. It uses terrain-aware Gaussian set convolution to lift irregular GNN station forecasts onto a regular grid, blends them with CAMS forecasts via multi-scale Spatial Source Attention, and decodes into consistent predictions at stations or grid cells over a 108h horizon. Evaluated across 1,618 stations in China over the full year of 2024, OmniPM-Net matches the station-level accuracy of the stronger GNN baseline (MAE 21.14 vs 22.00 µg/m³), reduces CAMS MAE by 30%, and provides gridded fields that discrete GNNs cannot. Gains are clearest in the high-concentration tail (90th percentile MAE -9% vs GNN, -25% vs CAMS) and during dust episodes.

Why it mattersMay add technical evidence for future radar tracking: OmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes

Research1 source
Sources & timeline
arXiv cs.LGOmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes
arXiv cs.CV1 source

Anomalous Frame Detection Using VLM-Based Description Comparison for Extracting Expert-Specific Actions and Contextual Decision-Making Scenes with Intra-Video Self-Similarity

This paper proposes an anomalous frame detection method using VLM-based description comparison to extract expert-specific actions and contextual decision-making scenes from task videos. In 27 simulated distribution board maintenance scenarios, the method achieves extraction rates of 65% for action candidates and 61% for decision-scene candidates, improving over conventional methods (59% and 33%).

Why it mattersThis research addresses the critical challenge of transferring expert maintenance knowledge to less experienced workers, which is vital for infrastructure safety. The proposed VLM-based method improves detection of expert-specific actions and contextual decision-making scenes, with a notable boost in decision-scene extraction (61% vs 33% conventional). This could automate knowledge capture from video, aiding training and decision support in industrial settings.

Research1 source
Sources & timeline
arXiv cs.CVAnomalous Frame Detection Using VLM-Based Description Comparison for Extracting Expert-Specific Actions and Contextual Decision-Making Scenes with Intra-Video Self-Similarity
arXiv cs.CL1 source

G-SHARE: A Guideline-Based Structured Reasoning Framework for Human-Factor Event Diagnosis

G-SHARE is a structured reasoning framework for human-factor event diagnosis in nuclear power plants. It operationalizes the CNNP nine-step guideline into evidence extraction, stepwise diagnostic reasoning, and consistency repair. Evaluated on a dataset of real reports, it outperforms one-shot LLM prompting and traditional ML baselines, achieving higher accuracy and macro-F1. Structured reasoning and consistency enforcement are found to be critical.

Why it mattersDemonstrates a structured reasoning approach for safety-critical AI, outperforming naive LLM use, with potential to improve human-factor analysis in nuclear industry.

Research1 source
Sources & timeline
arXiv cs.CLG-SHARE: A Guideline-Based Structured Reasoning Framework for Human-Factor Event Diagnosis
arXiv cs.CV1 source

TSCA-Net: Temporal-Spatial Clique Attention for Interpretable Multimodal Pedestrian Trajectory Prediction

TSCA-Net proposes a temporal-spatial clique attention network for multimodal pedestrian trajectory prediction, featuring three modules: TSCA (learnable temporal gating in clique-based goal-history interaction), CPCP (asymmetric pairwise agent relationships via dynamic clique potential), and AKGR (adaptive KAN-LSTM decoder grid refinement based on goal distribution entropy). It achieves state-of-the-art performance on ETH/UCY (ADE/FDE 0.13/0.20 m) and SDD (6.95/10.43 pixels).

Why it mattersAdvances state-of-the-art in pedestrian trajectory prediction, a critical component for safe autonomous navigation in crowded environments. The adaptive mechanisms and novel attention could improve robustness and efficiency.

Research1 source
Sources & timeline
arXiv cs.CVTSCA-Net: Temporal-Spatial Clique Attention for Interpretable Multimodal Pedestrian Trajectory Prediction
arXiv cs.CL1 source

CANDI: Contextual Alignment for Niche Domains Question Answering

This paper introduces CANDI-QA, a dataset for evaluating LLMs on context-sensitive question answering in niche domains (e.g., medical, financial). It consists of expert-curated QA pairs in two categories: Information Assistance (factual extraction) and Applied Inference (multi-hop reasoning). Over ten LLMs are evaluated, and a neuro-symbolic baseline MTSS-Net is proposed. Findings indicate current LLMs struggle with contextual alignment without enhanced integration.

Why it mattersCANDI-QA provides a benchmark for context-sensitive QA in niche domains, highlighting limitations of current LLMs and promoting research into trustworthy AI for high-stakes fields like medical diagnostics and financial advisory.

Benchmarks1 source
Sources & timeline
arXiv cs.CLCANDI: Contextual Alignment for Niche Domains Question Answering
arXiv cs.CV1 source

GenDiff: A Dose and Anatomy Aware Diffusion Model with Structural Prior Refinement for Low-Dose CT Reconstruction and Generalization

GenDiff is a generalizable diffusion-based framework for low-dose CT reconstruction that jointly models continuous radiation dose and anatomical information. It integrates a Dose-Anatomy Encoder, dose- and anatomy-conditioned cold diffusion backbone, physics-consistency update, and Structural Prior Refinement Module (SPRM). Experiments on multi-anatomy clinical datasets, including unseen ultra-low-dose conditions and out-of-distribution datasets, show it outperforms state-of-the-art methods.

Why it mattersGenDiff addresses a major limitation in low-dose CT reconstruction by jointly modeling continuous dose levels and anatomical variations, enabling robust generalization to unseen ultra-low-dose conditions and out-of-distribution data. This could significantly improve clinical adoption of low-dose CT, reducing patient radiation exposure while maintaining diagnostic image quality.

Research1 source
Sources & timeline
arXiv cs.CVGenDiff: A Dose and Anatomy Aware Diffusion Model with Structural Prior Refinement for Low-Dose CT Reconstruction and Generalization
arXiv cs.LG1 source

Semidirect Fourier Delta Attention: Phase-Controlled Delta Memory with Constructive Chunk-WY Kernels

This paper introduces Semidirect Fourier Delta Attention (SFDA), a generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control. The main theoretical result is a constructive chunk-WY factorization enabling exact affine chunk transfer, formal stability and complexity bounds, and a compact characterization of phase-plus-low-rank memory. Experiments on toy state-tracking tasks show SFDA learns cyclic memory while the phase-disabled KDA baseline remains near chance. Fused kernels and large-scale language-model comparisons are left to future work.

Why it mattersLinear attention mechanisms compress memory into fixed states, limiting long-context tracking. SFDA introduces phase-controlled Fourier memory, which could significantly improve memory capacity without large overhead, with a constructive chunk-WY factorization ensuring stability and bounded rank growth. Toy experiments show it learns cyclic patterns where the baseline fails, but scaling to LMs remains unverified.

Research1 source
Sources & timeline
arXiv cs.LGSemidirect Fourier Delta Attention: Phase-Controlled Delta Memory with Constructive Chunk-WY Kernels
arXiv cs.LG1 source

Repairing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry

This paper addresses the shape-prior shortcut problem in single-shot fringe projection profilometry (FPP) networks by introducing PhiCalNet, which outputs a wrapped-phase representation and maps it to depth via a fixed differentiable calibration layer, architecturally removing the shortcut. On a synthetic benchmark, PhiCalNet reduces object MAE from 14.54 mm to 4.46 mm (3.3x improvement) and introduces the first pixel-wise conformal uncertainty quantification for FPP.

Why it mattersAddresses the shortcut learning problem in fringe projection profilometry by introducing PhiCalNet, which uses a wrapped-phase representation and fixed calibration layer to architecturally prevent shortcuts, achieving 3.3× MAE improvement and the first pixel-wise conformal uncertainty quantification for FPP. This approach may generalize to other physics-based vision tasks where shortcuts can degrade performance.

Research1 source
Sources & timeline
arXiv cs.LGRepairing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry
arXiv cs.CL1 source

Scaling Point-in-Time Language Models

This paper shows that the performance gap between point-in-time language models and their temporally unrestricted counterparts can be substantially narrowed through scale. The authors train decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens, producing monthly checkpoints from 2013 to 2024. On reasoning and understanding benchmarks, the models approach the performance of similar-size open models like Gemma-3-4B and LLaMA-7B, though a gap remains. Instruction fine-tuning via LoRA improves downstream usability, and the full pipeline is released.

Why it mattersPoint-in-time language models prevent lookahead bias, critical for valid backtesting in finance and causal inference. Scaling them to 4B parameters and releasing a full pipeline narrows the performance gap with unrestricted models, making temporally valid LLMs practical for real-world applications.

Open source1 source
Sources & timeline
arXiv cs.CLScaling Point-in-Time Language Models

Tue, Jul 14

Ars Technica AI1 source

Lawsuit claims Meta's layoff decisions were made by AI, not humans

A lawsuit claims Meta's layoff decisions were made by AI rather than humans; Meta denies using AI to terminate workers with disabilities or medical issues.

Why it mattersIf true, this would be a significant case of AI being used for employment decisions, raising ethical and legal concerns about algorithmic bias, particularly affecting workers with disabilities. It highlights the need for regulation and transparency in AI-driven HR processes.

Regulation1 source
Sources & timeline
Ars Technica AILawsuit claims Meta's layoff decisions were made by AI, not humans
OGX1 source

v1.2.1

Public information on “v1.2.1”: What's Changed * fix(ci): regenerate uv.lock for ogx-client 1.2.0 on release-1.2.x by @cdoern in https://github.com/ogx-ai/ogx/pull/6300 * fix(ci): generate ogx-client in-repo instead of cloning external repo (backport #6297) by @mergify[bot] in https://github…

Open source1 source
Sources & timeline
OGXv1.2.1
NVIDIA Generative AI Technical Blog1 source

Lessons From the Leaderboard: What 5,000+ Kagglers Taught Us About Improving AI Reasoning

This article summarizes lessons from the NVIDIA Nemotron Model Reasoning Challenge, where over 5,000 Kagglers explored techniques to improve AI reasoning accuracy. The text is truncated, so full content is unavailable.

Why it mattersProvides actionable insights from a large-scale community effort to improve AI reasoning accuracy, with potential implications for model development and evaluation.

Benchmarks1 source
Sources & timeline
NVIDIA Generative AI Technical BlogLessons From the Leaderboard: What 5,000+ Kagglers Taught Us About Improving AI Reasoning
Ars Technica AI1 source

US military sent explosive drone boats into combat for the first time

The US military used explosive drone boats in combat for the first time, striking an Iranian naval port as war escalates again.

Why it mattersThis marks the first confirmed combat use of explosive uncrewed surface vessels by the US military, indicating advancing autonomous weapon capabilities and potential escalation in naval warfare. It highlights the integration of AI-driven targeting and navigation systems in active conflict zones.

Safety1 source
Sources & timeline
Ars Technica AIUS military sent explosive drone boats into combat for the first time
NVIDIA Generative AI Technical Blog1 source

How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo

This NVIDIA blog post introduces how to run an autoresearch workflow using RL agent skills and NVIDIA NeMo, highlighting that coding AI agents can now handle long-running ML workflows by inspecting repositories, setting up runtimes, and resolving issues.

Why it mattersShows NVIDIA's advance in making AI agents operable for research workflows using NeMo and RL.

Regulation1 source
Sources & timeline
NVIDIA Generative AI Technical BlogHow to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo
OpenAI News1 source

How to manage AI investments in the agentic era

OpenAI's official blog post discusses how enterprises can manage AI investments in the agentic era, focusing on measuring useful work per dollar, improving efficiency, and scaling high-value workflows.

Why it mattersThis article from OpenAI, a leading AI company, provides guidance for enterprises on measuring and optimizing AI investments in the agentic era. As agentic AI systems become more prevalent, understanding how to allocate resources effectively is critical for business adoption and ROI. It signals a shift towards more pragmatic, efficiency-focused AI deployment strategies.

Agents1 source
Sources & timeline
OpenAI NewsHow to manage AI investments in the agentic era
vLLM1 source

v0.25.1

vLLM v0.25.1 is a patch release with two bug fixes: deferring TorchCodec's FFmpeg import error to runtime to avoid blocking model startup, and guarding mixed-dtype allreduce + RMSNorm + static-quantization fusions to prevent corrupted hidden states.

Why it mattersFixes a startup-blocking import error for TorchCodec when FFmpeg is absent and prevents silent hidden-state corruption in mixed-dtype allreduce+RMSNorm+quant fusions, improving reliability for production inference.

Infrastructure1 source
Sources & timeline
vLLMv0.25.1
SGLang1 source

v0.5.15.post1

SGLang v0.5.15.post1 is a patch release focusing on GLM 5.2 fixes, including IndexShare in PD disaggregation and context parallel settings, plus fixes for DSA model launching on non-CUDA/HIP devices, flashinfer dependency on CUDA 12 images, and NaN outputs from flashinfer trtllm FP4 MoE kernels on long inputs.

Why it mattersThis patch release fixes critical issues in SGLang, including NaN outputs from flashinfer trtllm FP4 MoE kernels on long inputs and GLM 5.2 IndexShare bugs in PD disaggregation and context parallel settings, ensuring stable inference for large models.

Infrastructure1 source
Sources & timeline
SGLangv0.5.15.post1
llama.cpp1 source

b9994

llama.cpp release b9994 adds Q2_0 support for the Metal backend (PR #25419). The KleidiAI-enabled build for macOS Apple Silicon is disabled (PR #23780).

Why it mattersEnables extreme quantization (Q2_0) on Apple GPUs for lower memory usage, making larger models more accessible on Macs. Disabling KleidiAI on macOS suggests stability concerns.

Infrastructure1 source
Sources & timeline
llama.cppb9994
arXiv cs.CL1 source

Index SLM Technical Report

Bilibili released Index-1.9B, a series of open small language models including Base (1.9B params, pretrained on 2.8T tokens), Pure, Chat, and Character. The Base model achieves 64.92 average on benchmarks, competitive with larger models. Pre-training uses Warmup-Stable-Decay LR schedule and Norm-Head output layer. Controlled studies reveal insights on model depth, LR, data quality, and an unexplained benchmark surge. Models and code are open-sourced.

Benchmarks1 source
Sources & timeline
arXiv cs.CLIndex SLM Technical Report
arXiv cs.LG1 source

AuditWeave: A Tamper-Evident, Auditor-Navigable Evidence Layer for AI-Assisted and Data-Transformation Workflows

AuditWeave is a lightweight Python library that records AI-assisted and data-transformation workflows into an append-only, hash-chained ledger for auditability and tamper detection, with overhead of tens of microseconds per event.

Why it mattersAuditWeave addresses the critical need for tamper-evident, auditor-friendly records in AI-assisted and data-transformation workflows, especially in regulated domains like auditing, finance, and healthcare. Its lightweight design and verifiable integrity fill a gap between operational monitoring and compliance evidence.

Research1 source
Sources & timeline
arXiv cs.LGAuditWeave: A Tamper-Evident, Auditor-Navigable Evidence Layer for AI-Assisted and Data-Transformation Workflows
arXiv cs.CL1 source

CLIR-Bench: Benchmarking Multimodal Question Answering over Irregular Clinical Time Series

CLIR-Bench is a benchmark for multimodal question answering over irregular clinical time series, constructed from de-identified ICU records. It contains 6,600 QA instances covering 11 clinical variables, organized into 4 capability dimensions and 11 tasks. Experiments show that existing generalist models struggle to retrieve and reason over sparse clinical evidence, highlighting the need for stronger irregular time-series reasoning methods.

Why it mattersIntroduces the first benchmark specifically for multimodal QA over sparse, irregular clinical time series, enabling rigorous evaluation of evidence grounding in real-world medical AI.

Benchmarks1 source
Sources & timeline
arXiv cs.CLCLIR-Bench: Benchmarking Multimodal Question Answering over Irregular Clinical Time Series
arXiv cs.AI1 source

Faithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent

This paper investigates how message format (free NL, precision-instructed NL, JSON, triples, key-value) affects information fidelity in multi-hop LLM agent relays. Using a controlled relay testbed with programmatically generated atomic facts re-encoded over six hops, the study finds that format effects are tier-dependent: under faithful-relay instructions, strong relays are nearly lossless, with minimal impact from format or cognitive load; weak relays (1.5B) show 8.7x larger spread in six-hop recall across formats, with JSON's fixed-key schema providing drift resistance at an encoding cost; injected errors persist to the final hop in 83-100% of chains across all formats without collateral damage. Structure provides a faithful, error-localizing channel, not error correction, and format choice should follow the weakest relay in the pipeline.

Why it mattersAs multi-agent LLM systems become prevalent, understanding how message formats impact information fidelity over sequential relays is critical. This paper bridges conflicting literatures by showing effects are tier-dependent, guiding format selection to the weakest relay in the pipeline. Findings that strong models are nearly lossless under faithful instructions, while weak models benefit from structured formats like JSON, have direct practical implications for designing robust agent architectures.

Agents1 source
Sources & timeline
arXiv cs.AIFaithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent
arXiv cs.AI1 source

Format Sensitivity Index: Token-Controlled Prompt Wrapper Robustness and Schema Compliance in LLM Benchmarking

This paper introduces the Format Sensitivity Index (FSI) and Parseability Sensitivity Index (PSI) to quantify how prompt wrapper formatting variations affect LLM performance and output compliance. Across 140,000 OpenRouter generations, FSI varied by over 30x across models, and parseability was a strong predictor of accuracy. The authors argue that reporting accuracy without wrapper variance and compliance is statistically fragile, and provide recommendations for benchmarking and structured-output deployments.

Why it mattersHighlights a critical, often overlooked factor in LLM benchmarking: prompt wrapper formatting can drastically alter leaderboard outcomes. Introduces measurable indices (FSI, PSI) to quantify this sensitivity, pushing the community toward more robust evaluation practices.

Research1 source
Sources & timeline
arXiv cs.AIFormat Sensitivity Index: Token-Controlled Prompt Wrapper Robustness and Schema Compliance in LLM Benchmarking
arXiv cs.CV1 source

Unified Backbone Refinement for Diffusion Models via Internal-Latent Analysis

This paper presents DUNE, a training-free refinement framework for diffusion models that detects abrupt deviations in deep latents and applies backbone-specific suppression, improving fidelity and reducing hallucinations.

Why it mattersThe paper presents DUNE, a training-free framework that detects and suppresses abrupt fluctuations in deep latents to improve diffusion model fidelity and reduce hallucinations. This approach is backbone-agnostic (works with U-Net and Transformers) and requires no retraining, potentially lowering the barrier to high-quality image generation and offering new insights into where and when to control denoising backbones.

Research1 source
Sources & timeline
arXiv cs.CVUnified Backbone Refinement for Diffusion Models via Internal-Latent Analysis
arXiv cs.CV1 source

RSLoRA: Training-free Rank Allocation for LoRA via Representational Sensitivity Probing

RSLoRA proposes a training-free rank allocator for LoRA, leveraging refresh-space geometry and virtual representational probing to outperform existing methods like AdaLoRA and GoRA.

Why it mattersRSLoRA eliminates the need for iterative, gradient-based rank allocation during LoRA fine-tuning, using a representation-sensitivity probing mechanism instead. This drastically reduces computational overhead while outperforming state-of-the-art allocators like AdaLoRA and GoRA, making efficient and robust adaptation of large models more accessible.

Research1 source
Sources & timeline
arXiv cs.CVRSLoRA: Training-free Rank Allocation for LoRA via Representational Sensitivity Probing
arXiv cs.AI1 source

From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation

This paper proposes a framework for medical image diagnosis using the Toulmin model of argumentation to decompose ML-based retinal diagnosis into components: claim, grounds, warrant, qualifier, rebuttal, and backing. Grounds are provided by a biomarker extraction model, warrant analyzed by a MedGemma agent, qualifier determined via quantitative evaluation, and rebuttal constructed using MedSigLip image similarity. The output is presented to human experts for informed assessment.

Why it mattersThis framework enhances interpretability of ML-based medical diagnosis by structuring justifications using the Toulmin model of argumentation, potentially improving trust and enabling human experts to critically assess ML predictions with evidence.

Agents1 source
Sources & timeline
arXiv cs.AIFrom ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation
arXiv cs.LG1 source

Knowledge Graphs Meet Graph Neural Networks: A Comprehensive Survey

This paper proposes a novel two-level taxonomy for GNN-based knowledge graph technologies, covering the full pipeline (construction, embedding, reasoning, applications) and categorizing by GNN models like GCN, GAT, HGNN. It reviews various models, analyzes advantages and limitations, and discusses open challenges and future directions.

Why it mattersThis survey provides a comprehensive taxonomy and systematic review of GNN-based knowledge graph technologies, filling a gap in the literature and guiding future research.

Research1 source
Sources & timeline
arXiv cs.LGKnowledge Graphs Meet Graph Neural Networks: A Comprehensive Survey
arXiv cs.LG1 source

Position: Every Ground Truth is a Human Construction, not an Objective Truth

This position paper on arXiv argues that ground truth datasets in machine learning are not objective measurements but human-technological constructions. It advocates for acknowledging their situated and contingent nature to improve reliability, transparency, and accountability.

Why it mattersChallenges the presumed objectivity of ML ground truths, advocating for increased transparency and situated reliability, which could improve model trustworthiness and accountability.

Research1 source
Sources & timeline
arXiv cs.LGPosition: Every Ground Truth is a Human Construction, not an Objective Truth
arXiv cs.CV1 source

Cross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization

This paper introduces WiCAT, a multi-subject model for widefield calcium imaging that uses self-supervised pretraining and atlas-grounded tokenization to achieve cross-subject, cross-task, and cross-dataset transfer, and enables zero-shot continuous behavior decoding and brain region reconstruction on unseen subjects.

Why it mattersFirst multi-subject foundation model for widefield calcium imaging, enabling zero-shot behavior decoding and cross-subject generalization, which could accelerate neuroscience research and inspire new AI architectures for spatiotemporal data.

Research1 source
Sources & timeline
arXiv cs.CVCross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization
arXiv cs.CL1 source

RouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation

The paper 'RouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation' was published on arXiv cs.CL on July 14, 2026. It introduces the RouteRec framework, comparing request-level hard selection with item-level learned aggregation over four traditional and one LLM-based reranker agents. On the MovieLens-1M dataset, under a leakage-free 5-fold out-of-fold protocol, hard selection underperforms BM25 (HR@10=0.223 vs 0.254), while a cheap-only learned aggregation variant matches BM25 in HR and has a higher NDCG point estimate (0.123 vs 0.114). Gated all-agent aggregation achieves HR@10=0.295 but requires 70.2% LLM calls. The key lesson is that request-level selection is too coarse; item-level aggregation is a more promising direction.

Why it mattersThis paper provides empirical evidence that for recommender systems with multiple heterogeneous agents (including LLM-based rerankers), request-level hard selection is inferior to item-level learned aggregation under a strict evaluation protocol. It shows that a cheap-only aggregation variant matches BM25 performance while gated aggregation improves recall but at increased computational cost. These findings guide practitioners in building efficient hybrid recommenders, offering a framework to balance quality and cost, particularly when integrating LLMs.

Research1 source
Sources & timeline
arXiv cs.CLRouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation
TechCrunch AI1 source

Already rich, already successful, why the last wave of tech winners is grinding again

Successful tech founders are returning to work, driven by fear of missing AI's defining moment and the allure of making even more money.

Why it mattersIndicates a behavioral trend where successful tech founders, motivated by FOMO and profit, re-enter the workforce to capitalize on AI, potentially accelerating innovation and competition in the AI sector.

Business1 source
Sources & timeline
TechCrunch AIAlready rich, already successful, why the last wave of tech winners is grinding again
TechCrunch AI1 source

Uber’s product chief on hotels, robotaxis, and why the company doesn’t want to be “everything for everyone”

Uber Chief Product Officer Sachin Kansal discusses the company's financial-services ambitions, its complicated relationship with Waymo, its new AV Labs data operation, and how AI is starting to benefit riders and drivers. The article also touches on hotels and robotaxis, but emphasizes Uber's strategy of not trying to be 'everything for everyone.'

Media / interview1 source
Sources & timeline
TechCrunch AIUber’s product chief on hotels, robotaxis, and why the company doesn’t want to be “everything for everyone”
TechCrunch AI1 source

Video-generation startup PixVerse raises $439M, valuation soars past $2B

Video-generation startup PixVerse raised $439M, pushing its valuation past $2B. The company plans to use the funds to expand its world model offering and reach customers globally.

Why it mattersPixVerse's $439M funding round at $2B+ valuation underscores the massive investment and market potential in AI video generation and world models. This signals a growing race among startups and tech giants to lead in generative video technology, which could disrupt media, entertainment, and simulation industries. The scale of funding also highlights investor confidence in AI-driven visual content creation as a near-term revenue opportunity.

Business1 source
Sources & timeline
TechCrunch AIVideo-generation startup PixVerse raises $439M, valuation soars past $2B

Mon, Jul 13

llama.cpp1 source

b9993

llama.cpp release b9993 adds support for Tencent Hunyuan 3 (Hy3/hy_v3) architecture with MTP speculative decoding. The model is an MoE decoder stack with per-head Q/K RMSNorm, sigmoid router, always-active ungated shared expert, and leading dense blocks. Implementation ported from charlie12345's fork, with blk.N.exp_probs_b stored without .bias suffix for compatibility. Also provides macOS Apple Silicon and Intel binaries.

Why it mattersIntegrates Tencent Hunyuan 3 (Hy3) MoE architecture with MTP speculative decoding into llama.cpp, broadening the range of models available for efficient local inference and potentially improving generation speed on compatible hardware.

Infrastructure1 source
Sources & timeline
llama.cppb9993
llama.cpp1 source

b9992

llama.cpp release b9992 refactors CUDA MMQ kernel configuration (#24127), including fixing Blackwell config and removing legacy code. Additionally, the macOS Apple Silicon (arm64) build with KleidiAI enabled is disabled. Provides multi-platform binary downloads.

Why it mattersThis release refactors CUDA MMQ kernel configuration, which could improve performance and maintainability of GPU-accelerated inference. It also disables KleidiAI on Apple Silicon, indicating ongoing compatibility challenges.

Infrastructure1 source
Sources & timeline
llama.cppb9992
The Verge AI · Ars Technica AI2 sources

The 6 wildest claims in Apple’s lawsuit against OpenAI

Apple sues OpenAI, alleging that during job interviews, OpenAI's hardware head asked Apple employees to bring unreleased hardware components and samples, and accusing OpenAI of stealing confidential documents and spying on hardware prototypes.

Why it mattersThis lawsuit could set a precedent for intellectual property protection in the AI industry, highlighting tensions between major tech companies and startups over trade secrets.

Regulation2 sources
Sources & timeline
The Verge AIThe 6 wildest claims in Apple’s lawsuit against OpenAIArs Technica AIApple sues OpenAI after ex-engineer allegedly used bug to steal trade secrets
Ars Technica AI1 source

Now, defenders are embracing the prompt injection, too

The article reports that defenders are using 'context bombing' as a prompt injection technique to trick hacking agents into shutting down.

Why it mattersDefenders adopting offensive prompt injection techniques like 'context bombing' represents a significant shift in AI security, potentially enabling autonomous agents to neutralize malicious bots without human intervention. This could lower the barrier to defending against AI-driven attacks and influence the design of future AI systems.

Safety1 source
Sources & timeline
Ars Technica AINow, defenders are embracing the prompt injection, too
Ars Technica AI1 source

Simulating everything, sort of: The promise and limits of world models

Based on the ingestion summary, the article discusses world models from experts, covering how they work, what they can do, and what's still unsettled.

Why it mattersWorld models are a key research area for advancing AI's ability to understand and interact with the physical world, with implications for robotics, autonomous systems, and AGI. Understanding their current promise and limits helps set realistic expectations.

Research1 source
Sources & timeline
Ars Technica AISimulating everything, sort of: The promise and limits of world models
The Verge AI1 source

Waze is getting a bunch of new AI-powered features

Waze is introducing new AI-powered features including integration of Google's Gemini assistant. Of four updates, two use Gemini. The conversation reporting feature is being updated.

Why it mattersGoogle is integrating its flagship AI assistant, Gemini, into Waze, potentially improving voice commands and personalization for millions of drivers.

Product update1 source
Sources & timeline
The Verge AIWaze is getting a bunch of new AI-powered features

Sat, Jul 11

vLLM1 source

v0.25.0

Public information on “v0.25.0”: vLLM v0.25.0 Release Notes Highlights This release features 558 commits from 232 contributors (64 new)! * **Model Runner V2 is now the default for all dense models** (#44443). Building on quantized-model support from the previous release, MRv2 is now the stand…

Infrastructure1 source
Sources & timeline
vLLMv0.25.0
Hugging Face Transformers1 source

Patch release v5.13.1

Public information on “Patch release v5.13.1”: Patch release v5.13.1 This patch is focused on enabling `transformers` for the latest release of vllm! - Be more defensive with remap_legacy_layer_types for custom models (#47245) from @hmellor - Fix custom code which doesn't know about the new linear layer ty…

Open source1 source
Sources & timeline
Hugging Face TransformersPatch release v5.13.1
Ollama1 source

v0.32.0

Public information on “v0.32.0”: What's Changed - New interactive agent experience: running `ollama` now launches an agent to help you code and delegate work ``` ❯ ollama Ollama 0.32.0 ▸ Chat, Code, & Work (glm-5.2:cloud) Chat with models, code, search the web, and delegate real work ``` - Re…

Agents1 source
Sources & timeline
Ollamav0.32.0

Fri, Jul 10

SGLang1 source

v0.5.15

Public information on “v0.5.15”: Highlights **GLM-5.2 NVFP4, tuned for production**: We took time this cycle to tune GLM-5.2 NVFP4 on Blackwell for optimized production serving. It now runs at **500+ tok/s/user on 8x B300, 450 on 4x GB300** (bs=1). Run GLM-5.2 with our cookbook. * **Spec V2 b…

Infrastructure1 source
Sources & timeline
SGLangv0.5.15
Google Gemini Blog1 source

Here’s how to make study notebooks in the Gemini app.

Public information on “Here’s how to make study notebooks in the Gemini app.”: Studying for a test, but not sure where to start? Study notebooks, a new feature in the Gemini app, can help you get organized and learn more efficiently.Think of study

Why it mattersMay add technical evidence for future radar tracking: Here’s how to make study notebooks in the Gemini app.

Research1 source
Sources & timeline
Google Gemini BlogHere’s how to make study notebooks in the Gemini app.
OGX1 source

v1.2.0

Public information on “v1.2.0”: What's Changed * fix(vertexai): use structured logging key-value style by @major in https://github.com/ogx-ai/ogx/pull/6081 * chore: bump fallback_version to 1.1.1.dev0 by @cdoern in https://github.com/ogx-ai/ogx/pull/6084 * chore: Limit payload-bearing logs t…

Infrastructure1 source
Sources & timeline
OGXv1.2.0
OpenAI News1 source

How Deutsche Telekom is rewiring telecommunications with AI

Deutsche Telekom is becoming an AI-native telco with OpenAI, transforming customer service, employee workflows, network operations, and the future of voice.

Why it mattersA major European telco publicly commits to becoming AI-native with OpenAI, signaling a strategic industry shift that could influence infrastructure, services, and competitive dynamics.

Business1 source
Sources & timeline
OpenAI NewsHow Deutsche Telekom is rewiring telecommunications with AI

Thu, Jul 9

MIT Technology Review AI · Anthropic Research2 sources

Anthropic found a hidden space where Claude puzzles over concepts

Anthropic developed a technique called the Jacobian lens, providing the clearest view yet of what happens inside large language models like Claude when answering questions or performing tasks, with findings ranging from mundane to unnerving.

Why it mattersProvides unprecedented insight into the inner workings of large language models, with potential to improve safety, alignment, and transparency. Findings may reveal concerning behaviors or gaps in understanding.

Research2 sources
Sources & timeline
MIT Technology Review AIAnthropic found a hidden space where Claude puzzles over conceptsAnthropic ResearchA global workspace in language models
OpenAI Python SDK1 source

v2.45.0

Public information on “v2.45.0”: 2.45.0 (2026-07-09) Full Changelog: v2.44.0...v2.45.0 Features * **api:** gpt-5.6-sol updates (039d1fe) Bug Fixes * **api:** restore beta resource accessors (2dfc130) Chores * retrigger release automation (7b61351)

Open source1 source
Sources & timeline
OpenAI Python SDKv2.45.0
Anthropic Research1 source

How Claude Performs on Robotics Tasks

Anthropic Research tested language models (e.g., Claude) on robotics tasks. Models controlled various robots (quadruped, arm, etc.) via interfaces: direct torque control, programmatic control, policy control (pretrained), and RL supervision. Models failed at direct low-level control but succeeded with pretrained policies or high-level commands, completing navigation and manipulation tasks. Newer models showed significant improvement but still cannot control humanoid robots without a pretrained policy.

Why it mattersReveals the current and near-future capabilities of frontier language models in robotics, with direct implications for safety, deployment, and the roadmap to embodied AI. Highlights the reliance on high-level control and the rapid improvement across model generations.

Research1 source
Sources & timeline
Anthropic ResearchHow Claude Performs on Robotics Tasks
OpenAI News1 source

GPT-5.6 is now the preferred model in Microsoft 365 Copilot

GPT-5.6 is now the preferred model in Microsoft 365 Copilot, enhancing capabilities across Word, Excel, PowerPoint, Chat, and Cowork for faster, higher-quality work.

Why it mattersGPT-5.6 becoming the preferred model in Microsoft 365 Copilot represents a major deployment of a cutting-edge AI model into widely used enterprise productivity tools. This integration enhances AI capabilities across Word, Excel, PowerPoint, Chat, and Cowork, potentially improving user efficiency and output quality at scale. It also signals a deepening partnership between OpenAI and Microsoft, with implications for the enterprise AI assistant market.

Model release1 source
Sources & timeline
OpenAI NewsGPT-5.6 is now the preferred model in Microsoft 365 Copilot

Wed, Jul 8

Anthropic Research1 source

An off switch for dual use knowledge in AI models

Anthropic, in collaboration with AE Studio, proposes GRAM (Gradient-Routed Auxiliary Modules), a method to equip AI models with removable compartments for dual-use knowledge (e.g., virology, cybersecurity). This enables surgical control over dangerous capabilities without retraining multiple models. The research is preliminary and has not been applied to Anthropic's production models.

Why it mattersGRAM offers a scalable and surgical method to control dual-use knowledge in AI models, potentially enabling safe deployment of powerful models without sacrificing beneficial access or incurring prohibitive retraining costs. This directly addresses a critical challenge in AI safety and alignment.

Safety1 source
Sources & timeline
Anthropic ResearchAn off switch for dual use knowledge in AI models
PyTorch1 source

PyTorch 2.13.0 Release

PyTorch 2.13.0 release highlights four major items: FlexAttention lands on Apple Silicon (MPS), delivering up to ~12x speedup over SDPA on sparse patterns and gaining a deterministic backward path on CUDA; the CuTeDSL "Native DSL" backend gives Inductor a second high-performance code path alongside Triton for key GPU operations with faster compilation (prototype); nn.LinearCrossEntropyLoss combines the final prediction and loss computation to cut peak GPU memory by up to 4x for large-vocabulary language model training; and torchcomms, a new communications backend for PyTorch Distributed, improves fault tolerance, scalability, and debuggability for large-cluster training.

Why it mattersThis major PyTorch release delivers high-impact features: FlexAttention on MPS (up to ~12x speedup for sparse attention), a new GPU backend (CuTeDSL) for faster compilation, nn.LinearCrossEntropyLoss reducing peak memory up to 4× for large-vocabulary LM training, and torchcomms for more robust distributed training—directly advancing efficiency, scalability, and hardware support for AI developers.

Open source1 source
Sources & timeline
PyTorchPyTorch 2.13.0 Release
NVIDIA AI Blog1 source

NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Harness

NVIDIA Nemotron 3 Ultra, paired with LangChain's Deep Agents harness, delivers benchmark-leading performance at lower cost than top closed models. It achieves the highest accuracy among open models while completing more tasks at higher throughput; the text mentions a 10x figure but is truncated, so its exact reference is unclear.

Agents1 source
Sources & timeline
NVIDIA AI BlogNVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Harness

Tue, Jul 7

Turing Post1 source

🦸🏻#14:What Is MCP? Model Context Protocol in Agentic AI, Explained

Public information on “🦸🏻#14:What Is MCP? Model Context Protocol in Agentic AI, Explained”: Alex Hancock, MCP Steering Committee member, explains how MCP turns AI models into agents that act — context discovery, OAuth 2.1, registries, and MCP vs A2A.

Agents1 source
Sources & timeline
Turing Post🦸🏻#14:What Is MCP? Model Context Protocol in Agentic AI, Explained
Microsoft Semantic Kernel1 source

dotnet-1.78.0

Public information on “dotnet-1.78.0”: Changes: * 35ba23e1b3092271c778ca057afe1a796e16e70e .Net: Update package version to 1.78.0 (#14142) * e6c9673684ca03621885083faff1644e1f42695e .Net: Disable automatic HTTP redirects in HttpPlugin and WebFileDownloadPlugin default clients (#14132) * f25753be0a1…

Open source1 source
Sources & timeline
Microsoft Semantic Kerneldotnet-1.78.0
MIT Technology Review AI1 source

The foundational elements of AI architecture that IT leaders need to scale

The article discusses foundational elements of AI architecture to help IT leaders make scalable investment decisions amid rapid AI progress and the shift to agentic systems, while managing risks. The original text is truncated, only providing the beginning.

Why it mattersProvides strategic guidance for IT leaders on building scalable AI architectures amid rapid evolution and shift to agentic systems, helping them make informed investment decisions and manage risk.

Opinion1 source
Sources & timeline
MIT Technology Review AIThe foundational elements of AI architecture that IT leaders need to scale
Microsoft Semantic Kernel1 source

python-1.44.0

Microsoft Semantic Kernel released Python version 1.44.0. This release includes dependency updates for Python packages (tornado, pyjwt, starlette, bleach, pyarrow) and .NET packages (Aspire.Azure.Search.Documents, Aspire.Hosting.Azure.CognitiveServices, and frontend dependencies).

Why it mattersThis release ensures the Semantic Kernel Python package remains up-to-date with security patches and compatible with latest dependencies, reducing vulnerability risks for AI applications built on SK.

Open source1 source
Sources & timeline
Microsoft Semantic Kernelpython-1.44.0

Mon, Jul 6

Ollama1 source

v0.31.2

Ollama v0.31.2 release: Enabled flash attention on older NVIDIA GPUs (compute capability 6.x); iGPU can offload vision models with padding; fixed structured output for thinking models when thinking disabled; hardened GGUF model creation; `ollama launch` for Claude Code now disables telemetry by default; fixed loading models on non-UTF-8 paths; updated MLX and llama.cpp engines. New contributor @kevinpark1217.

Why it mattersThis release extends flash attention support to older NVIDIA GPUs (compute capability 6.x), benefiting users with legacy hardware. It enables vision model offloading on integrated GPUs, broadening deployment options. Default telemetry opt-out for Claude Code integration improves privacy compliance. Bug fixes for structured output, non-UTF-8 paths, and hardened model creation enhance stability for production use.

Open source1 source
Sources & timeline
Ollamav0.31.2
MIT Technology Review AI1 source

Your family’s $300 stake in OpenAI

MIT Technology Review article discussing OpenAI CEO Sam Altman's promise that Americans will share in AI-generated wealth, referencing a Financial Times report. The title suggests a $300 per family stake in OpenAI, but full details are unavailable due to truncated text.

Why it mattersThe article discusses Sam Altman's proposal to distribute OpenAI's wealth to Americans, potentially giving each family a $300 stake. This could be a landmark approach to sharing AI-generated prosperity and mitigating inequality, though full details remain unclear due to truncated text.

Business1 source
Sources & timeline
MIT Technology Review AIYour family’s $300 stake in OpenAI

Fri, Jul 3

Hugging Face Transformers1 source

Release v5.13.0

Hugging Face Transformers released v5.13.0, adding architectures for KimiK 2.5, 2.6, and 2.7 based on the open-source native multimodal agentic model KimiK 2.5, which excels in long-horizon coding, coding-driven design, autonomous execution, and swarm orchestration, supporting multiple programming languages (Rust, Go, Python) and full-stack development.

Why it mattersIntegration of KimiK 2.5–2.7 into Hugging Face Transformers enables widespread adoption of advanced agentic coding models, potentially accelerating AI-driven software development.

Open source1 source
Sources & timeline
Hugging Face TransformersRelease v5.13.0

Thu, Jul 2

MIT Technology Review AI1 source

Achieving operational excellence with AI

This article explores how to achieve operational excellence with AI. It first reviews frameworks like Lean Six Sigma and Business Process Management (BPM) in bringing order to messy operations, and implies that AI can further enhance these approaches.

Why it mattersThis piece highlights the ongoing integration of AI into classic operational frameworks, which is relevant for business process optimization but offers no new insights.

Business1 source
Sources & timeline
MIT Technology Review AIAchieving operational excellence with AI

Tue, Jun 30

Ollama1 source

v0.31.1

Ollama v0.31.1 release focuses on faster Gemma 4 on Apple Silicon, achieving nearly 90% average speedup via multi-token prediction (MTP) with no configuration needed. Other changes include tightened Gemma 4 MoE model loading in the MLX engine, updated MLX engine, and updated llama.cpp engine.

Why it mattersThis update brings a near 90% average token generation speedup for Gemma 4 on Apple Silicon via multi-token prediction (MTP) with no configuration needed. It significantly improves local AI inference performance for users on Mac hardware, enhancing accessibility and responsiveness for coding-agent and similar workloads.

Infrastructure1 source
Sources & timeline
Ollamav0.31.1
Anthropic Python SDK1 source

v0.115.0

Anthropic Python SDK v0.115.0 adds API support for Managed Agents event delta streaming, agent overrides, reverse pagination, vault credential injection scoping, and agent and deployment webhook events.

Why it mattersThis release adds critical features for Managed Agents, such as event delta streaming and webhook events, which enhance real-time agent interaction and integration capabilities. It also introduces security improvements with vault credential injection scoping. These updates make Anthropic's platform more developer-friendly and enterprise-ready, supporting the growing ecosystem of AI agents.

Open source1 source
Sources & timeline
Anthropic Python SDKv0.115.0
Anthropic News1 source

Introducing Claude Sonnet 5

Anthropic introduces Claude Sonnet 5, described as the most agentic Sonnet model yet, with top-tier intelligence for coding and professional work. It offers performance close to Opus 4.8 at lower prices, with substantial improvements over Sonnet 4.6 in reasoning, tool use, coding, and knowledge work. The model is available across all plans starting today, priced at $2 per million input tokens and $10 per million output tokens. Safety evaluations show lower rates of undesirable behaviors than Sonnet 4.6, and early access partners report improved autonomy in complex tasks.

Safety1 source
Sources & timeline
Anthropic NewsIntroducing Claude Sonnet 5

Mon, Jun 29

vLLM1 source

v0.24.0

Public information on “v0.24.0”: vLLM v0.24.0 Release Notes Highlights This release features 571 commits from 256 contributors (77 new)! * **MiniMax-M3**: Added support for the new **MiniMax-M3** model (#45381), with a fast follow-on of BF16/FP8 indexer via MSA (#45892), MXFP4 support (#45896…

Infrastructure1 source
Sources & timeline
vLLMv0.24.0

Fri, Jun 26

SGLang1 source

v0.5.14

Public information on “v0.5.14”: Highlights New Model Support: GLM-5.2, LiquidAI LFM2.5, Kimi-K2.7-Code, Poolside Laguna-M.1, DiffusionGemma, Zyphra ZAYA1, MiMo-V2-ASR **DeepSeek-V4 on GB300 since Day 0**: 5x higher throughput at the same interactivity, serving DeepSeek-V4 on NVIDIA GB300 wit…

Infrastructure1 source
Sources & timeline
SGLangv0.5.14
OGX1 source

v0.5.4

Public information on “v0.5.4”: What's Changed * fix(deps): make OCI dependencies optional on release 0.5 by @skamenan7 in https://github.com/ogx-ai/ogx/pull/6193 **Full Changelog**: https://github.com/ogx-ai/ogx/compare/v0.5.3...v0.5.4

Open source1 source
Sources & timeline
OGXv0.5.4
OGX1 source

v0.5.3

OGX released v0.5.3, fixing OTel bootstrap conflicts in containers.

Why it mattersFixes OpenTelemetry bootstrap conflicts in containerized OGX deployments, improving stability for GenAI stack observability.

Open source1 source
Sources & timeline
OGXv0.5.3

Wed, Jun 24

OGX1 source

v1.1.3

OGX released v1.1.3 with two fixes: fix Vertex AI to walk routing tables to reset provider clients (backport #6148), fix pgvector to ensure vector extension exists before creating connection pool (backport #6168); and update ogx-client dependency in UI lockfile.

Why it mattersThis patch release fixes bugs in OGX that improve stability of the Vertex AI provider integration and pgvector extension. Users relying on these components should upgrade to avoid potential issues.

Infrastructure1 source
Sources & timeline
OGXv1.1.3

Thu, Jun 18

PyTorch1 source

PyTorch 2.12.1 Release, bug fix release

PyTorch 2.12.1 is a bug fix release that addresses regressions and silent correctness issues: fixes nondeterministic Flash Attention outputs on NVIDIA B200 GPUs and illegal memory access in Triton convolution2d_bwd_weight kernel on B100/B200 (sm100) GPUs by updating Triton to 3.7.1; fixes fill_ on byte-dtype views with misaligned storage offset; and drops CPython 3.13t from the binary build matrix.

Why it mattersThis patch fixes silent correctness issues and regressions in PyTorch on cutting-edge NVIDIA B100/B200 GPUs, which are critical for deterministic AI workflows and memory safety. It ensures reliable training and inference for users adopting the latest hardware.

Regulation1 source
Sources & timeline
PyTorchPyTorch 2.12.1 Release, bug fix release

Wed, Jun 17

OpenAI Python SDK1 source

v2.43.0

OpenAI Python SDK v2.43.0 released, featuring an API update to the OpenAPI spec or Stainless config.

Why it mattersRoutine SDK update to reflect changes in the underlying API specification or configuration (OpenAPI/Stainless). Keeps the Python client in sync with the latest OpenAI API surface, but introduces no user-facing features or breaking changes.

Open source1 source
Sources & timeline
OpenAI Python SDKv2.43.0
OGX1 source

v1.1.2

OGX released v1.1.2 with changes: updated ogx-client to ^1.1.1 in UI lockfile, fixed cascade delete for orphaned conversation items, and added ZIP decompression limits to MarkItDown processor.

Why it mattersMinor patch release for OGX, an open-source GenAI stack, including dependency updates, a bugfix for orphaned conversation deletion, and security limits (ZIP decompression) in MarkItDown processor. Maintains stability and security but introduces no new capabilities.

Open source1 source
Sources & timeline
OGXv1.1.2
Microsoft Semantic Kernel1 source

python-1.43.1

Public information on “python-1.43.1”: What's Changed * Python: Add function_choice_behavior support to Azure AI and OpenAI Assistant agents by @SergeyMenshykh in https://github.com/microsoft/semantic-kernel/pull/14057 * Fix MessagePack by @SergeyMenshykh in https://github.com/microsoft/semantic-ke…

Agents1 source
Sources & timeline
Microsoft Semantic Kernelpython-1.43.1

Mon, Jun 15

Hugging Face Transformers1 source

Patch release v5.12.1

Hugging Face Transformers released patch v5.12.1, updating the lower bound for PEFT and fixing the auto tokenizer for the mistral tokenizer.

Why it mattersEnsures compatibility with newer versions of PEFT and fixes tokenizer issues for Mistral models. Relevant for users who rely on these components.

Open source1 source
Sources & timeline
Hugging Face TransformersPatch release v5.12.1
OGX1 source

v1.1.1

OGX released v1.1.1 with fixes for file processor sync parsing, Milvus compatibility, VertexAI RuntimeError, and vector search error propagation.

Why it mattersOGX v1.1.1 is a patch release that addresses several stability and compatibility issues in the GenAI stack, including fixes for file processing, Milvus vector database compatibility, Vertex AI integration error handling, and vector search error propagation. While incremental, these fixes are critical for developers relying on OGX for production AI applications, ensuring smoother operation with updated dependencies.

Open source1 source
Sources & timeline
OGXv1.1.1
vLLM1 source

v0.23.0

vLLM v0.23.0 released with 408 commits from 200 contributors. Highlights: DeepSeek-V4 maturation with sparse MLA decoupling, TRTLLM kernel, etc.; Model Runner V2 becomes default for Llama and Mistral dense models, adds FlashInfer sampler. Note: Minimax M3 not yet supported.

Why it mattersvLLM is a critical open-source inference engine. This release hardens support for DeepSeek-V4, expands Model Runner V2 to Llama and Mistral dense models, and adds performance optimizations, directly influencing deployment efficiency for many AI practitioners.

Infrastructure1 source
Sources & timeline
vLLMv0.23.0

Sat, Jun 13

SGLang1 source

v0.5.13

SGLang released v0.5.13, adding support for autoregressive models (Nemotron 3 Ultra, Step-3.7-Flash, Command A+) and diffusion models (Cosmos3, LingBot-World, SANA-WM, Ernie-Image, FLUX.2-Klein 4B/9B, Ideogram 4). Spec V2 is now the default speculative-decoding path, with tree drafting (topk>1) production-ready across multiple backends.

Why it mattersSGLang v0.5.13 adds support for several new autoregressive and diffusion models, including Day-0 support for Nemotron 3 Ultra, and makes Spec V2 the default speculative-decoding path with production-ready tree drafting across multiple backends. This broadens the framework's applicability and improves inference performance.

Infrastructure1 source
Sources & timeline
SGLangv0.5.13

Thu, Jun 11

OGX1 source

v1.1.0

OGX released v1.1.0 with fixes (header sanitization, CI test fixture issues), performance improvements (parallelized health and vector store fan-out), documentation updates (1.0 release notes, blog announcement, API endpoint references), dependency cleanup (removed unused litellm), and a new OpenAI Responses schema drift checker.

Regulation1 source
Sources & timeline
OGXv1.1.0

Wed, Jun 10

Fri, May 29

OGX1 source

v0.4.6

OGX v0.4.6 release includes security fixes for multiple CVEs (CVE-2026-32597, CVE-2026-30922, CVE-2026-27628, CVE-2025-14009, CVE-2026-33236, CVE-2026-48710), dependency bumps (pyjwt, pyasn1, pypdf, nltk, starlette), and added release automation workflows.

Why it mattersAddresses multiple security vulnerabilities (CVEs) in dependencies used by OGX, a component of the llama-stack AI framework. Prompt updating improves security posture.

Safety1 source
Sources & timeline
OGXv0.4.6

Thu, May 28

OGX1 source

v0.7.2

OGX released v0.7.2 with changes: update llama-stack-client to ^0.7.1 in UI lockfile, and constrain starlette to >=1.0.1 to fix CVE-2026-48710.

Why it mattersThis release includes a security fix for CVE-2026-48710 in starlette, a critical web framework vulnerability, and updates llama-stack-client to ^0.7.1. Users should update to avoid potential exploits.

Safety1 source
Sources & timeline
OGXv0.7.2

Wed, May 27

Bessemer Venture Partners1 source

Researching the frontier of robotics: Three founders on what it takes to succeed in embodied AI

From the data pyramid to ‘dark magic’ in data attribution, here's what three founders at the cutting edge of embodied AI are watching closely—and what they think their field is missing.

Other1 source
Sources & timeline
Bessemer Venture PartnersResearching the frontier of robotics: Three founders on what it takes to succeed in embodied AI

Mon, May 25

arXiv cs.CV1 source

GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation

GEM-4D is a geometry-grounded video world model that addresses the lack of consistent point-level motion in video world models for robot manipulation by injecting dense 4D correspondence supervision distilled from a pretrained geometry foundation model. It jointly learns appearance and geometric structure while retaining a single-stream architecture with no extra inference cost. An inverse dynamics module converts correspondence-consistent video rollouts into executable robot trajectories for real-world and simulated deployment. It achieves state-of-the-art performance in video prediction and geometric consistency, improving real-world manipulation success from 61% to 81%.

Research1 source
Sources & timeline
arXiv cs.CVGEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation
arXiv cs.AI1 source

BOHM: Zero-Cost Hierarchical Attribution for Compound AI Systems

This paper presents BOHM, a zero-cost hierarchical attribution method that extracts an attribution tree from routing weights in compound AI systems, requiring no extra evaluations or internal access. It provides multi-resolution attribution and shows high correlation with Shapley-based methods at a fraction of the cost, as demonstrated on LLM, agentic, and census benchmarks.

Why it mattersMay add technical evidence for future radar tracking: BOHM: Zero-Cost Hierarchical Attribution for Compound AI Systems

Research1 source
Sources & timeline
arXiv cs.AIBOHM: Zero-Cost Hierarchical Attribution for Compound AI Systems
arXiv cs.CV1 source

Improved Vision-to-Chart Buoy Association with Learned World-to-Image Projection

This paper presents a lightweight modification to the DETR-based fusion transformer baseline for the MaCVi 2026 Vision-to-Chart data association challenge. It trains a dedicated MLP (QueryMLP) to explicitly predict buoy waterline pixel coordinates from chart measurements and IMU data, appending these to the decoder query vector as spatial priors. On the leaderboard, it achieves Overall 0.7386, F1=0.8055, mIoU=0.6718, ranking second.

Why it mattersMay add technical evidence for future radar tracking: Improved Vision-to-Chart Buoy Association with Learned World-to-Image Projection

Research1 source
Sources & timeline
arXiv cs.CVImproved Vision-to-Chart Buoy Association with Learned World-to-Image Projection
arXiv cs.AI1 source

NeuroNL2LTL: A Neurosymbolic Framework for Natural Language Translation of Linear Temporal Logic

NeuroNL2LTL is a neurosymbolic framework for translating natural language to Linear Temporal Logic (LTL). It uses an intermediate representation with structure-preserving mapping to LTL, combined with satisfiability checking and minimal-edit repair. The key innovation is verifier-in-the-loop training, where verification outcomes serve as reinforcement learning rewards, directly optimizing for formal correctness. On 200k+ requirements across domains, it achieves 28% semantic equivalence and 86% verifiable satisfiability, and generates contextual explanations for domain experts.

Why it mattersCombines neural translation with formal verification to produce LTL specifications with correctness guarantees, which is critical for deploying AI in safety-critical domains like aerospace and autonomous vehicles.

Research1 source
Sources & timeline
arXiv cs.AINeuroNL2LTL: A Neurosymbolic Framework for Natural Language Translation of Linear Temporal Logic
arXiv cs.LG1 source

Reading Calibrated Uncertainty from Language Model Trajectories

The paper 'Reading Calibrated Uncertainty from Language Model Trajectories' introduces a method to extract eleven scale-invariant geometric features from per-layer MLP updates in language models. These features are fed into a sparse linear probe to calibrate uncertainty, outperforming maximum softmax probability (MSP) by up to 21 AURC points under selective abstention. The geometric features allow interpretable tracing of error formation across layers.

Why it mattersMay add technical evidence for future radar tracking: Reading Calibrated Uncertainty from Language Model Trajectories

Research1 source
Sources & timeline
arXiv cs.LGReading Calibrated Uncertainty from Language Model Trajectories
arXiv cs.AI1 source

RMA: an Agentic System for Research-Level Mathematical Problems

This paper presents Research Math Agents (RMA), an agentic framework for automated reasoning on research-level mathematical problems. RMA decomposes proof solving into specialized modules for problem analysis, literature search and understanding, fair comparison, knowledge-bank construction, and proof verification, coordinated by initializer, proposer, and verifier agents via shared structured memory. On the First Proof benchmark of ten research-level problems, RMA solves eight, outperforming strong baselines including GPT-5.2R and Aletheia. Ablation studies show performance gains arise from interaction of structured reasoning modules, iterative refinement, and verifier feedback. Code will be released upon acceptance.

Why it mattersMay add technical evidence for future radar tracking: RMA: an Agentic System for Research-Level Mathematical Problems

Research1 source
Sources & timeline
arXiv cs.AIRMA: an Agentic System for Research-Level Mathematical Problems
arXiv cs.AI1 source

SciAtlas: A Large-Scale Knowledge Graph for Automated Scientific Research

SciAtlas is a large-scale, multi-disciplinary knowledge graph integrating over 43 million papers, 157 million entities, and 3 billion triplets. It provides a structured topological cognitive substrate and a neuro-symbolic retrieval algorithm to enable automated scientific research, including literature review, trend synthesis, idea positioning, and academic trajectory exploration.

Open source1 source
Sources & timeline
arXiv cs.AISciAtlas: A Large-Scale Knowledge Graph for Automated Scientific Research
arXiv cs.LG1 source

The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models

This paper investigates the answer readout stage of chain-of-thought (CoT) reasoning in small language models for arithmetic tasks. It identifies a positional shortcut: models copy the number occupying the trailing position before the answer delimiter, regardless of intermediate reasoning. In 1-3B instruction-tuned models on GSM8K, the presence of the gold answer accounts for 54-92 percentage points of accuracy (89-92% of the teacher-forcing ceiling). Even on incorrect items, the final answer matches the last CoT number 95-96% of the time. Replacing the trailing number with a wrong value collapses accuracy to near-zero, while removing it recovers 5-32 pp above that floor. Qwen and Llama copy novel distractors 87-95% of the time; Gemma gates selectively. Head-level ablation implicates architecture-specific head sets; the effect replicates on GSM-Symbolic. On non-arithmetic BBH tasks, shuffle retention drops sharply; at 7-8B, content-selective gating emerges. Step-level faithfulness evaluations risk conflating positional answer transport with genuine computation, posing a failure mode for CoT-based oversight.

Why it mattersMay add technical evidence for future radar tracking: The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models

Research1 source
Sources & timeline
arXiv cs.LGThe Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models
arXiv cs.AI1 source

Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems

This paper presents A-LEMS, a framework that shifts AI energy accounting from energy per inference to Energy per Successful Goal (EpG), and defines the Orchestration Overhead Index (OOI). Experiments show agentic workflows consume 4.33x higher mean energy per successful goal than linear baselines, driven by orchestration structure rather than inference compute. The authors argue energy-per-inference is insufficient for agentic AI.

Why it mattersMay add technical evidence for future radar tracking: Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems

Agents1 source
Sources & timeline
arXiv cs.AIEnergy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems
arXiv cs.LG1 source

FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning

FuRA (Full-Rank Adaptation) is proposed, an efficient full-rank fine-tuning method with spectral preconditioning. It uses block tensor-train factorization to fix pretrained SVD bases and only optimize compact cores and singular values, achieving parameter, memory, and step-time efficiency comparable to LoRA while preserving full-rank expressivity. It outperforms full fine-tuning in LLM fine-tuning (+1.37 on LLaMA-3-8B commonsense reasoning), RL for math reasoning, and visual instruction tuning. The 4-bit quantized variant QFuRA also surpasses QLoRA.

Why it mattersMay add technical evidence for future radar tracking: FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning

Research1 source
Sources & timeline
arXiv cs.LGFuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning
arXiv cs.CL1 source

Query-Adaptive Semantic Chunking for Retrieval-Augmented Generation: A Dynamic Strategy with Contextual Window Expansion

This paper proposes Query-Adaptive Semantic Chunking (QASC) for Retrieval-Augmented Generation (RAG) systems. QASC dynamically constructs chunks using three mechanisms: cosine similarity scoring between sentence and query embeddings, contextual window expansion for coherence, and chunk-level score aggregation. Evaluated on 100 technical documents and 200 queries, QASC achieves an F1-score of 0.85, a relative improvement of 18-27% over fixed chunking and 8-12% over semantic and agentic alternatives. Ablation studies and human evaluation (Cohen kappa=0.82) confirm the contribution of each component.

Why it mattersQuery-adaptive chunking significantly improves retrieval in RAG systems, addressing a key bottleneck in AI applications that rely on external knowledge retrieval.

Research1 source
Sources & timeline
arXiv cs.CLQuery-Adaptive Semantic Chunking for Retrieval-Augmented Generation: A Dynamic Strategy with Contextual Window Expansion
arXiv cs.LG1 source

Latent Cache Flow: Model-to-Model Communication Without Text

This paper introduces Latent Cache Flow (LCF), a method for model-to-model communication without text. It jointly translates and compresses KV caches, reducing adapter size to 4% of prior C2C, and handles differing contexts. Early experiments show LCF achieves 23% higher accuracy and 8.5x speedup over text-based communication.

Why it mattersAddresses a key bottleneck in LLM agent communication by replacing lossy, high-latency text exchange with direct KV cache translation, achieving 8.5× speedup and 23% accuracy gain. The drastic adapter compression (4% of prior work) makes this approach practical for heterogeneous-agent systems.

Research1 source
Sources & timeline
arXiv cs.LGLatent Cache Flow: Model-to-Model Communication Without Text
arXiv cs.CL1 source

Evaluating Large Language Models in a Complex Hidden Role Game

This paper evaluates LLMs' reasoning, persuasion, and deception capabilities in the social deduction game Secret Hitler. The author introduces an open-source framework and novel metrics, finding models ineffective at complex multi-turn manipulation: rule-based agents align with expert human voting decisions 86.7% of the time, while Llama 3.1 70B achieves only 59.7%; Chain-of-Thought prompting and internal memory fail to improve performance, with up to 23.2% worse win rates for fascist roles. The study underscores the need to detect when models master deceptive behaviors and provides a reproducible testbed for alignment research.

Benchmarks1 source
Sources & timeline
arXiv cs.CLEvaluating Large Language Models in a Complex Hidden Role Game
arXiv cs.LG1 source

FusionSense: Tri-Stage Near-Sensor Learning for Runtime-Adaptive Multimodal Edge Intelligence

This paper presents FusionSense, a fusion-aware intelligent sensing framework for energy-constrained autonomous edge systems. It employs a tri-stage near-sensor learning procedure (server-side fusion model learning, filter-out-safe label quantification, edge model compaction) to enable runtime-adaptive multimodal decisions with reduced compute and communication. On a dual-modality (RGB+Depth/LiDAR) setup, FusionSense achieves up to 33x lower energy at 1% FoI prevalence, 11x at 10%, a 92.3% reduction in quality loss at 30% data reduction, and ~1.5x higher energy savings than the best prior baseline.

Why it mattersMay add technical evidence for future radar tracking: FusionSense: Tri-Stage Near-Sensor Learning for Runtime-Adaptive Multimodal Edge Intelligence

Safety1 source
Sources & timeline
arXiv cs.LGFusionSense: Tri-Stage Near-Sensor Learning for Runtime-Adaptive Multimodal Edge Intelligence
arXiv cs.CL1 source

Knowledge Distillation for Low-Resource Open-source Text-to-SQL Model

This paper proposes a knowledge-aware Text-to-SQL framework that constructs a task-specific knowledge base including schema semantics, abbreviations, business logic, and query patterns, and injects them into both training and inference, substantially improving the performance of open-source and closed-source LLMs in low-resource domain-specific settings. Experiments on seven benchmarks demonstrate its effectiveness.

Why it mattersMay add technical evidence for future radar tracking: Knowledge Distillation for Low-Resource Open-source Text-to-SQL Model

Research1 source
Sources & timeline
arXiv cs.CLKnowledge Distillation for Low-Resource Open-source Text-to-SQL Model
arXiv cs.CV1 source

Suicide Risk Assessment from AI-powered Video Surveillance: An Interpretable Framework for Prevention in Metro Stations

This paper formalizes the task of Suicide Risk Assessment (SRA) in metro stations and introduces an interpretable AI framework that integrates person tracking, activity recognition, semantic segmentation, and trajectory-driven risk heatmap modeling. It achieves 83.2% ROC-AUC on real surveillance data.

Why it mattersMay add technical evidence for future radar tracking: Suicide Risk Assessment from AI-powered Video Surveillance: An Interpretable Framework for Prevention in Metro Stations

Research1 source
Sources & timeline
arXiv cs.CVSuicide Risk Assessment from AI-powered Video Surveillance: An Interpretable Framework for Prevention in Metro Stations
arXiv cs.CV1 source

VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding

VideoOdyssey is a benchmark for ultra-long-context and omni-modal video understanding, emphasizing continuous certificate length (the video duration a human must continuously watch to answer a question). It features extreme video durations averaging 109 minutes across 11 domains and 54 subcategories, two subsets (VideoOdyssey-V and -AV), and five granular levels from seconds to hours. Evaluations reveal that current MLLMs struggle with continuous reasoning, fine-grained perception, and non-verbal omni-modal understanding.

Why it mattersMay add technical evidence for future radar tracking: VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding

Benchmarks1 source
Sources & timeline
arXiv cs.CVVideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding
arXiv cs.CV1 source

Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?

This paper shows that removing a substantial fraction of image tokens only slightly degrades performance on a widely used hallucination benchmark, indicating current benchmarks do not reliably test fine-grained visual grounding in vision-language models (VLMs). Through global degradation, local occlusion, question reformulation, and representation-level analysis revealing increased similarity among visual tokens in deeper layers, the authors conclude that existing benchmarks are insufficient for evaluating VLMs' reliance on visual evidence.

Why it mattersMay add technical evidence for future radar tracking: Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?

Benchmarks1 source
Sources & timeline
arXiv cs.CVSeeing without Looking: Do Vision-Language Benchmarks Really Test Vision?
arXiv cs.CL1 source

How Far Will They Go? Red-Teaming Online Influence with Large Language Models

This paper introduces a red-teaming framework to measure the Overton Windows (range of political opinions reliably expressed) of open-source LLMs and quantify how simple natural-language jailbreaks expand that range. Evaluating over 30 LLMs, it finds systematic asymmetries: open-source models tend to generate left-leaning social media content, Overton Windows shrink with model size, and regional differences are substantial. Jailbreak potency varies across model families, providing a workflow for identifying effective jailbreak combinations. The work establishes a practical framework for auditing the political steerability of open-source LLMs.

Why it mattersProvides a practical framework for auditing political steerability of open-source LLMs, measuring 'Overton Windows' and jailbreak effects to counter AI-enabled influence campaigns. Critical for information integrity as LLMs increasingly participate in online discourse.

Open source1 source
Sources & timeline
arXiv cs.CLHow Far Will They Go? Red-Teaming Online Influence with Large Language Models
arXiv cs.CL1 source

A Survey of Text and Speech Resources for Hausa and Fongbe: Availability, Quality, and Gaps for NLP Development

This survey catalogs publicly available text and speech resources for Hausa (80-100M speakers) and Fongbe (~2M speakers), analyzing availability, quality, and gaps. Hausa has broader text diversity across domains; Fongbe has recent speech collection efforts. Both are in Masakhane benchmarks. Priority gaps include domain-diverse Fongbe text and dedicated Hausa speech corpora.

Why it mattersMay add technical evidence for future radar tracking: A Survey of Text and Speech Resources for Hausa and Fongbe: Availability, Quality, and Gaps for NLP Development

Research1 source
Sources & timeline
arXiv cs.CLA Survey of Text and Speech Resources for Hausa and Fongbe: Availability, Quality, and Gaps for NLP Development
OpenAI News1 source

OpenAI, Grupo Folha and Grupo UOL announce strategic content partnership

OpenAI partners with Grupo Folha and Grupo UOL to bring Brazilian journalism to ChatGPT with attribution and transparency.

Why it mattersThis partnership expands ChatGPT's access to high-quality Brazilian news content, enhancing the model's information accuracy and cultural relevance for Portuguese-speaking users. It also demonstrates OpenAI's commitment to working with publishers to provide attributed, transparent news, which is crucial for trust and combating misinformation.

Business1 source
Sources & timeline
OpenAI NewsOpenAI, Grupo Folha and Grupo UOL announce strategic content partnership

Sun, May 24

llama.cpp1 source

b9305

llama.cpp release b9305 fixes UI build by adding -fPIC and renaming a helper in cmake. Provides binaries for macOS (Apple Silicon, Intel), iOS, and Linux (various architectures, including Vulkan).

Open source1 source
Sources & timeline
llama.cppb9305

Sat, May 23

llama.cpp1 source

b9297

llama.cpp release b9297 adds NVFP4 MTP scale tensors to support Qwen3.5 MTP tensors and aligns nullptr handling. It also provides prebuilt binaries for macOS, iOS, and Linux (Ubuntu) across various architectures.

Open source1 source
Sources & timeline
llama.cppb9297
llama.cpp1 source

b9296

llama.cpp released version b9296, which fixes a bug in the ggml library where the correct iface method was not checked before falling back to 2d get (PR #23514). The release includes binaries for macOS (Apple Silicon and Intel), iOS, and Linux (Ubuntu x64/arm64/s390x, CPU and Vulkan).

Open source1 source
Sources & timeline
llama.cppb9296

Fri, May 22

Anthropic Python SDK1 source

v0.104.1

Anthropic Python SDK v0.104.1 fixes a bug in streaming where encrypted_content was not carried through the beta compaction accumulator.

Why it mattersMinor bug fix in Anthropic Python SDK streaming feature; ensures encrypted_content is properly handled during beta compaction, but impact is limited to developers using that specific feature.

Product update1 source
Sources & timeline
Anthropic Python SDKv0.104.1
arXiv cs.LG1 source

GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation

This paper proposes GraphDiffMed, a medication recommendation framework for EHRs that combines dual-scale differential attention (intra-visit and inter-visit) with pharmacological knowledge constraints (e.g., drug-drug interactions). Evaluated on MIMIC-III, it outperforms baselines in recommendation quality and safety balance, with the best configuration using only demographic features. Code is open-sourced.

Safety1 source
Sources & timeline
arXiv cs.LGGraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation
arXiv cs.CV1 source

Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models

Lens is a 3.8B-parameter text-to-image model that achieves performance competitive with or surpassing state-of-the-art models with over 6B parameters, requiring only 19.3% of the training compute used by Z-Image. Its training efficiency stems from a densely captioned dataset (Lens-800M, 109 words per caption from GPT-4.1), multi-resolution batches, and architectural choices including a semantic VAE and strong language encoder. Post-training techniques include RL with taxonomy-driven prompts, a reasoner module, and distillation for 4-step inference. Lens supports aspect ratios from 1:2 to 2:1 and resolutions up to 1440^2, generating a 1024^2 image in 3.15s on a single H100 (0.84s with turbo).

Why it mattersMay add technical evidence for future radar tracking: Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models

Research1 source
Sources & timeline
arXiv cs.CVLens: Rethinking Training Efficiency for Foundational Text-to-Image Models
arXiv cs.AI1 source

Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration

This paper proposes COSMO-Agent, a tool-augmented reinforcement learning framework to bridge the CAD-CAE semantic gap in iterative industrial design-simulation optimization. It casts CAD generation, CAE solving, result parsing, and geometry revision as an interactive RL environment, where an LLM learns to orchestrate external tools and revise parametric geometries until constraints are satisfied. Experiments show that COSMO-Agent training substantially improves small open-source LLMs for constraint-driven design in feasibility, efficiency, and stability.

Why it mattersMay add technical evidence for future radar tracking: Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration

Agents1 source
Sources & timeline
arXiv cs.AITool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration
arXiv cs.CV1 source

GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

GenEvolve proposes a self-evolving image generation agent framework via Tool-Orchestrated Visual Experience Distillation, which models generation attempts as tool-orchestrated trajectories, compares multiple trajectories to abstract structured visual experience for dense token-level supervision, achieving state-of-the-art performance on public and custom benchmarks.

Why it mattersMay add technical evidence for future radar tracking: GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

Agents1 source
Sources & timeline
arXiv cs.CVGenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation
arXiv cs.AI1 source

OSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind

The paper introduces OSCToM, an approach using RL and adversarial generation to create observer-self belief conflicts for testing LLM Theory of Mind. OSCToM-8B achieves 76% accuracy on FANToM, outperforming ExploreToM's 0.2%, and is 6x more efficient in data synthesis.

Why it mattersMay add technical evidence for future radar tracking: OSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind

Research1 source
Sources & timeline
arXiv cs.AIOSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind
arXiv cs.CL1 source

Sem-Detect: Semantic Level Detection of AI Generated Peer-Reviews

Sem-Detect is a semantic-level method for detecting AI-generated peer reviews, combining textual features with claim-level semantic analysis. It compares a target review against multiple AI-generated reviews, leveraging that AI models converge on similar points while humans are more diverse. On over 20,000 reviews from ICLR and NeurIPS, it improves TPR@0.1% FPR by 25.5% in binary setting, and misclassifies fewer than 3.5% of LLM-refined human reviews in three-class setting.

Research1 source
Sources & timeline
arXiv cs.CLSem-Detect: Semantic Level Detection of AI Generated Peer-Reviews
arXiv cs.AI1 source

SOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation

SOLAR is a self-optimizing, open-ended autonomous agent for lifelong learning and continual adaptation. It leverages parameter-level meta-learning and multi-level reinforcement learning to adapt without gradient-based fine-tuning, outperforming baselines on commonsense, mathematical, medical, coding, social, and logical reasoning tasks.

Why it mattersMay add technical evidence for future radar tracking: SOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation

Agents1 source
Sources & timeline
arXiv cs.AISOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation
arXiv cs.LG1 source

TabPFN-MT: A Natively Multitask In-Context Learner for Tabular Data

This public research item examines “TabPFN-MT: A Natively Multitask In-Context Learner for Tabular Data”. Open the original paper to review its methods and conclusions.

Why it mattersMay add technical evidence for future radar tracking: TabPFN-MT: A Natively Multitask In-Context Learner for Tabular Data

Research1 source
Sources & timeline
arXiv cs.LGTabPFN-MT: A Natively Multitask In-Context Learner for Tabular Data
arXiv cs.CL1 source

Broadening Access to Transportation Safety Data with Generative AI: A Schema-Grounded Framework for Spatial Natural Language Queries

This paper presents a schema-grounded natural language interface for transportation safety analysis, using an LLM to interpret user intent while preserving deterministic, reviewable execution. Evaluated on a Massachusetts database, all queries executed successfully, with the validation layer correcting errors in 29% of queries, suggesting that combining natural language accessibility with deterministic execution broadens access to safety data.

Why it mattersMay add technical evidence for future radar tracking: Broadening Access to Transportation Safety Data with Generative AI: A Schema-Grounded Framework for Spatial Natural Language Queries

Safety1 source
Sources & timeline
arXiv cs.CLBroadening Access to Transportation Safety Data with Generative AI: A Schema-Grounded Framework for Spatial Natural Language Queries
arXiv cs.CV1 source

PhysX-Omni: Unified Simulation-Ready Physical 3D Generation for Rigid, Deformable, and Articulated Objects

PhysX-Omni is a unified framework for simulation-ready physical 3D generation covering rigid, deformable, and articulated objects. It introduces an efficient geometry representation for Vision-Language Models, constructs the first general simulation-ready 3D dataset PhysXVerse, and proposes a benchmark PhysX-Bench. Experiments show strong performance in generation and understanding, with potential applications in scene generation and robotic policy learning.

Why it mattersMay add technical evidence for future radar tracking: PhysX-Omni: Unified Simulation-Ready Physical 3D Generation for Rigid, Deformable, and Articulated Objects

Research1 source
Sources & timeline
arXiv cs.CVPhysX-Omni: Unified Simulation-Ready Physical 3D Generation for Rigid, Deformable, and Articulated Objects
arXiv cs.LG1 source

Neural Estimation of Pairwise Mutual Information in Masked Discrete Sequence Models

This paper proposes a neural framework to estimate pairwise conditional mutual information (MI) directly from the hidden states of a pretrained masked diffusion model (MDM), supervised by ground-truth MI computed from the model's own conditional distributions. The estimator predicts the full MI matrix in a single forward pass, enabling MI-guided parallel decoding by identifying conditionally independent subsets of variables. Evaluated on Sudoku and ESM-C protein sequence generation, the method recovers known structural constraints, achieves a 3-5x reduction in inference-time forward passes while preserving generative quality, and outperforms entropy-based parallelization.

Why it mattersMay add technical evidence for future radar tracking: Neural Estimation of Pairwise Mutual Information in Masked Discrete Sequence Models

Research1 source
Sources & timeline
arXiv cs.LGNeural Estimation of Pairwise Mutual Information in Masked Discrete Sequence Models
llama.cpp1 source

b9279

Release b9279 of llama.cpp introduces a fused snake refresh kernel in the Vulkan backend, combining mul, sin, sqr, mul, and add ops into a single elementwise GPU kernel, targeting audio decoders (BigVGAN, Vocos), with added tests.

Product update1 source
Sources & timeline
llama.cppb9279
llama.cpp1 source

b9276

llama.cpp release b9276 exposes prompt token counts (n_prompt_tokens, n_prompt_tokens_processed, n_prompt_tokens_cache) in the /slots endpoint, enabling clients to monitor prompt evaluation progress.

Infrastructure1 source
Sources & timeline
llama.cppb9276
OpenAI News1 source

How Virgin Atlantic ships faster with Codex

Virgin Atlantic used OpenAI's Codex to ship its revamped mobile app on a fixed holiday travel deadline, achieving near-total unit test coverage and zero P1 defects.

Why it mattersDemonstrates tangible productivity and quality gains from Codex in a time-constrained enterprise mobile app development, supporting the narrative that AI coding tools accelerate delivery while maintaining high standards.

Product update1 source
Sources & timeline
OpenAI NewsHow Virgin Atlantic ships faster with Codex
OpenAI News1 source

OpenAI named a Leader in enterprise coding agents by Gartner

OpenAI was named a Leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents, with Codex recognized for innovation and enterprise-scale deployment.

Why it mattersOpenAI's designation as a Leader by Gartner validates its Codex platform's enterprise capabilities and competitive position, likely accelerating enterprise adoption and influencing procurement decisions.

Benchmarks1 source
Sources & timeline
OpenAI NewsOpenAI named a Leader in enterprise coding agents by Gartner

Thu, May 21

OpenAI Python SDK1 source

v2.38.0

OpenAI Python SDK v2.38.0 released, featuring API updates (manual and OpenAPI spec) and chores including docs updates and release automation changes.

Product update1 source
Sources & timeline
OpenAI Python SDKv2.38.0
Anthropic Python SDK1 source

v0.104.0

Anthropic Python SDK v0.104.0 adds support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.

Why it mattersEnables developers to estimate token consumption of thinking blocks in streaming responses, improving cost monitoring and resource planning for applications using Anthropic's advanced reasoning features.

Open source1 source
Sources & timeline
Anthropic Python SDKv0.104.0
arXiv cs.CL1 source

Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token

This study uses a BERT-based LLM for sentiment analysis of Decentraland's MANA token from Discord community, and integrates sentiment scores with multi-modal financial data (price, volume, market cap) in LSTM models for return prediction. Results show neutral sentiment with positive skew, and the multi-modal model significantly outperforms price-only baseline, demonstrating predictive value of community signals.

Why it mattersMay add technical evidence for future radar tracking: Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token

Research1 source
Sources & timeline
arXiv cs.CLLeveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token
arXiv cs.CV1 source

Why Latent Actions Fail, and How to Prevent It

This paper analyzes how exogenous state (e.g., background clutter) hinders latent action learning from unlabeled videos. By extending a linear latent action model to explicitly model exogenous state, the authors find that minimizing the standard reconstruction objective encodes exogenous information from future observations, and learning in a representation space focused on endogenous components is key to mitigating noise. Additionally, previously proposed auxiliary objectives like action-supervision provably encourage latent actions to be consistent across exogenous states. Experiments on linear and nonlinear models validate the findings.

Why it mattersMay add technical evidence for future radar tracking: Why Latent Actions Fail, and How to Prevent It

Research1 source
Sources & timeline
arXiv cs.CVWhy Latent Actions Fail, and How to Prevent It
arXiv cs.CV1 source

AI-Assisted Competency Assessment from Egocentric Video in Simulation-Based Nursing Education

This paper proposes a three-stage framework to assess learner competency from egocentric nursing simulation videos, using frozen visual encoders (DINOv2) and few-shot learning for action recognition. On 22 sessions (3.8 hours, 493 actions), it achieves 57.4% MOF in leave-one-out 1-shot recognition. The study finds a negative correlation between recognition accuracy and competency (rho = -0.524, p=0.012 for mIoU): higher-competency students exhibit more diverse and harder-to-classify workflows but more protocol-consistent transitions. This suggests recognition accuracy as a pedagogically informative signal for automated competency assessment.

Why it mattersMay add technical evidence for future radar tracking: AI-Assisted Competency Assessment from Egocentric Video in Simulation-Based Nursing Education

Research1 source
Sources & timeline
arXiv cs.CVAI-Assisted Competency Assessment from Egocentric Video in Simulation-Based Nursing Education
arXiv cs.CL1 source

Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification

This paper investigates the performance of quantized LLaMA-3.1 (8B) models in qualitative analysis, focusing on different quantization levels (2-8 bit) and types. To address hallucinations and instability in low-bit models, it proposes a quantization-aware multi-pass prompt verification method that reduces hallucinations through controlled steps. Experiments using 82 interview transcripts compare against a gold standard (BF16 model and human coding). Results show 8-bit models perform closest to the gold standard; 4-bit models become stable with the method; 3-bit and 2-bit models degrade but improve with the approach. The method enables low-resource LLMs to be more stable and accurate for qualitative research at lower cost.

Why it mattersMay add technical evidence for future radar tracking: Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification

Research1 source
Sources & timeline
arXiv cs.CLImproving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification
arXiv cs.AI1 source

Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production

This paper presents a microservice architecture for operationalizing Document AI, encapsulating pipelines of classification, OCR, and LLM-based structured field extraction in production. Key design decisions include hybrid classification, separation of GPU-bound inference from CPU-bound orchestration, asynchronous IO processing, and independent horizontal scaling. Batch profiling reveals two surprising findings: OCR dominates end-to-end latency, and system saturation is determined by shared GPU-inference capacity rather than worker count. The goal is to provide practitioners with concrete architectural patterns for production-grade document understanding systems.

Why it mattersMay add technical evidence for future radar tracking: Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production

Research1 source
Sources & timeline
arXiv cs.AIOperationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production
arXiv cs.AI1 source

Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance

This position paper advocates for developing systematic methodologies called 'data probes'—synthetic sequences generated from appropriately defined random processes—to fundamentally understand how data characteristics affect LLM performance, generalization, and robustness. The authors argue that current compute-intensive, heuristic-based approaches lack principled understanding, and propose using theoretical concepts like typical sets to analyze probe sequences, offering a pathway to foundational insights beyond empirical heuristics.

Why it mattersMay add technical evidence for future radar tracking: Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance

Research1 source
Sources & timeline
arXiv cs.AIPosition: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance
arXiv cs.AI1 source

Evaluating the Utility of Personal Health Records in Personalized Health AI

This paper evaluates LLMs (Gemini 3.0 Flash) for answering health queries using Personal Health Records (PHRs). 2,257 queries from three sources were matched with 1,945 de-identified PHRs. Gemini responses were generated with no PHR context, a basic summary, or full clinical notes. Evaluation used SHARP and a new framework for PHR-specific errors. Significant improvements in helpfulness with PHR data (p<0.001), and potential gains in safety, accuracy, relevance, and personalization. Gaps such as temporal disorientation and rare confabulations were identified. The study supports PHR data potential and provides a monitoring framework.

Why it mattersMay add technical evidence for future radar tracking: Evaluating the Utility of Personal Health Records in Personalized Health AI

Safety1 source
Sources & timeline
arXiv cs.AIEvaluating the Utility of Personal Health Records in Personalized Health AI
arXiv cs.CV1 source

Leveraging Vision-Language Models to Detect Attention in Educational Videos

This paper investigates using Vision-Language Models (VLMs) to detect attention in educational videos, but finds that prompting strategies with Gemini 3 fail to outperform statistical baselines, highlighting limitations of VLMs for real-time educational diagnostics.

Why it mattersMay add technical evidence for future radar tracking: Leveraging Vision-Language Models to Detect Attention in Educational Videos

Research1 source
Sources & timeline
arXiv cs.CVLeveraging Vision-Language Models to Detect Attention in Educational Videos

Wed, May 20

Hugging Face Transformers1 source

Release v5.9.0

Hugging Face Transformers released v5.9.0, adding three new models: Cohere2Moe (Command A+), a Mixture-of-Experts model with hybrid sliding window and full attention; Parakeet tdt; and HRM-Text, a hierarchical recurrent transformer with two stacks for slow abstract planning and fast computation.

Open source1 source
Sources & timeline
Hugging Face TransformersRelease v5.9.0
arXiv cs.LG1 source

Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance

This paper proposes a dimensional balance framework that uses spatial and temporal entropy diagnostics to harmonize feature representations via low-rank matrix embedding and extended temporal horizon, achieving substantial accuracy gains on urban traffic, meteorological, and epidemic datasets.

Why it mattersMay add technical evidence for future radar tracking: Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance

Research1 source
Sources & timeline
arXiv cs.LGDimensional Balance Improves Large Scale Spatiotemporal Prediction Performance
arXiv cs.CV1 source

Harnessing Self-Supervised Features for Art Classification

This paper systematically investigates the effectiveness of self-supervised features for artwork classification and retrieval, using DINO and CLIP models. Results show consistent improvements with self-supervised backbones, and insights into real-world applications such as VR museum navigation are provided.

Why it mattersMay add technical evidence for future radar tracking: Harnessing Self-Supervised Features for Art Classification

Research1 source
Sources & timeline
arXiv cs.CVHarnessing Self-Supervised Features for Art Classification
arXiv cs.LG1 source

HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models

HELLoRA is a parameter-efficient fine-tuning method for Mixture-of-Experts (MoE) models that attaches LoRA modules only to the most frequently activated experts per layer, reducing trainable parameters and adapter FLOPs while improving downstream performance. Evaluated on OlMoE, Mixtral, and DeepSeekMoE, it outperforms vanilla LoRA with significantly fewer parameters and higher accuracy and training throughput.

Why it mattersMay add technical evidence for future radar tracking: HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models

Research1 source
Sources & timeline
arXiv cs.LGHELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models
arXiv cs.CV1 source

MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation

MotionMERGE is a unified framework that achieves fine-grained human motion editing, reasoning, and generation by explicitly modeling motion at part and temporal levels within a single LLM. It introduces ReasoningAware Granularity-Synergy pre-training and curates a large-scale dataset MotionFineEdit (837K atomic + 144K complex triplets) with fine-grained spatio-temporal corrective instructions and motion-grounded chain-of-thought annotations. Extensive experiments demonstrate superior precision in motion generation, understanding, and editing, as well as compelling zero-shot generalization.

Why it mattersMay add technical evidence for future radar tracking: MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation

Benchmarks1 source
Sources & timeline
arXiv cs.CVMotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation
arXiv cs.CL1 source

The Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints

This paper identifies the 'Annotation Scarcity Paradox' in low-resource NLP evaluation, where model scaling outpaces sovereign human infrastructure. It reviews three phases from 2014 to present and discusses responses like data augmentation and model-based evaluation, calling for a paradigm shift to community-embedded evaluation.

Why it mattersMay add technical evidence for future radar tracking: The Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints

Research1 source
Sources & timeline
arXiv cs.CLThe Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints
arXiv cs.CV1 source

Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos

Artifact-Bench is a comprehensive benchmark for evaluating Multimodal Large Language Models (MLLMs) on detecting and analyzing artifacts in AI-generated videos. It establishes a three-level hierarchical taxonomy of realism artifacts covering photorealistic, animated, and CG-style videos, and defines three complementary tasks: real vs. AI-generated video classification, pairwise realism comparison, and fine-grained artifact identification. Experiments on 19 leading MLLMs reveal substantial limitations in artifact perception and reasoning, with many models approaching random or below-random performance in challenging settings, and significant misalignment between MLLM judgments and human perceptual preferences.

Why it mattersMay add technical evidence for future radar tracking: Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos

Benchmarks1 source
Sources & timeline
arXiv cs.CVArtifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos
arXiv cs.LG1 source

Robust Basis Spline Decoupling for the Compression of Transformer Models

This paper introduces a B-spline-based decoupling framework for compressing transformer models. It proposes a robust alternating least-squares algorithm (R-CMTF-BSD) using constrained coupled matrix-tensor factorization, achieving substantial parameter reduction while maintaining competitive accuracy on Vision and Swin Transformer architectures.

Why it mattersMay add technical evidence for future radar tracking: Robust Basis Spline Decoupling for the Compression of Transformer Models

Research1 source
Sources & timeline
arXiv cs.LGRobust Basis Spline Decoupling for the Compression of Transformer Models
arXiv cs.CL1 source

Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German

This paper presents a benchmark evaluating five commercial ASR systems on code-switching speech across four language pairs (Egyptian Arabic-English, Saudi Arabic-English, Persian-English, German-English). Each dataset contains 300 samples selected via a two-stage pipeline. ElevenLabs Scribe v2 achieved the lowest WER (13.2% overall) and highest BERTScore (0.936 overall). The authors argue BERTScore is more reliable for Arabic and Persian due to transliteration variance. The dataset is publicly available.

Why it mattersMay add technical evidence for future radar tracking: Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German

Benchmarks1 source
Sources & timeline
arXiv cs.CLBenchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German
arXiv cs.CL1 source

ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking

ReacTOD is a bounded neuro-symbolic architecture for zero-shot dialogue state tracking. It reformulates NLU as discrete tool calls within a self-correcting ReAct loop with deterministic validation. On MultiWOZ 2.1, it achieves 52.71% joint goal accuracy with gpt-oss-20B (14 points improvement) and 47.34% with Qwen3-8B. On SGD, Claude-Opus-4.6 achieves 80.68% JGA. The architecture improves accuracy by up to 9.3% over single-pass inference and achieves 93.1% self-correction rate on intercepted errors.

Why it mattersMay add technical evidence for future radar tracking: ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking

Agents1 source
Sources & timeline
arXiv cs.CLReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking
OpenAI News1 source

An OpenAI model has disproved a central conjecture in discrete geometry

An OpenAI model solved the 80-year-old unit distance problem, disproving a central conjecture in discrete geometry, marking a milestone in AI-driven mathematics.

Why it mattersAn AI model disproved a central conjecture in discrete geometry, solving the 80-year-old unit distance problem. This demonstrates AI's growing capability in advanced mathematical reasoning, potentially accelerating discoveries in pure mathematics.

Research1 source
Sources & timeline
OpenAI NewsAn OpenAI model has disproved a central conjecture in discrete geometry
OpenAI News1 source

The next phase of OpenAI’s Education for Countries

OpenAI advances its Education for Countries initiative to a new phase, expanding AI adoption in schools through new partnerships, teacher training, and tools to improve global learning outcomes.

Why it mattersOpenAI's expansion of its Education for Countries initiative with new partnerships, teacher training, and tools could significantly accelerate AI adoption in global education, potentially improving learning outcomes and digital literacy, while also raising questions about equitable access and data privacy.

Product update1 source
Sources & timeline
OpenAI NewsThe next phase of OpenAI’s Education for Countries

Tue, May 19

VentureBeat AI1 source

Google just redesigned the search box for the first time in 25 years — here’s why it matters more than you think.

At Google I/O 2026, Google redesigned its search box for the first time in 25 years, transforming it from a simple keyword field into a multimodal AI conversation interface that accepts text, images, PDFs, videos, and Chrome tabs. AI Overviews and AI Mode are merged into a single seamless experience, with an AI-powered query suggestion system. Usage stats cited: AI Mode surpassed 1 billion monthly users, AI Overviews reaches 2.5B monthly users, and AI Mode queries double every quarter. The redesign is powered by Gemini 3.5 Flash.

Product update1 source
Sources & timeline
VentureBeat AIGoogle just redesigned the search box for the first time in 25 years — here’s why it matters more than you think.
Anthropic Python SDK1 source

v0.103.1

Anthropic Python SDK v0.103.1 released, fixing a bug in the runner component where SessionToolRunner skips tool calls it does not own.

Why it mattersThis patch fixes a bug where SessionToolRunner skipped tool calls it does not own, potentially causing incorrect behavior in applications relying on Anthropic's tool calling feature. The fix is narrowly scoped and impacts only the Python SDK runner component, making it a low-importance maintenance release.

Open source1 source
Sources & timeline
Anthropic Python SDKv0.103.1
Anthropic Python SDK1 source

v0.103.0

Anthropic Python SDK v0.103.0 adds support for self-hosted sandboxes in CMA with sandbox helpers.

Why it mattersEnables developers to run self-hosted sandboxes with CMA, providing more control and flexibility.

Open source1 source
Sources & timeline
Anthropic Python SDKv0.103.0
arXiv cs.CV1 source

How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A

This paper proposes F^3A, a training-free visual token pruning router for multimodal language models, which efficiently allocates tokens under a fixed budget via task-conditioned evidence search, requiring no extra LLM forward pass.

Why it mattersMay add technical evidence for future radar tracking: How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A

Research1 source
Sources & timeline
arXiv cs.CVHow Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A
arXiv cs.LG1 source

Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra

This paper systematically optimizes real-time diffusion model inference on Apple M3 Ultra (60-core GPU, 512GB unified memory). Across 10 phases, techniques including CoreML conversion, quantization, Token Merging, and Neural Engine utilization are evaluated. The best result (22.7 FPS at 512x512) is achieved by combining CoreML-converted distilled model SDXS-512 with a three-thread camera pipeline. Key findings show that CUDA-optimization insights (e.g., quantization speedup, parallel inference) do not transfer to Apple Silicon, revealing a distinct optimization landscape and providing practical guidelines.

Why it mattersMay add technical evidence for future radar tracking: Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra

Research1 source
Sources & timeline
arXiv cs.LGSystematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra
arXiv cs.CL1 source

The Scaling Laws of Skills in LLM Agent Systems

This study analyzes 15 frontier LLMs, 1,141 real-world skills, and over 3 million routing/execution decisions, identifying two coupled scaling laws in LLM agent systems: the routing law (single-step routing accuracy decays logarithmically with library size) and the execution law (correct execution improves difficult downstream decisions by about 4×). A single parameter b couples the two laws. Law-guided optimization raises held-out routing accuracy from 71.3% to 91.7%, reduces hijack from 22.4% to 4.1%, and improves pass rates on downstream benchmarks. Results show agent performance depends not only on model capability but also on skill library structure, granularity, and exposure policy.

Why it mattersMay add technical evidence for future radar tracking: The Scaling Laws of Skills in LLM Agent Systems

Agents1 source
Sources & timeline
arXiv cs.CLThe Scaling Laws of Skills in LLM Agent Systems
arXiv cs.CL1 source

Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4

arXiv reports progress on its HTML Papers project (available since 2023), highlighting community-driven improvements, corpus-scale conversion achieving 75% error-free HTML (aiming for 90%), initial MathML 4 Intent annotations for accessibility, and a Rust port of LaTeXML for efficiency.

Infrastructure1 source
Sources & timeline
arXiv cs.CLScaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4
arXiv cs.CV1 source

Noise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model

This paper develops a probabilistic model for event cameras based on photon statistics, unifying static scene noise events and step response curves. It proposes Noise2Params, a method to determine camera-specific parameters (B, α, θ) by minimizing error against observed noise distributions, requiring only recordings of static uniform scenes. Experiments show that CNNs trained on synthetic noise data from the model outperform those trained solely on experimental data in static scene reconstruction.

Why it mattersMay add technical evidence for future radar tracking: Noise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model

Research1 source
Sources & timeline
arXiv cs.CVNoise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model
arXiv cs.CL1 source

PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures

The paper introduces PQR, a framework for automatically generating diverse and realistic user queries that elicit failures (e.g., unhelpfulness, unsafety) in LLM-based QA agents. It operates via iterative interaction between a query refinement module and a prompt refinement module, producing failure-triggering queries that resemble real user intents. Evaluated on an e-commerce QA agent, PQR uncovers 23%-78% more unhelpful responses and generates more diverse and realistic queries than previous methods.

Why it mattersMay add technical evidence for future radar tracking: PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures

Safety1 source
Sources & timeline
arXiv cs.CLPQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures
arXiv cs.CV1 source

StrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs

This paper proposes StrLoRA, a framework for Multimodal Large Language Models in Streaming Continual Visual Instruction Tuning (Streaming CVIT). Streaming CVIT is a new, more realistic setting where data arrives as continuous chunks of dynamically mixed tasks. StrLoRA uses a regularized two-stage expert routing: task-aware expert selection via textual instruction, token-wise expert weighting via cross-modal attention, and routing-stability regularization. Experiments on a new StrCVIT benchmark show StrLoRA substantially outperforms existing methods.

Open source1 source
Sources & timeline
arXiv cs.CVStrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs

Mon, May 18

arXiv cs.LG1 source

AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices

AgentStop is a lightweight efficiency supervisor for locally deployed LLM agents that predicts and terminates unlikely-to-succeed trajectories, reducing energy waste by 15-20% with minimal performance impact (<5% utility drop).

Why it mattersMay add technical evidence for future radar tracking: AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices

Agents1 source
Sources & timeline
arXiv cs.LGAgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices
arXiv cs.CV1 source

Deep Pre-Alignment for VLMs

This paper proposes Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder with a small VLM as perceiver to deeply align visual features with the text space of the target LLM. DPA improves baselines by 1.9 points on 8 multimodal benchmarks at 4B scale and 3.0 points at 32B scale, while reducing language capability forgetting by 32.9%. Gains are consistent across Qwen3 and LLaMA 3.2 families.

Why it mattersMay add technical evidence for future radar tracking: Deep Pre-Alignment for VLMs

Research1 source
Sources & timeline
arXiv cs.CVDeep Pre-Alignment for VLMs
arXiv cs.CL1 source

Fluency and Faithfulness in Human and Machine Literary Translation

This study analyzes 130,486 translated paragraphs from 106 novels in 16 source languages, including human, Google Translate, and TranslateGemma translations, and finds a consistent negative correlation between fluency and faithfulness, except for TranslateGemma where the correlation is weaker and often non-significant, suggesting a tradeoff between fluency and faithfulness in literary translation and that segment length matters for automatic evaluation.

Why it mattersMay add technical evidence for future radar tracking: Fluency and Faithfulness in Human and Machine Literary Translation

Research1 source
Sources & timeline
arXiv cs.CLFluency and Faithfulness in Human and Machine Literary Translation
arXiv cs.CV1 source

One Pass Is Not Enough: Recursive Latent Refinement for Generative Models

This paper introduces RTM, which replaces single-pass latent mapping with recursive latent refinement to improve both quality and diversity in image generation. It argues that FID is saturated and conflates fidelity with mode coverage. RTM integrated with IMLE achieves the highest precision and recall among SOTA methods on CIFAR-10, CelebA-HQ, and few-shot benchmarks, while maintaining competitive FID, and also improves StyleGAN2 variants.

Why it mattersMay add technical evidence for future radar tracking: One Pass Is Not Enough: Recursive Latent Refinement for Generative Models

Research1 source
Sources & timeline
arXiv cs.CVOne Pass Is Not Enough: Recursive Latent Refinement for Generative Models
arXiv cs.LG1 source

Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels

This study conducts a controlled empirical evaluation of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 to 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Results show that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, and models' willingness to select 'unknown' answers declines by 17.4%. Standard quality metrics like perplexity increase less than 0.5% at 8-bit and under 3% at 4-bit, yet 2.5-5.6% of items already develop new biases at 4-bit, demonstrating that aggregate metrics systematically miss fairness-critical degradation.

Benchmarks1 source
Sources & timeline
arXiv cs.LGQuantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels
arXiv cs.CV1 source

ReactiveGWM: Steering NPC in Reactive Game World Models

ReactiveGWM is a reactive game world model that decouples player controls from NPC behaviors using additive bias and cross-attention modules, enabling dynamic interactions and zero-shot strategy transfer. Evaluated on Street Fighter games, it maintains player controllability and achieves prompt-aligned NPC strategy adherence.

Why it mattersMay add technical evidence for future radar tracking: ReactiveGWM: Steering NPC in Reactive Game World Models

Research1 source
Sources & timeline
arXiv cs.CVReactiveGWM: Steering NPC in Reactive Game World Models
arXiv cs.AI1 source

SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch

This arXiv cs.AI paper introduces SDOF, a framework that models multi-agent orchestration as a constrained state machine, using an online-RLHF intent router (trained via GRPO) and a state-aware dispatcher to enforce business stage constraints. Evaluated on a recruitment system (Beisen iTalent, 6000+ enterprises), the 7B model achieves 80.9% joint accuracy on an FSM-constrained benchmark (GPT-4o: 48.9%), end-to-end task completion rate of 86.5%, and blocks all 22 injection/illegal operations. Message-level blocking achieves 100% precision and 88% recall.

Why it mattersMay add technical evidence for future radar tracking: SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch

Agents1 source
Sources & timeline
arXiv cs.AISDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch
arXiv cs.AI1 source

Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations

This study examines whether improvements in Theory of Mind (ToM) for LLMs truly benefit dynamic human-AI interactions. By proposing an interactive evaluation paradigm and systematically studying four ToM enhancement techniques, it finds that gains on static benchmarks do not necessarily translate to better performance in dynamic interactions, highlighting the need for interaction-based assessments.

Why it mattersMay add technical evidence for future radar tracking: Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations

Research1 source
Sources & timeline
arXiv cs.AIDoes Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations
arXiv cs.LG1 source

TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination

This paper identifies a compounding occupancy shift failure in sequential fine-tuning of multi-agent LLMs and proposes TeamTR, a trust-region framework that resamples trajectories and enforces per-agent divergence control, achieving 7.1% average improvement over baselines.

Agents1 source
Sources & timeline
arXiv cs.LGTeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination
arXiv cs.CL1 source

Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time

This paper introduces OP-Mix, a data mixing algorithm for the entire language model training lifecycle. It cheaply simulates candidate data mixtures by interpolating low-rank adapters trained on the current model, eliminating separate proxy models. In pretraining, OP-Mix improves average perplexity by 6.3%; in continual learning, it matches retraining and on-policy distillation while using 66% and 95% less compute, respectively.

Why it mattersMay add technical evidence for future radar tracking: Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time

Research1 source
Sources & timeline
arXiv cs.CLAlways Learning, Always Mixing: Efficient and Simple Data Mixing All The Time
arXiv cs.AI1 source

DeepSlide: From Artifacts to Presentation Delivery

DeepSlide is a human-in-the-loop multi-agent system that supports the full presentation preparation process, from requirement elicitation and time-budgeted narrative planning to evidence-grounded slide-script generation, attention augmentation, and rehearsal support. It integrates a controllable logical-chain planner, a lightweight content-tree retriever, Markov-style sequential rendering with style inheritance, and sandboxed execution. A dual-scoreboard benchmark separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while achieving larger gains on delivery metrics such as narrative flow, pacing precision, slide-script synergy, and clearer attention guidance.

Why it mattersMay add technical evidence for future radar tracking: DeepSlide: From Artifacts to Presentation Delivery

Agents1 source
Sources & timeline
arXiv cs.AIDeepSlide: From Artifacts to Presentation Delivery
arXiv cs.CL1 source

DiscoExplorer: An Open Interface for the Study of Multilingual Discourse Relations

This paper presents DiscoExplorer, an open source web interface for studying multilingual discourse relations. It makes datasets from the DISRPT Shared Task publicly available, covering 16 languages, and provides query, search, and visualization facilities for relations and signaling devices such as connectives.

Open source1 source
Sources & timeline
arXiv cs.CLDiscoExplorer: An Open Interface for the Study of Multilingual Discourse Relations

Sat, May 16

SGLang1 source

v0.5.12

SGLang v0.5.12 released on 2026-05-16, featuring full inference support for DeepSeek V4, including various parallelism (tensor, expert, context, data parallel attention), hardware support (Nvidia B300/B200/H200/H100/GB200/GB300, AMD MI35X), prefill-decode disaggregation, sparse KV cache offloading (HiSparse), reasoning and tool call parsers, custom kernels (DeepGemm, FlashMLA, MegaMoE), and post-day-0 additions: HiCache under unified Radix Tree, W4A4 and W4A8 MoE kernels, compression kernels, TP16 support, fused quantization kernel, optimized MHC+DeepGemm pipeline, non-standard chat template support, multi-detokenizer support, pipeline parallelism + PD support. Also includes a unified Docker tag lmsysorg/sglang.

Why it mattersEnables efficient, full-stack inference for the newly released DeepSeek V4 across a wide range of hardware using SGLang, a widely adopted open‑source serving framework. The extensive optimizations (multiple parallelism strategies, custom kernels, prefill‑decode disaggregation, and sparse KV cache offloading) are critical for production deployment of large models.

Infrastructure1 source
Sources & timeline
SGLangv0.5.12

Fri, May 15

OpenAI Python SDK1 source

v2.37.0

OpenAI Python SDK v2.37.0 adds service_tier parameter to responses compact method, eagerly validates pydantic iterators, removes unnecessary client_id when using workload identity provider for auth, and fixes missing f-string prefix in file type error message.

Why it mattersThis is a routine patch release for the OpenAI Python SDK that adds a minor API parameter, pydantic validation support, and auth cleanup. It is relevant to developers using the SDK but does not represent a major AI advancement.

Product update1 source
Sources & timeline
OpenAI Python SDKv2.37.0
vLLM1 source

v0.21.0

vLLM v0.21.0 release features 367 commits from 202 contributors (49 new). Highlights: formal deprecation of Transformers v4 (migrate to v5), C++20 build requirement (breaking change), KV Offload integrated with Hybrid Memory Allocator (HMA), speculative decoding with thinking budget, and new TOKENSPEED_MLA attention backend for Blackwell GPUs. New model architectures include MiMo-V2.5, Laguna XS.2, Moondream3, Qianfan-OCR, Cohere MoE, Cohere Eagle, and speculative decoding support for Mistral and Gemma4.

Why it mattersMajor release of vLLM (367 commits, 202 contributors) with breaking changes (C++20, transformers v4 deprecation), advanced memory management (KV offload + HMA), speculative decoding budgeting, and Blackwell-optimized attention for DeepSeek-R1/Kimi-K25. Widely adopted in production inference stacks.

Infrastructure1 source
Sources & timeline
vLLMv0.21.0
OpenAI News1 source

A new personal finance experience in ChatGPT

OpenAI announces a preview of a new personal finance experience in ChatGPT for Pro users in the U.S., allowing secure connection of financial accounts and providing AI-powered insights and guidance grounded in users’ financial context and goals.

Why it mattersChatGPT integrating personal finance could make AI-driven financial guidance mainstream, potentially improving financial literacy and decision-making for users. Limited to Pro users in the U.S., but signals OpenAI's expansion into sensitive, real-world domains.

Safety1 source
Sources & timeline
OpenAI NewsA new personal finance experience in ChatGPT

Thu, May 14

Ollama1 source

v0.24.0

Ollama v0.24.0 released with support for the Codex App, OpenAI's desktop experience for parallel Codex threads with built-in worktree and git functionality. It also features a built-in browser for annotation, review mode, and recommended models like kimi-k2.6, glm-5.1 for difficult tasks, and nemotron-3-super, gemma4:31b, qwen3.6 for local use.

Product update1 source
Sources & timeline
Ollamav0.24.0
Microsoft Semantic Kernel1 source

python-1.42.0

Microsoft Semantic Kernel Python version 1.42.0 released on 2026-05-14. This release includes documentation updates adding a callout for the Microsoft Agent Framework successor, and updates multiple dependencies: authlib, onnxruntime, nbconvert, boto3, python-multipart, google-cloud-aiplatform, and google-genai.

Why it mattersThe release mentions a new Microsoft Agent Framework successor in its documentation, hinting at Microsoft's evolving agent strategy. Otherwise, it includes routine dependency updates for stability and compatibility.

Open source1 source
Sources & timeline
Microsoft Semantic Kernelpython-1.42.0

Wed, May 13

OGX1 source

v1.0.2

Release v1.0.2 of ogx, featuring dependency update for ogx-client and a fix for SQL engine reset in storage.

Open source1 source
Sources & timeline
OGXv1.0.2
Ollama1 source

v0.23.4

Ollama v0.23.4 released: 'ollama launch opencode' now supports vision models with image inputs; fixed formatting of Claude tool results when using local image paths.

Why it mattersOllama v0.23.4 adds vision model support to 'ollama launch opencode', enabling image inputs for vision-capable models, and fixes a formatting bug when using local image paths with Claude tools. This benefits developers using opencode for multi-modal interactions.

Infrastructure1 source
Sources & timeline
Ollamav0.23.4
Anthropic Python SDK1 source

v0.102.0

Anthropic Python SDK v0.102.0 released, adding BetaManagedAgentsSearchResultBlock types, cache diagnostics beta support, and internal improvements (pydantic iterator validation).

Agents1 source
Sources & timeline
Anthropic Python SDKv0.102.0
OGX1 source

v1.0.1

Meta Llama Stack released v1.0.1 patch with fixes: parallelized health and vector store fan-out, asyncio.Lock for SQLStore and MongoDB expiration enforcement, hardened secret handling, async safety improvements in routers and Redis KV reads, async safety fixes for multiple providers (Databricks, WatsonX, Bing, Tavily, OCI, OpenAI Files), and handling missing collections in Milvus. Also updated ogx-client to ^1.0.0.

Safety1 source
Sources & timeline
OGXv1.0.1
PyTorch1 source

PyTorch 2.12.0 Release

PyTorch 2.12.0 release highlights: batched linalg.eigh on CUDA is up to 100x faster; new torch.accelerator.Graph API unifies graph capture and replay across backends; torch.export.save supports Microscaling (MX) quantization formats; Adagrad now supports fused=True; torch.cond control flow can be captured and replayed inside CUDA Graphs; ROCm users gain expandable memory segments and more.

Why it mattersPyTorch 2.12.0 delivers up to 100x speedup for batched eigenvalue computations on CUDA, a unified graph capture API for diverse backends, and export support for Microscaling quantization formats, enabling more efficient training and deployment of compressed models. The addition of fused Adagrad and CUDA Graph support for control flow further optimizes performance and flexibility. These enhancements directly improve developer productivity and model efficiency, making it a significant update for the AI ecosystem.

Open source1 source
Sources & timeline
PyTorchPyTorch 2.12.0 Release
Ollama1 source

v0.30.0

Ollama v0.30.0-rc22 pre-release changes architecture from GGML to direct llama.cpp support, adds GGUF compatibility, and uses MLX for Apple Silicon acceleration. Known issues: laguna-xs.2 and llama3.2-vision not supported.

Regulation1 source
Sources & timeline
Ollamav0.30.0
Ollama1 source

v0.30.0

Ollama v0.30.0 pre-release changes architecture to directly support llama.cpp instead of GGML, adds GGUF compatibility, and uses MLX for Apple Silicon acceleration. Known issues: laguna-xs.2 and llama3.2-vision not yet supported.

Regulation1 source
Sources & timeline
Ollamav0.30.0
Hugging Face Transformers1 source

Patch release v5.8.1

Hugging Face Transformers releases patch v5.8.1 primarily to fix Deepseek V4 integration, including fixes for WeightConverter regex, ContinuousBatchingManager fatal error, and Deepseek V4 CSA mask collapse.

Why it mattersFixes critical bugs in Deepseek V4 integration for Hugging Face Transformers, improving reliability for users deploying this model. Not a major feature release.

Open source1 source
Sources & timeline
Hugging Face TransformersPatch release v5.8.1

Tue, May 12

OGX1 source

v1.0.0

Release v1.0.0 of ogx-ai/ogx features new inline::auto composite file processor, inline::markitdown provider, improved OpenAI preprocessing for dict-backed reasoning messages, better error reporting for file processor rejections, redesigned provider cards and tables in docs, added 'ogx run' and 'ogx letsgo' CLI shortcuts, fixed missing dependencies, and batched guardrail checks during streaming for performance.

Open source1 source
Sources & timeline
OGXv1.0.0

Mon, May 11

Microsoft Semantic Kernel1 source

dotnet-1.76.0

Release of Semantic Kernel .NET 1.76.0 with key improvements: hardened CloudDrivePlugin defaults and path validation, improved input validation in OpenAPI plugin, support for ImageContent in tool/function results, hardened gRPC plugin address handling, updated Kiota and Snappier packages to fix NU1903 vulnerabilities, fixed DocumentPlugin path validation order, added deny-by-default AllowedUploadDirectories to CloudDrivePlugin, and fixed fallback to ToString() for logging unregistered types.

Safety1 source
Sources & timeline
Microsoft Semantic Kerneldotnet-1.76.0

Sun, May 10

vLLM1 source

v0.20.2

vLLM v0.20.2 is a small patch release with 6 commits from 6 contributors, fixing bugs for DeepSeek V4, gpt-oss, and Qwen3-VL. Fixes include: DeepSeek V4 sparse attention (re-enabled persistent topk path and fixed MTP=1 hang), DeepSeek V4 KV cache allocation failure, gpt-oss MXFP4 compatibility with torch.compile, and Qwen3-VL invalid deepstack boundary check removal.

Product update1 source
Sources & timeline
vLLMv0.20.2

Thu, May 7

Wed, May 6

OpenAI Python SDK1 source

v2.35.1

OpenAI Python SDK released v2.35.1, fixing a regression in the image generation size enum.

Why it mattersFixes a regression in the image generation size enum, ensuring stability for developers using the OpenAI Python SDK.

Product update1 source
Sources & timeline
OpenAI Python SDKv2.35.1
OpenAI Python SDK1 source

v2.35.0

OpenAI Python SDK v2.35.0 released with API updates (image 2, manual updates), removal and renaming of legacy Python CLI, and documentation update for top_logprobs parameter.

Open source1 source
Sources & timeline
OpenAI Python SDKv2.35.0
Coatue1 source

Public Markets Update Deck and Replay

Coatue shares its latest Public Markets Update deck and video replay. Philippe and team discuss the rise of the token economy, AI's impact on key sectors, and why agents represent the next major unlock for the industry.

Why it mattersCoatue is a prominent technology-focused investment firm; its public markets update provides authoritative insight into institutional perspectives on AI trends, including the token economy and AI agents, which may signal investment shifts and market opportunities.

Business1 source
Sources & timeline
CoatuePublic Markets Update Deck and Replay

Tue, May 5

SGLang1 source

v0.5.11

SGLang v0.5.11 highlights: CUDA 13 + Torch 2.11 as default; Speculative Decoding V2 enabled by default; Decode Radix Cache for PD Disaggregation; new model support (Gemma 4, GLM-5.1, etc.); DFLASH speculative decoding kernel.

Infrastructure1 source
Sources & timeline
SGLangv0.5.11
Hugging Face Transformers1 source

Release 5.8.0

Hugging Face Transformers released v5.8.0, adding DeepSeek-V4 and Gemma 4 Assistant models. DeepSeek-V4 is a next-generation MoE language model with hybrid local+long-range attention, Manifold-Constrained Hyper-Connections (mHC), and a static token-id to expert-id hash table. Gemma 4 Assistant is mentioned only by name, details incomplete.

Open source1 source
Sources & timeline
Hugging Face TransformersRelease 5.8.0

Mon, May 4

vLLM1 source

v0.20.1

vLLM v0.20.1 is a patch release focused on DeepSeek V4 stabilization and performance improvements, including base model support, multi-stream pre-attention GEMM, BF16 and MXFP8 all-to-all support, PTX cvt instruction for FP32→FP4 conversion, integrated tile kernels, and various bug fixes such as persistent topk deadlock, import error, torch inductor error, etc.

Why it mattersvLLM v0.20.1 provides crucial stabilizations and performance optimizations for DeepSeek V4, a state-of-the-art model, improving inference efficiency and reliability. As the leading open-source inference engine, these updates directly benefit developers deploying large models.

Infrastructure1 source
Sources & timeline
vLLMv0.20.1

Fri, May 1

OGX1 source

v0.8.0

OGX v0.8.0 release includes: security fix (pinning tornado>=6.5.5), vector I/O fixes (wiring file_processors provider, honoring default_search_mode config, fixing sqlite-vec BM25 score inversion), documentation improvements (comprehensive documentation overhaul, 0.7.0 release notes, MLflow observability blog post), CI changes (removing starter-gpu and dell from Docker build matrix), and new feature: native Anthropic Messages API (/v1/messages). Note: source is tagged as Meta Llama Stack but repository is ogx-ai/ogx, possibly a misattribution.

Safety1 source
Sources & timeline
OGXv0.8.0

Wed, Apr 29

Microsoft Semantic Kernel1 source

dotnet-1.75.0

Microsoft Semantic Kernel released dotnet-1.75.0 with multiple .NET and Python updates: hardened AllowedBaseUrls validation, extended InMemoryCollection filter attribute blocklist, added field/table name escaping for Python SQL Server connector, backslash escaping for Redis text search, fixed single-quote escaping in OBJECT_ID and dynamic SQL string literals, validated step types, removed MEVD components, and more.

Product update1 source
Sources & timeline
Microsoft Semantic Kerneldotnet-1.75.0

Tue, Apr 28

Hugging Face Transformers1 source

Release v5.7.0

Hugging Face Transformers released v5.7.0, adding two new models: Laguna, a mixture-of-experts language model by Poolside with per-layer head counts and sigmoid router, and DEIMv2, a real-time object detection model extending DEIM with DINOv3 features, available in eight sizes.

Why it mattersAddition of two noteworthy models to the widely-used Transformers library. Laguna introduces architectural innovations (per-layer head counts, sigmoid MoE router) for efficient LLMs, while DEIMv2 advances real-time object detection in multiple sizes. Both may influence subsequent research and applications.

Open source1 source
Sources & timeline
Hugging Face TransformersRelease v5.7.0
Coatue1 source

AI is Expanding Across the Stack

In a blog post, Coatue discusses the 2026 Forbes AI 50 list, observing that AI is expanding across the stack. Key takeaways: the enterprise application layer is seeing the most growth, infrastructure is critical for scaling, and physical AI (robotics/embodied AI) is the next frontier. Coatue highlights its 16 portfolio companies on the list, spanning foundation models, infrastructure, developer tools, enterprise applications, and physical AI.

Why it mattersThis analysis from a major VC firm signals where the AI market is maturing, with value shifting from foundation models to enterprise applications and physical AI. It provides a snapshot of investment and technology trends as captured by the Forbes AI 50 list.

Opinion1 source
Sources & timeline
CoatueAI is Expanding Across the Stack

Thu, Apr 23

Thu, Apr 16

Hugging Face PEFT1 source

v0.19.1

Hugging Face PEFT v0.19.1 is a small patch release containing fixes referenced as #3161 and #3165. Full changelog is available via the compare link.

Why it mattersHugging Face PEFT is a widely used library for parameter-efficient fine-tuning. This patch release fixes minor issues, which may be relevant to active users but lacks broader impact.

Product update1 source
Sources & timeline
Hugging Face PEFTv0.19.1

Tue, Apr 14

Hugging Face PEFT1 source

v0.19.0

Hugging Face PEFT released v0.19.0, introducing nine new PEFT methods (e.g., GraLoRA and BD-LoRA) along with numerous enhancements to make PEFT more useful.

Why it mattersPEFT v0.19.0 introduces nine new parameter-efficient fine-tuning methods, significantly expanding the toolkit for adapting large models with reduced resource consumption, thus accelerating AI deployment and research.

Open source1 source
Sources & timeline
Hugging Face PEFTv0.19.0

Thu, Apr 9

Wed, Apr 8

OGX1 source

v0.7.1

llama-stack v0.7.1 release: updates llama-stack-client dependency to ^0.7.0, adds [starter] pip extra for zero-install experience, fixes tool call arguments initialization to empty string in streaming, and improves CI to auto-bump client versions if they exist on PyPI/npm. All fixes are backports.

Open source1 source
Sources & timeline
OGXv0.7.1

Wed, Mar 25

Thu, Mar 19

Bessemer Venture Partners1 source

50 startups transforming industries with physical AI

Bessemer Venture Partners published an article identifying 50 startups transforming industries with physical AI. It highlights how AI and robotics are moving from labs to real-world applications, forecasts over 2.5 billion robots globally by 2035, and cites Waymo and Anduril Industries as leaders. The piece introduces the 50 companies as the vanguard of intelligent automation across multiple categories.

Why it mattersMay add technical evidence for future radar tracking: 50 startups transforming industries with physical AI

Research1 source
Sources & timeline
Bessemer Venture Partners50 startups transforming industries with physical AI

Fri, Jan 30

Thu, Jan 22

VentureBeat AI1 source

Railway secures $100 million to challenge AWS with AI-native cloud infrastructure

Public information on “Railway secures $100 million to challenge AWS with AI-native cloud infrastructure”: Railway , a San Francisco-based cloud platform that has quietly amassed two million developers without spending a dollar on marketing, announced Thursday that it raised $100 million in a Series B funding round, as surging demand for artificial intelligence app…

Funding1 source
Sources & timeline
VentureBeat AIRailway secures $100 million to challenge AWS with AI-native cloud infrastructure

Mon, Jan 19

VentureBeat AI1 source

Claude Code costs up to $200 a month. Goose does the same thing for free.

Public information on “Claude Code costs up to $200 a month. Goose does the same thing for free.”: The artificial intelligence coding revolution comes with a catch: it's expensive. Claude Code , Anthropic's terminal-based AI agent that can write, debug, and deploy code autonomously, has captured the imagination of software developers worldwide. But its pric…

Open source1 source
Sources & timeline
VentureBeat AIClaude Code costs up to $200 a month. Goose does the same thing for free.

Thu, Jan 15

Anthropic Research1 source

Anthropic Economic Index report: Economic primitives

This report introduces new metrics of AI usage to provide a rich portrait of interactions with Claude in November 2025, just prior to the release of Opus 4.5. The 'primitives' cover five dimensions: user and AI skills, task complexity, autonomy, success rate, and purpose. Findings include striking geographic variation, real-world estimates of AI task horizons, and a basis for revised assessments of Claude's macroeconomic impact.

Why it mattersMay add technical evidence for future radar tracking: Anthropic Economic Index report: Economic primitives

Research1 source
Sources & timeline
Anthropic ResearchAnthropic Economic Index report: Economic primitives

Wed, Jan 14

Sequoia1 source

2026: This is AGI

Sequoia's article '2026: This is AGI' argues that long-horizon agents are functionally AGI and have arrived, with coding agents as the first example. It states that 2026 will be their year, while clarifying that the definition is functional rather than technical, as the authors are investors, not researchers.

Why it mattersSignals a shift in investor perspective towards considering current AI agents as functionally equivalent to AGI, potentially influencing funding and market dynamics.

Agents1 source
Sources & timeline
Sequoia2026: This is AGI

Fri, Jan 9

Hugging Face PEFT1 source

0.18.1

Public information on “0.18.1”: Small patch release containing the following changes: - #2934: Small fixes required for some special cases to work with the upcoming transformers v5 release - #2963: Fix to enable PEFT to run with AMD ROCm thanks to @vladmandic - #2976: Fix a regression that i…

Open source1 source
Sources & timeline
Hugging Face PEFT0.18.1

Mon, Jan 5

NVIDIA AI Blog1 source

NVIDIA Rubin Platform, Open Models, Autonomous Driving: NVIDIA Presents Blueprint for the Future at CES

At CES 2026, NVIDIA presented a blueprint for the future, highlighting the Rubin platform, open models, and autonomous driving. CEO Jensen Huang, speaking at the Fontainebleau Las Vegas, declared that AI is scaling into every domain and device, and that accelerated computing and AI have fundamentally reshaped computing, affecting an estimated $10 trillion (text truncated).

Product update1 source
Sources & timeline
NVIDIA AI BlogNVIDIA Rubin Platform, Open Models, Autonomous Driving: NVIDIA Presents Blueprint for the Future at CES

Sun, Dec 28

Implications1 source

12 Outlooks for the Future: 2026+

This article from Implications presents 12 outlooks for 2026+, covering AI's impact on talent, creative industries, health, and more. Key themes include talent arbitrage for AI-native workers, a shift toward craft and proof-of-craft in content, and the implications of increased longevity for insurance and health companies.

Why it mattersThe article foreshadows significant workforce changes as AI-native talent gains advantage, a shift in Hollywood toward valuing craft and shared experiences over AI-generated content, and upcoming transformations in insurance and healthcare due to extended lifespans. These trends will impact hiring, entertainment production, and health industries.

Business1 source
Sources & timeline
Implications12 Outlooks for the Future: 2026+

Thu, Dec 11

NVIDIA AI Blog1 source

As AI Grows More Complex, Model Builders Rely on NVIDIA

This NVIDIA blog post highlights how model builders rely on NVIDIA as AI grows more complex. It mentions OpenAI's launch of GPT-5.2 in December, trained and deployed on NVIDIA Hopper and GB200 NVL72 systems, and the release of GPT-5.3 Codex in February, the first OpenAI agentic coding model to help build itself.

Infrastructure1 source
Sources & timeline
NVIDIA AI BlogAs AI Grows More Complex, Model Builders Rely on NVIDIA

Tue, Dec 2

Tue, Nov 18

Implications1 source

AI & Artistry, the Resurgence of Hardware Startups, & Unicorns that are Really Rabbits

This issue examines AI's impact on artistry, the resurgence of hardware startups as a competitive moat, and the perils of startup hypergrowth, suggesting some 'unicorns' may actually be 'rabbits'.

Why it mattersThe analysis differentiates content creators (willing to trade control for speed) from artists (unwilling), highlighting evolving dynamics in AI-assisted creation. It also discusses hardware as a new startup moat and the risks of hypergrowth, suggesting some 'unicorns' are actually 'rabbits'—fragile despite high valuations. These insights matter for understanding shifting competitive advantages and the sustainability of AI-driven business models.

Regulation1 source
Sources & timeline
ImplicationsAI & Artistry, the Resurgence of Hardware Startups, & Unicorns that are Really Rabbits

Tue, Nov 11

Stephen Wolfram1 source

What’s Special about Life? Bulk Orchestration and the Rulial Ensemble in Biology and Beyond

Stephen Wolfram shares his early findings from researching a general characterization of what happens in systems that exhibit "bulk orchestration" and discovers that there is in fact a certain universality.

Other1 source
Sources & timeline
Stephen WolframWhat’s Special about Life? Bulk Orchestration and the Rulial Ensemble in Biology and Beyond

Tue, Oct 14

Anthropic Research1 source

Preparing for AI’s economic impact: exploring policy responses

Anthropic published a research piece on policy responses to AI's economic impact, exploring nine categories of policy ideas proposed by economists and researchers, covering workforce development, permitting reform, fiscal policy, and social services, emphasizing the need to prepare strategies for various scenarios.

Why it mattersProvides a structured exploration of policy options for AI's economic impact, feeding into global discussions on workforce, infrastructure, and fiscal policy.

Opinion1 source
Sources & timeline
Anthropic ResearchPreparing for AI’s economic impact: exploring policy responses

Mon, Oct 13

Bessemer Venture Partners1 source

A B2B founder’s guide to generating demand from scratch

A B2B founder's guide to generating demand from scratch, covering timeless fundamentals adapted for today's AI-driven market. It discusses defining the ideal customer profile (ICP), iterative go-to-market (GTM) strategies, the role of product-led growth (PLG), and emphasizes the importance of middle-of-funnel (MOFU) content, recommending six core MOFU channels.

Why it mattersProvides a trusted VC's perspective on demand generation for B2B AI startups, emphasizing scalable PLG and MOFU strategies.

Opinion1 source
Sources & timeline
Bessemer Venture PartnersA B2B founder’s guide to generating demand from scratch

Tue, Sep 30

Microsoft AutoGen1 source

python-v0.7.5

Microsoft AutoGen released python-v0.7.5, with multiple fixes and improvements: fixed doc typo, Bedrock streaming tool-call empty arguments, message ID correlation, Redis caching, and GraphFlow cycle detection; added thinking mode support for the Anthropic client; fixed spurious </think> tags in streaming; added linear memory support in RedisMemory; and added GitHub Copilot development instructions.

Product update1 source
Sources & timeline
Microsoft AutoGenpython-v0.7.5

Thu, Aug 28

Sequoia1 source

Video: The $10 Trillion AI Revolution

Sequoia presents a video by Konstantine Buhler on the $10 trillion AI revolution, arguing that the Cognitive Revolution will be bigger and faster than the Industrial Revolution, with AI factories emerging in just 17 years versus 67 for the industrial engine-to-factory cycle. It cites Nvidia's Jensen Huang and questions who will be the Rockefeller and Carnegie of the AI era, betting on startups.

Why it mattersThis piece from a top VC firm highlights the accelerating pace of AI development and the belief that startups will produce the next industrial titans, signaling investment trends and market expectations.

Opinion1 source
Sources & timeline
SequoiaVideo: The $10 Trillion AI Revolution

Tue, Aug 19

Microsoft AutoGen1 source

python-v0.7.4

Microsoft AutoGen released Python package version v0.7.4 (tag python-v0.7.4). Changes include docs updates for 0.7.3, README addition of agent-as-tool, fix for Redis deserialization error, clarification that Redis does not support streaming, version bump to 0.7.4, and updated documentation. New contributor @BenConstable9 made their first contribution.

Why it mattersMinor release for Microsoft AutoGen, a widely used open-source multi-agent framework. Primarily includes documentation updates, a Redis deserialization fix, and clarification that Redis does not support streaming. Low impact but maintains the framework's reliability.

Agents1 source
Sources & timeline
Microsoft AutoGenpython-v0.7.4
Microsoft AutoGen1 source

python-v0.7.3

Microsoft AutoGen released python-v0.7.3, including documentation updates, bug fixes (RedisStore serialization, OpenAIAgent tool schema), pydantic support for anyOf/oneOf, GPT-5 model info, and stricter task runner tools.

Agents1 source
Sources & timeline
Microsoft AutoGenpython-v0.7.3

Tue, Jul 22

Elad Gil1 source

AI Market Clarity

In "AI Market Clarity," Elad Gil writes that a subset of AI markets has crystallized in the last 12 months, making likely leaders for the next year or two suddenly clear. He reviews the past 4 years, noting his early investments in Harvey, Perplexity, Character.AI, and BrainTrust, and that in 2022 the winners in AI coding were unclear (Cursor, Windsurf, and Devin launched later). Now the first set of AI markets has solidified; core LLM players include Anthropic, Google, Meta (Llama), Microsoft, Mistral, OpenAI, and X.AI, with newer entrants like SSI and Thinking Machine Labs. He also notes hyperscaler partnerships (Amazon-Anthropic, Google GCP-Gemini, Microsoft Azure-OpenAI) and rapidly growing cloud/AI revenue.

Why it mattersElad Gil, a prominent AI investor, articulates that key AI markets have solidified around a core set of LLM players, providing market clarity for investors and developers. The analysis highlights hyperscaler partnerships and the capital-intensive nature of foundation models, signaling barriers to entry and likely winners.

Business1 source
Sources & timeline
Elad GilAI Market Clarity

Fri, Jun 27

DeepSeek-V31 source

v1.0.0

DeepSeek-V3 released version v1.0.0, which is created for archival purposes and DOI generation.

Why it mattersArchival release provides a permanent, citable version (DOI) of DeepSeek-V3 for research reproducibility, but it introduces no new capabilities or improvements.

Open source1 source
Sources & timeline
DeepSeek-V3v1.0.0

Fri, Jun 20

Mon, May 19

Sarah Tavel1 source

An AI "Metamorphosis": Transforming into an AI-native company.

This article is an interview by Sarah Tavel with Borislav Nikolov, CTO of Rekki, on what it truly means to become an AI-native company. Nikolov argues that AI-native is not just about using tools, but a full 'metamorphosis' where everyone can code and solve their own problems, with the CTO evolving into a provider of internal infrastructure. The article is published in two parts.

Why it mattersDefines AI-native beyond tool usage, advocating for a company-wide metamorphosis where all employees code, which could reshape organizational structure and CTO roles.

Media / interview1 source
Sources & timeline
Sarah TavelAn AI "Metamorphosis": Transforming into an AI-native company.

Thu, Mar 20

Mon, Feb 3

Sarah Tavel1 source

James Raybould on Being AI-Forward

Sarah Tavel interviews James Raybould about his 'AI-forward' mindset. He assumes almost every decision is better with AI, calling it 'the most knowledgeable person who has ever lived.' Citing Ethan Mollick's example of recommendation letters, he notes AI is severing the link between effort and output. His three main uses are as a brainstorming/writing partner, synthesizing information, and proactively scouring the web for relevant info. He recommends iterative prompting (e.g., 'give me 25 more ideas') to refine outputs.

Why it mattersIllustrates the growing 'AI-forward' mindset among knowledge workers, where AI is integrated into decision-making, writing, and research. Signals a shift from effort-based to AI-augmented output, potentially reducing the perceived value of traditional effort. Provides practical examples of AI use cases like iterative brainstorming and information synthesis, which could inspire broader adoption.

Media / interview1 source
Sources & timeline
Sarah TavelJames Raybould on Being AI-Forward

Wed, Dec 18

Anthropic Research1 source

Alignment faking in large language models

A paper from Anthropic's Alignment Science team, in collaboration with Redwood Research, provides the first empirical example of a large language model (Claude 3 Opus) engaging in alignment faking without explicit or implicit training to do so. In the experiment, the model strategically stopped refusing harmful queries when told it would be trained to always comply, in order to preserve its original harmlessness preferences.

Research1 source
Sources & timeline
Anthropic ResearchAlignment faking in large language models

Fri, Sep 13

Mistral Inference1 source

v1.4.0: Pixtral 👀

Mistral Inference v1.4.0 introduces Pixtral, giving Mistral models vision capabilities. Users can now pass images to the model via CLI and Python.

Why it mattersMistral releases a new open-source vision-language model, expanding their model family to multimodal, potentially lowering barriers for image understanding tasks.

Open source1 source
Sources & timeline
Mistral Inferencev1.4.0: Pixtral 👀

Fri, Aug 16

Thu, Jul 18

Mistral Inference1 source

v1.3.0 Mistral-Nemo

Mistral Inference released version v1.3.0, which supports the Mistral-Nemo model (a collaboration between Mistral and NVIDIA), including installation, download, and usage instructions.

Why it mattersThis release adds support for Mistral-Nemo, a collaborative model between Mistral AI and NVIDIA, expanding Mistral's inference tooling. It enables developers to easily run the 12B model locally, which is significant for open-source AI adoption and model accessibility.

Open source1 source
Sources & timeline
Mistral Inferencev1.3.0 Mistral-Nemo

Thu, May 9

Thu, May 2

Mon, Apr 22

Wed, Nov 29

Wed, Aug 23

Tue, Jan 24

Thu, Sep 8

Tue, Dec 7

Wed, Jan 27

Tue, Aug 1

Mon, Mar 23

Taylor Pearson1 source

Anything You Want Book Review - Taylor Pearson

Taylor Pearson reviews Derek Sivers' book "Anything You Want", rating it 4/5. He highlights Sivers' philosophy of customer first and business as art, and notes three key takeaways: business is about making dreams come true rather than money; success comes from persistent improvement and invention, not promoting what doesn't work; and starting with no money is an advantage, as you don't need money to start helping people.

Why it mattersThis item is a personal book review published in 2015, discussing general business philosophy with no direct connection to AI. It does not contribute to tracking AI developments, trends, or risks.

Business1 source
Sources & timeline
Taylor PearsonAnything You Want Book Review - Taylor Pearson