{ "schema_version": 1, "generated_at": "2026-05-22T04:26:13.800Z", "reference_app_url": "https://ai-radar-web-luosongred-5507s-projects.vercel.app", "public_site": { "cloudflare_url": "https://ai-industry-radar.pages.dev", "reference_app_url": "https://ai-radar-web-luosongred-5507s-projects.vercel.app", "purpose": "Primary Cloudflare public read surface for AI Industry Radar data.", "read_only": true }, "source": { "data_source": "supabase_radar_items", "kind": "supabase_public_views", "local_data_used": false, "warnings": [] }, "freshness": { "latest_ingestion": "2026-05-22T04:11:48.668+00:00", "latest_timestamp": "2026-05-22T04:13:38.288+00:00", "latest_timestamp_source": "processed_at", "latest_understanding": "2026-05-22T04:14:10.821+00:00", "note": "Latest public radar timestamp is 2026-05-22T04:13:38.288+00:00 (processed_at)." }, "counts": { "citations": 254, "entities": 506, "excluded": 0, "failed": 0, "included": 101, "ingestion_runs": 26, "item_entities": 625, "needs_review": 5, "public_radar_items": 106, "radar_items": 121, "raw_items": 140, "report_candidates": 14, "report_snapshots": 14, "saved_report_candidates": 14, "scores": 1519, "sources": 312, "snapshot_radar_items": 106, "understanding_runs": 26, "visible_radar_items": 106 }, "top_categories": [ { "count": 50, "label": "research" }, { "count": 22, "label": "product update" }, { "count": 17, "label": "other" }, { "count": 12, "label": "open source" }, { "count": 9, "label": "agent" }, { "count": 9, "label": "model release" }, { "count": 8, "label": "opinion" }, { "count": 5, "label": "media interview" }, { "count": 4, "label": "benchmark" }, { "count": 3, "label": "infrastructure" } ], "top_source_tiers": [ { "count": 80, "label": "T1" }, { "count": 18, "label": "T2" }, { "count": 7, "label": "T1.5" }, { "count": 1, "label": "unreviewed" } ], "top_sources": [ { "count": 12, "label": "arXiv cs.CL" }, { "count": 12, "label": "arXiv cs.CV" }, { "count": 10, "label": "arXiv cs.LG" }, { "count": 9, "label": "arXiv cs.AI" }, { "count": 9, "label": "OpenAI News" }, { "count": 4, "label": "Anthropic Python SDK" }, { "count": 4, "label": "Lex Fridman" }, { "count": 3, "label": "Hugging Face Transformers" }, { "count": 2, "label": "DeepSeek-V3" }, { "count": 2, "label": "OpenAI Python SDK" } ], "radar_items": [ { "id": "radar_8b99b678f3681d37", "categories": [ "other" ], "collected_at": "2026-05-22T04:11:46.587+00:00", "confidence": 0.5167, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:13:38.288+00:00", "scores": { "ai_relevance": 1, "credibility": 0.25, "freshness": 1, "importance": 0.715, "novelty": 0.5075, "overall": 0.6891 }, "source_name": "Dan Hock's Essays", "source_tier": "unreviewed", "status": "needs_review", "summary_en": "The strategy and tactics of growing startups and growing your career. Less frequent, more rigorous essays. A Substack publication by Dan Hockenmaier with tens of thousands of subscribers.", "summary_zh": "关于初创公司成长和职业发展的策略与战术。文章较少但更加严谨。这是Dan Hockenmaier在Substack上的出版物，拥有数万订阅者。", "tags": [ "startup", "career", "essays", "substack" ], "title": "Dan Hock's Essays | Substack", "url": "https://www.danhock.co/", "why_it_matters": "Potentially relevant AI signal for review: Dan Hock's Essays | Substack" }, { "id": "radar_97fa436d3b72fdf2", "categories": [ "opinion" ], "collected_at": "2026-05-22T04:11:46.553+00:00", "confidence": 0.6366, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:12:21.288+00:00", "scores": { "ai_relevance": 0.6, "credibility": 0.57, "freshness": 1, "importance": 0.543, "novelty": 0.473, "overall": 0.6026 }, "source_name": "Tomasz Tunguz", "source_tier": "T2", "status": "included", "summary_en": "Tomasz Tunguz is a partner at Theory Ventures and a former director at Redpoint Ventures. He writes frequently about SaaS, with short, informal posts of variable quality.", "summary_zh": "Tomasz Tunguz 是 Theory Ventures 的合伙人，前红点创投董事，主要撰写 SaaS 相关文章，更新频繁，文章简短随意，质量参差不齐。", "tags": [ "investor", "vc-partner", "venture", "blog", "saas", "ai" ], "title": "Tomasz Tunguz", "url": "https://www.tomtunguz.com/", "why_it_matters": "Potentially relevant AI signal for review: Tomasz Tunguz" }, { "id": "radar_13e1e2f091841949", "categories": [ "opinion" ], "collected_at": "2026-05-22T04:08:36.208+00:00", "confidence": 0.7034, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:11:05.213+00:00", "scores": { "ai_relevance": 0.6, "credibility": 0.57, "freshness": 1, "importance": 0.543, "novelty": 0.473, "overall": 0.6026 }, "source_name": "Stratechery", "source_tier": "T2", "status": "included", "summary_en": "Stratechery is a tech media site founded by Ben Thompson, focusing on analyzing the strategies of US tech giants. Ben Thompson is recognized as one of the most well-known tech commentators overseas. This entry is a metadata summary of the Articles category page, not specific article content.", "summary_zh": "Stratechery 是一个由 Ben Thompson 创办的科技媒体网站，专注于分析美国科技巨头的战略走向。Ben Thompson 被誉为海外最知名的科技评论博主之一。该条目是 Articles 分类页面的元数据摘要，并非具体文章内容。", "tags": [ "stratechery", "ben thompson", "tech analysis", "tech strategy" ], "title": "Articles – Stratechery by Ben Thompson", "url": "https://stratechery.com/category/articles", "why_it_matters": "Potentially relevant AI signal for review: Articles – Stratechery by Ben Thompson" }, { "id": "radar_3dc3be5bc52fbf03", "categories": [ "other" ], "collected_at": "2026-05-22T04:08:36.181+00:00", "confidence": 0.57, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:09:50.114+00:00", "scores": { "ai_relevance": 0.4, "credibility": 0.57, "freshness": 1, "importance": 0.412, "novelty": 0.3875, "overall": 0.5035 }, "source_name": "Sequoia", "source_tier": "T2", "status": "needs_review", "summary_en": "Sequoia's Stories page features long-form founder profiles, market and technology perspectives, and news about portfolio companies. It includes articles such as 'AI Ascent 2026', 'From Hierarchy to Intelligence', 'Services: The New Software', and more.", "summary_zh": "红杉资本的故事页面包含创始人长篇介绍、市场和技术观点以及投资组合公司新闻。其中包括《AI Ascent 2026》、《From Hierarchy to Intelligence》、《Services: The New Software》等文章。", "tags": [ "sequoia", "vc-blog", "startup", "founder-stories" ], "title": "Stories", "url": "https://sequoiacap.com/stories", "why_it_matters": "Potentially relevant AI signal for review: Stories" }, { "id": "radar_96ec9f35fc86b349", "categories": [ "other" ], "collected_at": "2026-05-22T04:08:36.164+00:00", "confidence": 0.4366, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:09:15.362+00:00", "scores": { "ai_relevance": 1, "credibility": 0.57, "freshness": 1, "importance": 0.742, "novelty": 0.5075, "overall": 0.7675 }, "source_name": "SemiAnalysis", "source_tier": "T2", "status": "included", "summary_en": "SemiAnalysis is a tech media outlet bridging semiconductors and business, offering in-depth research and models on accelerators, HBM, AI cloud TCO, networking, datacenter, energy, and more.", "summary_zh": "SemiAnalysis 是一个专注于半导体与AI基础设施的技术媒体，提供加速器、HBM、AI云TCO、网络、数据中心、能源等领域的深度研究和模型。", "tags": [ "semianalysis", "semiconductors", "ai-infrastructure", "newsletter", "tech-media" ], "title": "SemiAnalysis", "url": "https://semianalysis.com/", "why_it_matters": "Potentially relevant AI signal for review: SemiAnalysis" }, { "id": "radar_c371086c111af941", "categories": [ "other" ], "collected_at": "2026-05-22T04:08:36.03+00:00", "confidence": 0.8034, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:08:38.282+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.57, "freshness": 1, "importance": 0.687, "novelty": 0.4875, "overall": 0.7235 }, "source_name": "Sarah Tavel", "source_tier": "T2", "status": "included", "summary_en": "Full archive page of Sarah Tavel's newsletter, listing all historical posts, including several articles on AI and venture capital.", "summary_zh": "Sarah Tavel通讯的完整档案页面，列出了所有历史文章，包括多篇关于人工智能和风险投资的文章。", "tags": [ "investor", "vc-partner", "venture", "newsletter", "archive" ], "title": "Archive - Sarah Tavel's Newsletter", "url": "https://www.sarahtavel.com/archive", "why_it_matters": "Potentially relevant AI signal for review: Archive - Sarah Tavel's Newsletter" }, { "id": "radar_89dcf8d523692dcf", "categories": [ "opinion", "other" ], "collected_at": "2026-05-22T04:05:04.001+00:00", "confidence": 0.7034, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T04:05:05.909+00:00", "scores": { "ai_relevance": 0.6, "credibility": 0.57, "freshness": 1, "importance": 0.543, "novelty": 0.473, "overall": 0.6026 }, "source_name": "Nnamdi", "source_tier": "T2", "status": "included", "summary_en": "A listing of essays by Nnamdi Iregbulem, including titles such as 'Tokens Aren't Fungible', 'Seed Valuations Aren’t Valuations', 'AI Benchmarking Is Broken', and 'The Venture Activity Index – Q4 2023'.", "summary_zh": "Nnamdi Iregbulem的博客文章列表页，包含多篇关于人工智能、风险投资等主题的文章标题与链接，如“Tokens Aren't Fungible”、“Seed Valuations Aren’t Valuations”等。", "tags": [ "investor", "vc-partner", "venture", "ai", "essays" ], "title": "Essays — Nnamdi Iregbulem", "url": "https://whoisnnamdi.com/essays", "why_it_matters": "Potentially relevant AI signal for review: Essays — Nnamdi Iregbulem" }, { "id": "radar_194d4ca6f9d502f8", "categories": [ "research" ], "collected_at": "2026-05-22T03:16:52.567+00:00", "confidence": 0.8034, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:18:44.138+00:00", "scores": { "ai_relevance": 1, "credibility": 0.57, "freshness": 1, "importance": 0.859, "novelty": 0.761, "overall": 0.8289 }, "source_name": "机器学习研究杂志（JMLR）", "source_tier": "T2", "status": "included", "summary_en": "JMLR (Journal of Machine Learning Research) is a machine learning research journal founded in 2000, with all published papers freely available online.", "summary_zh": "JMLR（Journal of Machine Learning Research）是一个成立于2000年的机器学习研究期刊，所有已发表的论文均可在线免费获取。", "tags": [ "machine learning", "journal", "open access", "publication" ], "title": "Journal of Machine Learning Research", "url": "http://www.jmlr.org/", "why_it_matters": "May add technical evidence for future radar tracking: Journal of Machine Learning Research" }, { "id": "radar_7fe3bd6d4f17fc49", "categories": [ "other" ], "collected_at": "2026-05-22T03:16:52.55+00:00", "confidence": 0.77, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:18:13.753+00:00", "scores": { "ai_relevance": 1, "credibility": 0.57, "freshness": 1, "importance": 0.742, "novelty": 0.5075, "overall": 0.7675 }, "source_name": "Elad Gil", "source_tier": "T2", "status": "included", "summary_en": "Archive page of Elad Gil's blog, listing links to all posts, including recent articles on AI, biotech, markets, etc.", "summary_zh": "Elad Gil博客的存档页面，列出了所有文章的链接，包括关于AI、生物科技、市场等主题的最新文章。", "tags": [ "blog", "archive", "investor", "venture" ], "title": "Archive - Elad Blog", "url": "https://blog.eladgil.com/archive", "why_it_matters": "Potentially relevant AI signal for review: Archive - Elad Blog" }, { "id": "radar_bad9f10761aefe5b", "categories": [ "opinion", "other" ], "collected_at": "2026-05-22T03:16:52.522+00:00", "confidence": 0.77, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:17:41.801+00:00", "scores": { "ai_relevance": 0.5, "credibility": 0.57, "freshness": 1, "importance": 0.488, "novelty": 0.453, "overall": 0.5585 }, "source_name": "Digital Native", "source_tier": "T2", "status": "needs_review", "summary_en": "This is the homepage summary of Digital Native, a Substack publication by Rex Woodbury, featuring weekly writing on the intersection of technology and people. Source is an investor blog.", "summary_zh": "这是Digital Native（由Rex Woodbury撰写）的Substack主页摘要，每周发布关于科技与人类交汇的写作。来源为投资人博客。", "tags": [ "investor", "product", "vc-partner", "venture", "newsletter", "technology" ], "title": "Digital Native | Rex Woodbury | Substack", "url": "https://www.digitalnative.tech/", "why_it_matters": "Potentially relevant AI signal for review: Digital Native | Rex Woodbury | Substack" }, { "id": "radar_7dc439ea357f4365", "categories": [ "research" ], "collected_at": "2026-05-22T03:16:52.49+00:00", "confidence": 0.7034, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:16:56.755+00:00", "scores": { "ai_relevance": 0.6, "credibility": 0.57, "freshness": 1, "importance": 0.639, "novelty": 0.681, "overall": 0.6529 }, "source_name": "Coatue", "source_tier": "T2", "status": "included", "summary_en": "The Coatue Insights page serves as a central hub for Coatue, a lifecycle investment platform, featuring their latest perspectives, portfolio updates, and industry analysis. Recent content includes a public markets update deck from May 6, 2026, a partnership announcement with Anthropic, and daily charts.", "summary_zh": "Coatue Insights页面是Coatue（一家生命周期投资平台）的见解中心，汇集了其公开市场更新、投资组合动态、行业分析等内容。页面包含近期文章，如2026年5月6日的公开市场更新、与Anthropic的合作公告，以及每日图表等。", "tags": [ "coatue", "insights", "investment", "technology" ], "title": "Our Insights | Coatue", "url": "https://www.coatue.com/insights", "why_it_matters": "May add technical evidence for future radar tracking: Our Insights | Coatue" }, { "id": "radar_8f9b517bb2f4740f", "categories": [ "research" ], "collected_at": "2026-05-22T03:13:49.456+00:00", "confidence": 0.7867, "evidence_notes": [ "raw_item.text字段明确描述：'谷歌研究员christopher olah的博客，原创质量很高，经常被国内ai技术媒体搬运。'", "元数据中last-modified为2024-01-09，表明网站活跃更新。", "links列表包含典型博客导航链接（about, contact, 文章等），确认是个人技术博客。", "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "mixed", "processed_at": "2026-05-22T03:16:08.546+00:00", "scores": { "ai_relevance": 1, "credibility": 0.57, "freshness": 1, "importance": 0.859, "novelty": 0.761, "overall": 0.8289 }, "source_name": "Christopher Olah", "source_tier": "T2", "status": "included", "summary_en": "Homepage of Christopher Olah's blog, featuring high-quality original technical articles often reposted by Chinese AI media.", "summary_zh": "克里斯托弗·奥拉（Christopher Olah）的个人博客首页，内容为高质量的原创技术文章，经常被国内AI技术媒体转载。", "tags": [ "blog", "research", "colah", "google" ], "title": "Home - colah's blog", "url": "http://colah.github.io/", "why_it_matters": "May add technical evidence for future radar tracking: Home - colah's blog" }, { "id": "radar_b71778e61e6d767e", "categories": [ "other" ], "collected_at": "2026-05-22T03:13:49.394+00:00", "confidence": 0.17, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:15:00.365+00:00", "scores": { "ai_relevance": 1, "credibility": 0.57, "freshness": 1, "importance": 0.742, "novelty": 0.5075, "overall": 0.7675 }, "source_name": "Apoorv’s notes", "source_tier": "T2", "status": "included", "summary_en": "This is the homepage of \"Tailwinds,\" a Substack publication by venture investor Apoorv Agrawal, focusing on the business of technology and the tailwinds that power it.", "summary_zh": "该条目为风险投资人Apoorv Agrawal的Substack博客\"Tailwinds\"的主页，内容聚焦科技商业及推动其发展的顺风趋势。", "tags": [ "investor", "vc-blog", "venture", "business", "technology" ], "title": "Tailwinds | Apoorv Agrawal | Substack", "url": "https://apoorv03.com/", "why_it_matters": "Potentially relevant AI signal for review: Tailwinds | Apoorv Agrawal | Substack" }, { "id": "radar_886fae3dd2af9c8a", "categories": [ "other" ], "collected_at": "2026-05-22T03:13:49.365+00:00", "confidence": 0.7367, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:14:25.776+00:00", "scores": { "ai_relevance": 1, "credibility": 0.57, "freshness": 1, "importance": 0.742, "novelty": 0.5075, "overall": 0.7675 }, "source_name": "Andrej Karpathy", "source_tier": "T2", "status": "included", "summary_en": "Andrej Karpathy's personal tech blog, former OpenAI and Tesla expert, featuring tutorials on large language models that are accessible to non-technical audiences.", "summary_zh": "Andrej Karpathy的个人技术博客，前OpenAI和Tesla专家，网站包含大量关于大语言模型的教程，对非技术人群友好。", "tags": [ "personal-blog", "technical-blog", "research", "education" ], "title": "Andrej Karpathy", "url": "https://karpathy.ai/", "why_it_matters": "Potentially relevant AI signal for review: Andrej Karpathy" }, { "id": "radar_6ca572668c630cd9", "categories": [ "model_release", "product_update", "infrastructure" ], "collected_at": "2026-05-22T03:13:49.337+00:00", "confidence": 0.8468, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:13:50.888+00:00", "scores": { "ai_relevance": 1, "credibility": 0.831, "freshness": 1, "importance": 0.919, "novelty": 0.813, "overall": 0.913 }, "source_name": "Kimi Platform Docs", "source_tier": "T1", "status": "included", "summary_en": "Kimi API Platform launches the K2.6 Open Platform, providing a trillion-parameter K2.5 large language model API, supporting 256K long context and Tool Calling, with professional code generation, intelligent dialogue, and visual reasoning capabilities to help developers build AI applications.", "summary_zh": "Kimi API平台推出K2.6开放平台，提供万亿参数K2.5大语言模型API，支持256K长上下文和工具调用，具备专业代码生成、智能对话、视觉推理能力，帮助开发者构建AI应用。", "tags": [ "kimi", "api", "llm", "moonshot", "trillion-parameter" ], "title": "Welcome to Kimi API Docs - Kimi API Platform", "url": "https://platform.kimi.ai/docs/overview", "why_it_matters": "May affect model capability tracking and product benchmarking: Welcome to Kimi API Docs - Kimi API Platform" }, { "id": "radar_08925fdae7e9e342", "categories": [ "other" ], "collected_at": "2026-05-22T03:11:47.239+00:00", "confidence": 0.57, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "mixed", "processed_at": "2026-05-22T03:13:10.187+00:00", "scores": { "ai_relevance": 0.7, "credibility": 0.57, "freshness": 1, "importance": 0.577, "novelty": 0.4475, "overall": 0.6355 }, "source_name": "A16Z", "source_tier": "T2", "status": "included", "summary_en": "This entry is the news and content page of Andreessen Horowitz (A16Z), aggregating links to its blog, investment areas (AI, Bio+Health, Crypto, etc.), and team. The page itself is a navigation page without specific articles. Metadata suggests comprehensive and rapid coverage of AI trends.", "summary_zh": "该条目为安德森·霍洛维茨基金（A16Z）的新闻与内容页面，汇总了其博客、投资领域（如AI、生物健康、加密等）及团队信息。页面本身为导航页，无具体文章内容。元数据中提及“AI趋势覆盖得多快好全”，暗示其AI领域内容更新迅速。", "tags": [ "a16z", "news", "content", "investor", "vc-blog" ], "title": "News and Content | Andreessen Horowitz", "url": "https://a16z.com/news-content", "why_it_matters": "Potentially relevant AI signal for review: News and Content | Andreessen Horowitz" }, { "id": "radar_7ec5cf5a916ad03e", "categories": [ "product_update" ], "collected_at": "2026-05-22T03:11:47.216+00:00", "confidence": 0.8488, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:12:37.825+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.843, "freshness": 1, "importance": 0.828, "novelty": 0.702, "overall": 0.8525 }, "source_name": "DeepSeek API Docs", "source_tier": "T1", "status": "included", "summary_en": "DeepSeek API is compatible with OpenAI/Anthropic API formats. By configuring the base URL, users can utilize OpenAI/Anthropic SDKs or compatible software to access DeepSeek API.", "summary_zh": "DeepSeek API兼容OpenAI/Anthropic的API格式。通过修改配置，用户可以使用OpenAI/Anthropic的SDK或兼容软件访问DeepSeek API。", "tags": [ "deepseek", "api", "documentation", "developer" ], "title": "Your First API Call | DeepSeek API Docs", "url": "https://api-docs.deepseek.com/", "why_it_matters": "Potentially relevant AI signal for review: Your First API Call | DeepSeek API Docs" }, { "id": "radar_4d065b1e19a95997", "categories": [ "product_update", "model_release" ], "collected_at": "2026-05-22T03:11:47.147+00:00", "confidence": 0.8488, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:12:00.885+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.843, "freshness": 1, "importance": 0.8975, "novelty": 0.803, "overall": 0.8965 }, "source_name": "Cohere Changelog", "source_tier": "T1", "status": "included", "summary_en": "The Cohere official changelog page for model, API, and developer platform updates. However, only page metadata was captured during this ingestion; no specific release notes were extracted.", "summary_zh": "Cohere官方更新日志页面，收录模型、API和开发者平台的更新，但本次抓取仅获取了页面元数据，未提取具体更新内容。", "tags": [ "cohere", "developer", "models", "official", "release-notes", "changelog", "api" ], "title": "Release Notes | Cohere", "url": "https://docs.cohere.com/changelog", "why_it_matters": "May affect model capability tracking and product benchmarking: Release Notes | Cohere" }, { "id": "radar_52aed8fac227edc4", "categories": [ "research" ], "collected_at": "2026-05-22T03:10:25.741+00:00", "confidence": 0.7964, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:11:08.7+00:00", "scores": { "ai_relevance": 1, "credibility": 0.728, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.8765 }, "source_name": "Yarin Gal", "source_tier": "T1.5", "status": "included", "summary_en": "Home page of Yarin Gal, a researcher at Oxford Machine Learning. The page serves as a portal with links to his research, publications, talks, software, blog, and other resources.", "summary_zh": "亚林·加尔（Yarin Gal）的个人主页，牛津大学机器学习研究员。页面提供其研究、出版物、讲座、软件、博客等链接入口。", "tags": [ "researcher", "homepage", "personal", "academic", "oxford" ], "title": "Yarin Gal - Home Page | Oxford Machine Learning", "url": "http://yarin.co/", "why_it_matters": "May add technical evidence for future radar tracking: Yarin Gal - Home Page | Oxford Machine Learning" }, { "id": "radar_5e6cf9ddee209b91", "categories": [ "product_update", "model_release" ], "collected_at": "2026-05-22T03:10:25.69+00:00", "confidence": 0.8488, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:10:29.623+00:00", "scores": { "ai_relevance": 1, "credibility": 0.843, "freshness": 1, "importance": 0.925, "novelty": 0.813, "overall": 0.9185 }, "source_name": "Alibaba Cloud Model Studio Release Notes", "source_tier": "T1", "status": "included", "summary_en": "Alibaba Cloud Model Studio release notes cover Qwen model updates, OpenAI-compatible endpoint changes, and LLM capability deprecation timelines. Consult them to avoid deprecated API call failures.", "summary_zh": "阿里云模型工作室的发布说明页面涵盖Qwen模型更新、OpenAI兼容端点变更及LLM功能弃用时间线，建议查阅以避免API调用失败。", "tags": [ "alibaba", "qwen", "model-studio", "api", "release-notes", "deprecation" ], "title": "Changelog - Alibaba Cloud", "url": "https://www.alibabacloud.com/help/en/model-studio/release-notes", "why_it_matters": "May affect model capability tracking and product benchmarking: Changelog - Alibaba Cloud" }, { "id": "radar_d29f26e5dce6f110", "categories": [ "other" ], "collected_at": "2026-05-22T03:08:07.611+00:00", "confidence": 0.8498, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:09:46.31+00:00", "scores": { "ai_relevance": 0.85, "credibility": 0.849, "freshness": 1, "importance": 0.7045, "novelty": 0.4775, "overall": 0.7813 }, "source_name": "NVIDIA Generative AI Technical Blog", "source_tier": "T1", "status": "included", "summary_en": "Category page for Generative AI and Agentic AI on the NVIDIA Developer Technical Blog, listing posts on generative AI, agents, acceleration, and deployment.", "summary_zh": "NVIDIA开发者技术博客中关于生成式AI和智能体AI的分类页面，列出了多篇技术文章，涵盖生成式AI、智能体、加速和部署。", "tags": [ "compute", "developer", "infrastructure", "nvidia", "official", "agentic_ai", "generative_ai" ], "title": "Category: Agentic AI / Generative AI | NVIDIA Technical Blog", "url": "https://developer.nvidia.com/blog/category/generative-ai", "why_it_matters": "Potentially relevant AI signal for review: Category: Agentic AI / Generative AI | NVIDIA Technical Blog" }, { "id": "radar_c4d96b5da0a9d2d5", "categories": [ "opinion" ], "collected_at": "2026-05-22T03:08:07.533+00:00", "confidence": 0.1963, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "mixed", "processed_at": "2026-05-22T03:09:03.609+00:00", "scores": { "ai_relevance": 0.85, "credibility": 0.728, "freshness": 1, "importance": 0.7105, "novelty": 0.523, "overall": 0.7601 }, "source_name": "Thesephist", "source_tier": "T1.5", "status": "included", "summary_en": "Thesephist.com is the personal site of a Notion AI product lead, featuring unique thoughts on product and interaction design. Source is an English blog/newsletter, weight 0.78.", "summary_zh": "Thesephist.com 是 Notion AI 产品负责人的个人网站，包含关于产品、交互设计的独特思考。源为英文博客/通讯，权重 0.78。", "tags": [ "ai-specific", "newsletter", "product", "blog", "opinion", "notion ai" ], "title": "thesephist.com", "url": "https://thesephist.com/", "why_it_matters": "Potentially relevant AI signal for review: thesephist.com" }, { "id": "radar_8d711b9b6ab07621", "categories": [ "other" ], "collected_at": "2026-05-22T03:08:07.325+00:00", "confidence": 0.6499, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:08:19.822+00:00", "scores": { "ai_relevance": 0.85, "credibility": 0.849, "freshness": 1, "importance": 0.7045, "novelty": 0.4775, "overall": 0.7813 }, "source_name": "NVIDIA AI Blog", "source_tier": "T1", "status": "included", "summary_en": "NVIDIA AI Blog's deep learning category page, listing recent AI and deep learning posts, including topics on Isambard-AI supercomputer, Vera CPU, Hermes AI agents, etc.", "summary_zh": "NVIDIA AI博客的深度学习分类页面，列出近期AI和深度学习相关博文，包括关于Isambard-AI超级计算机、Vera CPU、Hermes AI代理等。", "tags": [ "deep learning", "nvidia", "ai", "announcements" ], "title": "Deep Learning Archives", "url": "https://blogs.nvidia.com/blog/category/enterprise/deep-learning", "why_it_matters": "Potentially relevant AI signal for review: Deep Learning Archives" }, { "id": "radar_d1cc436771ce8a29", "categories": [ "opinion" ], "collected_at": "2026-05-22T03:06:56.039+00:00", "confidence": 0.8296, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:06:57.803+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.728, "freshness": 1, "importance": 0.738, "novelty": 0.533, "overall": 0.7822 }, "source_name": "Stephen Wolfram", "source_tier": "T1.5", "status": "included", "summary_en": "Collection of articles by Stephen Wolfram covering artificial intelligence, computational science, data science, education, future and historical perspectives, sciences, software design, technology, Wolfram products, and more. Source is his personal website, categorized as an AI-specific technical newsletter.", "summary_zh": "Stephen Wolfram 的文章合集，涵盖人工智能、计算科学、数据科学、教育、未来与历史视角、科学、软件设计、技术、Wolfram 产品等主题。来源为他的个人网站，属于 AI 相关的技术新闻通讯。", "tags": [ "artificial intelligence", "computational science", "data science", "education", "technology" ], "title": "Stephen Wolfram Writings", "url": "https://writings.stephenwolfram.com/", "why_it_matters": "Potentially relevant AI signal for review: Stephen Wolfram Writings" }, { "id": "radar_ac9e1bffbe78da71", "categories": [ "open_source" ], "collected_at": "2026-05-22T03:05:32.52+00:00", "confidence": 0.6175, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:06:17.808+00:00", "scores": { "ai_relevance": 0.8, "credibility": 0.855, "freshness": 1, "importance": 0.791, "novelty": 0.708, "overall": 0.8204 }, "source_name": "Hugging Face Blog", "source_tier": "T1", "status": "included", "summary_en": "The official blog page of Hugging Face, committed to advancing and democratizing AI through open source and open science. The page includes links to Models, Datasets, Spaces, and other products.", "summary_zh": "Hugging Face官方博客页面，致力于通过开源和开放科学推进和民主化人工智能。页面包含模型、数据集、Spaces等产品链接。", "tags": [ "huggingface", "models", "official", "open-source" ], "title": "Hugging Face – Blog", "url": "https://huggingface.co/blog", "why_it_matters": "May change available building blocks for teams evaluating open implementations: Hugging Face – Blog" }, { "id": "radar_87a13948102a46a4", "categories": [ "research" ], "collected_at": "2026-05-22T03:05:32.514+00:00", "confidence": 0.8296, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:05:43.227+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.728, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8326 }, "source_name": "Lilian Weng", "source_tier": "T1.5", "status": "included", "summary_en": "Homepage of Lilian Weng's personal blog 'Lil'Log', described as 'Document my learning notes.' It is a technical blog in AI domain, source weight 0.78, language English.", "summary_zh": "Lilian Weng的个人博客首页，标题为Lil'Log，描述为“记录我的学习笔记”。该站点为AI领域的技术博客，来源权重0.78，语言英文。", "tags": [ "newsletter", "blog", "ai research", "technical blog" ], "title": "Lil'Log", "url": "https://lilianweng.github.io/", "why_it_matters": "May add technical evidence for future radar tracking: Lil'Log" }, { "id": "radar_e922e8601641d54b", "categories": [ "product_update" ], "collected_at": "2026-05-22T03:03:02.54+00:00", "confidence": 0.8508, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "unknown", "processed_at": "2026-05-22T03:04:49.327+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.855, "freshness": 1, "importance": 0.834, "novelty": 0.702, "overall": 0.8581 }, "source_name": "Google Gemini API Changelog", "source_tier": "T1", "status": "included", "summary_en": "The changelog page for Google Gemini API (Korean language), collected on May 22, 2026. The title is '출시 노트 | Gemini API | Google AI for Developers', metadata indicates it is a raw HTML summary containing links to API docs, key acquisition, etc. No specific update content was extracted.", "summary_zh": "Google Gemini API 的发布说明页面（韩文版），收集于2026年5月22日。页面标题为“출시 노트 | Gemini API | Google AI for Developers”，元数据表明这是一个原始HTML摘要，包含指向API文档、密钥获取等链接。未提取具体的更新内容。", "tags": [ "developer", "gemini", "google", "models", "official", "release-notes", "changelog", "api" ], "title": "출시 노트  |  Gemini API  |  Google AI for Developers", "url": "https://ai.google.dev/gemini-api/docs/changelog?hl=ko", "why_it_matters": "Potentially relevant AI signal for review: 출시 노트  |  Gemini API  |  Google AI for Developers" }, { "id": "radar_57c2241ccf3b26cc", "categories": [ "product_update" ], "collected_at": "2026-05-22T03:03:02.499+00:00", "confidence": 0.8508, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:03:37.422+00:00", "scores": { "ai_relevance": 1, "credibility": 0.855, "freshness": 1, "importance": 0.889, "novelty": 0.722, "overall": 0.9021 }, "source_name": "Google AI for Developers", "source_tier": "T1", "status": "included", "summary_en": "This is the official documentation page for the Gemini API's generateContent endpoint on Google AI for Developers, with links to resources such as quickstart, API keys, libraries, pricing, and community.", "summary_zh": "这是Google AI for Developers上Gemini API的generateContent端点的官方文档页面，包含快速入门、API密钥、库、定价和社区等资源的链接。", "tags": [ "gemini", "api", "google", "generative-ai", "developer", "documentation", "reference" ], "title": "Gemini generateContent API  |  Google AI for Developers", "url": "https://ai.google.dev/gemini-api/docs", "why_it_matters": "Potentially relevant AI signal for review: Gemini generateContent API  |  Google AI for Developers" }, { "id": "radar_f13e235515c7869a", "categories": [ "media_interview" ], "collected_at": "2026-05-22T03:03:02.482+00:00", "confidence": 0.8463, "evidence_notes": [ "text from raw item: 'the ai engineer newsletter + top technical ai podcast. how leading labs build agents, models, infra, & ai for science. see https://latent.space/about for highlights from greg brockman, andrej karpathy...'", "metadata indicates source_homepage: https://www.latent.space/", "links include substack policies and other newsletters, but no further details about content.", "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:03:03.68+00:00", "scores": { "ai_relevance": 1, "credibility": 0.728, "freshness": 1, "importance": 0.823, "novelty": 0.618, "overall": 0.8419 }, "source_name": "Latent Space", "source_tier": "T1.5", "status": "included", "summary_en": "Latent Space is an AI Engineer newsletter and top technical AI podcast covering how leading labs build Agents, Models, Infra, & AI for Science. It features highlights from Greg Brockman, Andrej Karpathy, and others. The Substack publication has hundreds of thousands of subscribers.", "summary_zh": "Latent Space 是一份 AI Engineer 新闻通讯和顶级技术 AI 播客，涵盖代理、模型、基础设施和科学 AI 的构建方法。其嘉宾包括 Greg Brockman、Andrej Karpathy 等知名人士。该刊物在 Substack 上拥有数十万订阅者。", "tags": [ "newsletter", "podcast", "ai engineering", "agents", "models", "infrastructure" ], "title": "Latent.Space | Substack", "url": "https://www.latent.space/", "why_it_matters": "Potentially relevant AI signal for review: Latent.Space | Substack" }, { "id": "radar_65ef0c0339cca9d2", "categories": [ "model_release" ], "collected_at": "2026-05-22T02:59:32.516+00:00", "confidence": 0.8611, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:02:16.132+00:00", "published_at": "2025-08-28T03:24:37+00:00", "scores": { "ai_relevance": 1, "credibility": 0.917, "freshness": 0.25, "importance": 0.922, "novelty": 0.813, "overall": 0.8568 }, "source_name": "DeepSeek-V3", "source_tier": "T1", "status": "included", "summary_en": "This item is metadata of the official DeepSeek-V3 model repository on GitHub, including star count (103,578), forks (16,741), and open issues (216).", "summary_zh": "该条目为DeepSeek-V3官方模型仓库的GitHub元数据，包含星标数(103,578)、分支数(16,741)、开放问题数(216)等统计信息。", "tags": [ "deepseek", "github", "models", "open-source", "v3" ], "title": "deepseek-ai/DeepSeek-V3 repository metadata", "url": "https://github.com/deepseek-ai/DeepSeek-V3", "why_it_matters": "May affect model capability tracking and product benchmarking: deepseek-ai/DeepSeek-V3 repository metadata" }, { "id": "radar_55c7a0e63f4c612d", "categories": [ "model_release", "open_source" ], "collected_at": "2026-05-22T02:59:32.464+00:00", "confidence": 0.7955, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T03:00:08.534+00:00", "published_at": "2026-01-09T03:05:47+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.923, "freshness": 0.25, "importance": 0.87, "novelty": 0.793, "overall": 0.8156 }, "source_name": "Qwen3", "source_tier": "T1", "status": "included", "summary_en": "Qwen3 is a large language model series developed by Qwen team, Alibaba Cloud. Its GitHub repository metadata shows 27,246 stars, 1,981 forks, and 42 open issues as of January 9, 2026.", "summary_zh": "Qwen3是阿里云Qwen团队开发的大语言模型系列，其GitHub仓库元数据显示截至2026年1月9日已获得27246颗星、1981个分叉和42个开放问题。", "tags": [ "qwen3", "alibaba", "large language model", "open source", "github" ], "title": "QwenLM/Qwen3 repository metadata", "url": "https://github.com/QwenLM/Qwen3", "why_it_matters": "May affect model capability tracking and product benchmarking: QwenLM/Qwen3 repository metadata" }, { "id": "radar_59875977de6cf2b4", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:59:32.444+00:00", "confidence": 0.2185, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:59:35.499+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.861, "freshness": 1, "importance": 0.837, "novelty": 0.702, "overall": 0.8609 }, "source_name": "Google Gemini Blog", "source_tier": "T1", "status": "included", "summary_en": "The product page on the Google Gemini Blog serves as a central hub for news and updates about Gemini AI, including official information on writing, planning, learning, and other features.", "summary_zh": "Google Gemini 博客的产品页面，作为 Gemini AI 新闻和更新的中心枢纽，涵盖写作、规划、学习等功能的官方信息。", "tags": [ "gemini", "google", "ai", "product" ], "title": "Gemini", "url": "https://blog.google/products-and-platforms/products/gemini", "why_it_matters": "Potentially relevant AI signal for review: Gemini" }, { "id": "radar_560d9df478f45f89", "categories": [ "open_source" ], "collected_at": "2026-05-22T02:56:56.91+00:00", "confidence": 0.8621, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:58:52.871+00:00", "published_at": "2026-05-21T13:25:43+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.923, "freshness": 1, "importance": 0.84, "novelty": 0.728, "overall": 0.8748 }, "source_name": "Meta Llama Stack", "source_tier": "T1", "status": "included", "summary_en": "The ogx-ai/ogx repository is part of Meta Llama Stack on GitHub, with 8383 stars, 1309 forks, and 157 open issues, tagged as open-source and open-models.", "summary_zh": "ogx-ai/ogx 是 Meta Llama Stack 旗下的 GitHub 仓库，拥有 8383 颗星、1309 个复刻和 157 个开放问题，被标记为开源和开放模型。", "tags": [ "github", "meta", "official", "open-models", "open-source" ], "title": "ogx-ai/ogx repository metadata", "url": "https://github.com/ogx-ai/ogx", "why_it_matters": "May change available building blocks for teams evaluating open implementations: ogx-ai/ogx repository metadata" }, { "id": "radar_60606afc80a6fa60", "categories": [ "other" ], "collected_at": "2026-05-22T02:56:56.898+00:00", "confidence": 0.8523, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:58:11.865+00:00", "scores": { "ai_relevance": 0.7, "credibility": 0.864, "freshness": 1, "importance": 0.6295, "novelty": 0.4475, "overall": 0.7223 }, "source_name": "Meta AI Blog", "source_tier": "T1", "status": "included", "summary_en": "Official homepage of Meta AI Blog, featuring latest AI news and updates from Meta.", "summary_zh": "Meta AI 官方博客首页，提供 Meta 人工智能最新新闻与更新。", "tags": [ "meta", "official", "open-models", "research", "blog" ], "title": "AI at Meta Blog", "url": "https://ai.meta.com/blog", "why_it_matters": "Potentially relevant AI signal for review: AI at Meta Blog" }, { "id": "radar_057cc49ab17c43e0", "categories": [ "opinion" ], "collected_at": "2026-05-22T02:56:56.887+00:00", "confidence": 0.763, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:57:38.31+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.728, "freshness": 1, "importance": 0.7655, "novelty": 0.543, "overall": 0.8042 }, "source_name": "Fabricated Knowledge", "source_tier": "T1.5", "status": "included", "summary_en": "Fabricated Knowledge is a Substack publication by Doug O'Laughlin focusing on the world's most important manufactured product, providing insight, analysis, and occasional investment ideas.", "summary_zh": "Fabricated Knowledge 是 Doug O'Laughlin 在 Substack 上的出版物，专注于世界上最重要的制成品，提供见解、分析和偶尔的投资想法。", "tags": [ "semiconductors", "ai-infrastructure", "newsletter", "analysis" ], "title": "Fabricated Knowledge | Doug OLaughlin | Substack", "url": "https://www.fabricatedknowledge.com/", "why_it_matters": "Potentially relevant AI signal for review: Fabricated Knowledge | Doug OLaughlin | Substack" }, { "id": "radar_f84d222af7af5acf", "categories": [ "open_source", "infrastructure" ], "collected_at": "2026-05-22T02:56:56.872+00:00", "confidence": 0.8631, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:57:07.624+00:00", "published_at": "2026-05-21T21:23:37+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.929, "freshness": 1, "importance": 0.843, "novelty": 0.728, "overall": 0.8776 }, "source_name": "OpenAI Python SDK", "source_tier": "T1", "status": "included", "summary_en": "OpenAI Python SDK official repository updated metadata on May 21, 2026, with 30810 stars, 4796 forks, and 537 open issues.", "summary_zh": "OpenAI Python SDK 官方仓库在2026年5月21日更新元数据，拥有30810颗星、4796个分支和537个开放问题。", "tags": [ "developer", "github", "official", "open-source", "sdk" ], "title": "openai/openai-python repository metadata", "url": "https://github.com/openai/openai-python", "why_it_matters": "May change available building blocks for teams evaluating open implementations: openai/openai-python repository metadata" }, { "id": "radar_a783b82f2037fd9b", "categories": [ "research" ], "collected_at": "2026-05-22T02:49:38.39+00:00", "confidence": 0.8748, "evidence_notes": [ "arxiv:2605.20234v1, announce type: new, abstract: 'prior-data fitted networks (pfns) have been very successful... we propose tabpfn-mt, which is trained on an expanded multi-target synthetic prior to capture inter-task dependencies in context.'", "arxiv:2605.20234v1, abstract: 'extensive evaluations across 344 datasets demonstrate that tabpfn-mt establishes a new state-of-the-art for deep tabular multitask learning.'", "arxiv:2605.20234v1, abstract: 'on multitask datasets it achieves an overall accuracy rank of 4.89, the highest average rank among all models tested.'", "arxiv:2605.20234v1, abstract: 'reducing the inference cost for t tasks from o(t) to o(1) forward passes.'", "source url: https://arxiv.org/abs/2605.20234, collected at 2026-05-22t02:49:38.390z", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:56:10.597+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "TabPFN-MT is a natively multitask in-context learner for tabular data. It uses an expanded y-encoder and a shared decoder to enable simultaneous inference of multiple targets, reducing inference cost from O(T) to O(1). Evaluations on 344 datasets show it achieves state-of-the-art deep tabular multitask learning on small datasets (average <1000 samples), with an overall Accuracy rank of 4.89 on multitask datasets, while remaining competitive with top single-task ensembles.", "summary_zh": "TabPFN-MT是一种原生多任务上下文学习器，适用于表格数据。它通过扩展的y编码器和共享解码器头，实现了上下文中的多任务同时推理，将推理成本从O(T)降低到O(1)。在344个数据集上的评估显示，该模型在小样本（平均少于1000个样本）场景下达到了深度表格多任务学习的最优性能，在多任务数据集上取得了4.89的平均准确率排名，同时保持与顶尖单任务集成模型相当的竞争力。", "tags": [ "tabular data", "in-context learning", "multitask learning", "prior-data fitted networks", "tabpfn", "deep learning", "machine learning" ], "title": "TabPFN-MT: A Natively Multitask In-Context Learner for Tabular Data", "url": "https://arxiv.org/abs/2605.20234", "why_it_matters": "May add technical evidence for future radar tracking: TabPFN-MT: A Natively Multitask In-Context Learner for Tabular Data" }, { "id": "radar_79d879c0eafe1136", "categories": [ "research" ], "collected_at": "2026-05-22T02:49:38.39+00:00", "confidence": 0.2248, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:55:28.446+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "GraphDiffMed is a knowledge-constrained medication recommendation framework using dual-scale Differential Attention v2 to filter noise and incorporate pharmacological constraints (e.g., drug-drug interactions), outperforming baselines on MIMIC-III.", "summary_zh": "GraphDiffMed是一种基于知识约束的药物推荐框架，采用双尺度差分注意力v2，结合药理知识（如药物相互作用）来过滤噪声，在MIMIC-III数据集上优于基线模型。", "tags": [ "medication recommendation", "healthcare ai", "electronic health records", "differential attention", "graph priors", "open source" ], "title": "GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation", "url": "https://arxiv.org/abs/2605.20188", "why_it_matters": "May add technical evidence for future radar tracking: GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation" }, { "id": "radar_a6ce57d8dd29c96b", "categories": [ "research" ], "collected_at": "2026-05-22T02:49:38.39+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:54:50.735+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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), using ground-truth MI computed from the model's own conditional distributions for supervision. The estimator predicts the full MI matrix in a single forward pass, enabling MI-guided parallel decoding by identifying conditionally independent variable subsets. Evaluated on Sudoku and protein sequence generation with ESM-C, the method achieves a 3-5x reduction in inference-time forward passes while preserving generative quality and outperforming entropy-based parallelization methods.", "summary_zh": "该论文提出了一种从预训练的掩码扩散模型（MDM）的隐藏状态中估计成对条件互信息的神经框架，通过模型自身的条件分布计算真值互信息进行监督。该估计器能一次性预测完整的互信息矩阵，从而识别条件独立的变量子集，实现互信息引导的并行解码。在数独和ESM-C蛋白质序列生成任务上，该方法将推理时的前向传播次数减少了3-5倍，同时保持了生成质量，并优于基于熵的并行化方法。", "tags": [ "mutual information", "masked diffusion models", "neural estimation", "protein sequence generation", "parallel decoding", "interpretability", "arxiv", "paper" ], "title": "Neural Estimation of Pairwise Mutual Information in Masked Discrete Sequence Models", "url": "https://arxiv.org/abs/2605.20187", "why_it_matters": "May add technical evidence for future radar tracking: Neural Estimation of Pairwise Mutual Information in Masked Discrete Sequence Models" }, { "id": "radar_bbdaaeb7903db9de", "categories": [ "other" ], "collected_at": "2026-05-22T02:49:38.375+00:00", "confidence": 0.8195, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:54:08.634+00:00", "scores": { "ai_relevance": 0.5, "credibility": 0.867, "freshness": 1, "importance": 0.521, "novelty": 0.4075, "overall": 0.6357 }, "source_name": "Google DeepMind Blog", "source_tier": "T1", "status": "needs_review", "summary_en": "Google DeepMind's official blog homepage, featuring links to research, product pages, and AI tools such as Gemini, Google Labs, and Antigravity.", "summary_zh": "Google DeepMind官方博客首页，包含研究、产品页面及AI工具（如Gemini、Google Labs、Antigravity）的链接。", "tags": [ "google deepmind", "blog", "homepage", "research" ], "title": "News — Google DeepMind", "url": "https://deepmind.google/blog", "why_it_matters": "Potentially relevant AI signal for review: News — Google DeepMind" }, { "id": "radar_43125b0e2fc1ee24", "categories": [ "other" ], "collected_at": "2026-05-22T02:49:38.361+00:00", "confidence": 0.7297, "evidence_notes": [ "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:53:38.8+00:00", "scores": { "ai_relevance": 1, "credibility": 0.728, "freshness": 1, "importance": 0.772, "novelty": 0.5075, "overall": 0.8151 }, "source_name": "Every", "source_tier": "T1.5", "status": "included", "summary_en": "Every is a subscription service focused on AI, offering ideas, apps, and training from practitioners, including a newsletter, podcast, events, and more.", "summary_zh": "Every是一家专注于人工智能领域的订阅服务，提供来自AI实践者的思想、应用和培训，涵盖新闻通讯、播客、活动等。", "tags": [ "newsletter", "ai", "subscription" ], "title": "Every", "url": "https://every.to/", "why_it_matters": "Potentially relevant AI signal for review: Every" }, { "id": "radar_eb5c8c5aef1998b4", "categories": [ "model_release", "open_source" ], "collected_at": "2026-05-22T02:49:38.351+00:00", "confidence": 0.8999, "evidence_notes": [ "[release page](https://github.com/huggingface/transformers/releases/tag/v5.8.0) published at 2026-05-05t16:52:21z indicates deepseek-v4 addition and mentions gemma 4 assistant.", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:52:53.399+00:00", "published_at": "2026-05-05T16:52:21+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.929, "freshness": 0.65, "importance": 0.9005, "novelty": 0.803, "overall": 0.8803 }, "source_name": "Hugging Face Transformers", "source_tier": "T1", "status": "included", "summary_en": "Hugging Face Transformers released version 5.8.0 on May 5, 2026. This release adds support for DeepSeek-V4, a next-generation MoE language model with hybrid attention and other architectural innovations. It also includes Gemma 4 Assistant (details truncated in source).", "summary_zh": "Hugging Face Transformers 于 2026 年 5 月 5 日发布了 v5.8.0 版本。该版本新增了对 DeepSeek-V4 的支持，这是一个下一代 MoE 语言模型，采用混合注意力和其他架构创新。此外还包括 Gemma 4 Assistant（来源中描述被截断）。", "tags": [ "deepseek", "gemma", "transformers", "huggingface" ], "title": "Release 5.8.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.8.0", "why_it_matters": "May affect model capability tracking and product benchmarking: Release 5.8.0" }, { "id": "radar_d02811a704fa7cb5", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:49:38.351+00:00", "confidence": 0.8298, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:52:13.537+00:00", "published_at": "2026-05-13T03:21:23+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.929, "freshness": 0.65, "importance": 0.8585, "novelty": 0.712, "overall": 0.8583 }, "source_name": "Hugging Face Transformers", "source_tier": "T1", "status": "included", "summary_en": "Hugging Face Transformers released patch v5.8.1, primarily to fix Deepseek V4 integration, including fixes for WeightConverter regex, deepseek v4 CSA mask collapse, and other issues.", "summary_zh": "Hugging Face Transformers 发布了 v5.8.1 补丁版本，主要修复 Deepseek V4 集成问题，包括权重转换正则表达式错误和 CSA 掩码崩溃等问题。", "tags": [ "github", "huggingface", "models", "open-source", "patch", "fix", "deepseek" ], "title": "Patch release v5.8.1", "url": "https://github.com/huggingface/transformers/releases/tag/v5.8.1", "why_it_matters": "Potentially relevant AI signal for review: Patch release v5.8.1" }, { "id": "radar_e7b25d6c012d3d58", "categories": [ "model_release", "open_source" ], "collected_at": "2026-05-22T02:49:38.351+00:00", "confidence": 0.2465, "evidence_notes": [ "release v5.9.0 of hugging face transformers, published 2026-05-20.", "new model additions: cohere2moe (cohere command a+), parakeet tdt, hrm-text.", "cohere2moe: mixture-of-experts, hybrid attention (sliding window + full), shared and routed experts, large context window.", "parakeet tdt: mentioned but no details.", "hrm-text: improved autoregressive variant of hierarchical reasoning model, two transformer stacks (slow h, fast l), nested recurrence, prefixlm attention, per-head sigmoid gates, parameterless rmsnorm.", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:51:26.089+00:00", "published_at": "2026-05-20T14:12:54+00:00", "scores": { "ai_relevance": 1, "credibility": 0.929, "freshness": 0.85, "importance": 0.928, "novelty": 0.813, "overall": 0.9224 }, "source_name": "Hugging Face Transformers", "source_tier": "T1", "status": "included", "summary_en": "Hugging Face Transformers released v5.9.0, adding three new models: Cohere2Moe (Command A+, a Mixture-of-Experts with hybrid attention and large context), Parakeet tdt, and HRM-Text (a hierarchical recurrent autoregressive model with dual transformer stacks and PrefixLM attention).", "summary_zh": "Hugging Face Transformers 发布 v5.9.0 版本，新增三个模型：Cohere2Moe（Command A+，混合专家模型，混合滑动窗口和全注意力，支持长上下文）、Parakeet tdt 和 HRM-Text（分层循环自回归模型，使用双Transformer栈和PrefixLM注意力）。", "tags": [ "huggingface", "transformers", "v5.9.0", "cohere2moe", "parakeet", "hrm-text", "github-release" ], "title": "Release v5.9.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.9.0", "why_it_matters": "May affect model capability tracking and product benchmarking: Release v5.9.0" }, { "id": "radar_d842893e2c902cfe", "categories": [ "research" ], "collected_at": "2026-05-22T02:49:38.334+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:50:45.369+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文提出了一种三阶段框架，利用冻结的视觉编码器（DINOv2）和少样本学习从第一人称视角视频中提取护理模拟中的动作时间线，进而评估学习者能力。在22个会话（3.8小时，493个动作）上，通过留一法单样本识别达到57.4%的MOF。研究发现识别准确率与能力呈负相关（rho = -0.524, p = 0.012），能力更高的学生工作流程更多样且更难分类，但更符合协议。这表明识别准确率可作为自动化能力评估中具有教育意义的信号。", "tags": [ "ai", "vision-language-models", "competency-assessment", "nursing-education", "egocentric-video", "few-shot-learning", "dino", "hmm", "computer-vision" ], "title": "AI-Assisted Competency Assessment from Egocentric Video in Simulation-Based Nursing Education", "url": "https://arxiv.org/abs/2605.20233", "why_it_matters": "May add technical evidence for future radar tracking: AI-Assisted Competency Assessment from Egocentric Video in Simulation-Based Nursing Education" }, { "id": "radar_1eea04c5503e2f54", "categories": [ "research" ], "collected_at": "2026-05-22T02:49:38.334+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:50:11.242+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文分析了在无标签视频中学习潜在动作模型时，外生状态（如背景杂波）带来的干扰。通过扩展线性潜在动作模型显式建模外生状态，作者发现标准重建目标会导致潜在动作编码未来观察中的外生信息，而专注于内生成分的表征空间学习是减轻噪声干扰的关键。此外，先前提出的辅助目标（如动作监督）能在理论上保证潜在动作在不同外生状态间的一致性。实验在线性和非线性模型上验证了这些发现。", "tags": [ "latent action models", "exogenous state", "representation learning", "video analysis", "unsupervised learning", "computer vision" ], "title": "Why Latent Actions Fail, and How to Prevent It", "url": "https://arxiv.org/abs/2605.20223", "why_it_matters": "May add technical evidence for future radar tracking: Why Latent Actions Fail, and How to Prevent It" }, { "id": "radar_6e6fa6bee7eefa06", "categories": [ "research" ], "collected_at": "2026-05-22T02:49:38.334+00:00", "confidence": 0.8748, "evidence_notes": [ "source: arxiv cs.cv, url: https://arxiv.org/abs/2605.20211, published at 2026-05-21t04:00:00.000z, collected at 2026-05-22t02:49:38.334z", "abstract states: 'evaluate the performance of this vlm-based approach using several prompting strategies with gemini 3, but ultimately found that none of them could outperform statistical baselines.'", "only the abstract and metadata are available; full text not reviewed.", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:49:39.832+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.6, "credibility": 0.899, "freshness": 1, "importance": 0.669, "novelty": 0.681, "overall": 0.7348 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文探索使用视觉语言模型（VLM）检测教育视频中的注意力，但发现Gemini 3的多种提示策略均未能超越传统统计基线，揭示了VLM在实时教育诊断中的局限性。", "tags": [ "academic", "arxiv", "computer-vision", "vision-language-models", "education", "eye-tracking" ], "title": "Leveraging Vision-Language Models to Detect Attention in Educational Videos", "url": "https://arxiv.org/abs/2605.20211", "why_it_matters": "May add technical evidence for future radar tracking: Leveraging Vision-Language Models to Detect Attention in Educational Videos" }, { "id": "radar_2a46a79464ecf9bb", "categories": [ "other" ], "collected_at": "2026-05-22T02:44:08.603+00:00", "confidence": 0.7028, "evidence_notes": [ "source url: https://www.anthropic.com/research, collected at 2026-05-22t02:44:08.603z", "raw text: 'anthropic is an ai safety and research company that's working to build reliable, interpretable, and steerable ai systems.'", "metadata links include research teams and two specific research articles with titles and urls.", "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:48:55.98+00:00", "scores": { "ai_relevance": 0.5, "credibility": 0.867, "freshness": 1, "importance": 0.521, "novelty": 0.4075, "overall": 0.6357 }, "source_name": "Anthropic Research", "source_tier": "T1", "status": "needs_review", "summary_en": "Anthropic is an AI safety and research company focused on building reliable, interpretable, and steerable AI systems. Its research page lists research teams (e.g., Alignment, Interpretability, Economic Research, Societal Impacts) and recent projects (e.g., Natural Language Autoencoders, Teaching Claude).", "summary_zh": "Anthropic是一家专注于AI安全与研究的公司，致力于构建可靠、可解释和可控的AI系统。其研究页面列出了研究团队（如对齐、可解释性、经济研究、社会影响）和最新研究项目（如自然语言自编码器、教学Claude）。", "tags": [ "anthropic", "research", "safety", "alignment", "interpretability" ], "title": "Research", "url": "https://www.anthropic.com/research", "why_it_matters": "Potentially relevant AI signal for review: Research" }, { "id": "radar_9456669c322fdb75", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:44:08.587+00:00", "confidence": 0.8631, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:47:50.567+00:00", "published_at": "2026-05-19T07:07:51+00:00", "scores": { "ai_relevance": 1, "credibility": 0.929, "freshness": 0.85, "importance": 0.886, "novelty": 0.722, "overall": 0.9003 }, "source_name": "Anthropic Python SDK", "source_tier": "T1", "status": "included", "summary_en": "Anthropic Python SDK v0.103.0 released, adding support for self-hosted sandboxes in CMA with sandbox helpers.", "summary_zh": "Anthropic Python SDK v0.103.0 发布，主要新增功能：在 CMA 中支持自托管沙箱，并提供沙箱辅助工具。", "tags": [ "anthropic", "python-sdk", "self-hosted-sandboxes", "cma", "release" ], "title": "v0.103.0", "url": "https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.103.0", "why_it_matters": "Potentially relevant AI signal for review: v0.103.0" }, { "id": "radar_9dbbb2885ccf5e22", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:44:08.587+00:00", "confidence": 0.8298, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:47:14.111+00:00", "published_at": "2026-05-19T15:43:14+00:00", "scores": { "ai_relevance": 1, "credibility": 0.929, "freshness": 0.85, "importance": 0.886, "novelty": 0.722, "overall": 0.9003 }, "source_name": "Anthropic Python SDK", "source_tier": "T1", "status": "included", "summary_en": "Anthropic Python SDK released v0.103.1, a patch version that fixes a bug in the runner where tool calls not owned by SessionToolRunner were incorrectly skipped.", "summary_zh": "Anthropic Python SDK 发布 v0.103.1 补丁版本，修复了 runner 中跳过不属于 SessionToolRunner 的工具调用的问题。", "tags": [ "anthropic", "python-sdk", "bug-fix", "release", "sdk" ], "title": "v0.103.1", "url": "https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.103.1", "why_it_matters": "Potentially relevant AI signal for review: v0.103.1" }, { "id": "radar_cadc705c71d8e5d7", "categories": [ "product_update", "open_source" ], "collected_at": "2026-05-22T02:44:08.587+00:00", "confidence": 0.2298, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:46:40.459+00:00", "published_at": "2026-05-21T20:01:49+00:00", "scores": { "ai_relevance": 1, "credibility": 0.929, "freshness": 1, "importance": 0.898, "novelty": 0.748, "overall": 0.9216 }, "source_name": "Anthropic Python SDK", "source_tier": "T1", "status": "included", "summary_en": "Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "summary_zh": "Anthropic Python SDK v0.104.0 发布，新增流式传输时思考块增量中的思考令牌计数 beta 支持。", "tags": [ "developer", "github", "official", "open-source", "sdk", "thinking-token-count", "streaming", "api" ], "title": "v0.104.0", "url": "https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.104.0", "why_it_matters": "May change available building blocks for teams evaluating open implementations: v0.104.0" }, { "id": "radar_0e12774457b51ae2", "categories": [ "research" ], "collected_at": "2026-05-22T02:44:08.577+00:00", "confidence": 0.2248, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:45:58.223+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文研究了量化LLaMA-3.1（8B）模型在定性分析中的性能，尤其是不同量化级别（2-8位）和类型的影响。针对低比特模型出现幻觉和不稳定结果的问题，提出了一种量化感知的多轮提示验证方法，通过控制步骤减少幻觉。实验使用82份访谈记录，以BF16模型和人工编码的黄金标准为基准。结果表明，8位模型最接近标准；4位模型在应用方法后变得稳定；3位和2位模型性能下降但有所改善。该方法有助于低资源LLM在低成本下实现更稳定、准确的定性分析。", "tags": [ "academic", "arxiv", "language-models", "papers", "research", "quantization", "llm", "qualitative analysis", "hallucination", "llama" ], "title": "Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification", "url": "https://arxiv.org/abs/2605.20193", "why_it_matters": "May add technical evidence for future radar tracking: Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification" }, { "id": "radar_d9de0c996dcddce8", "categories": [ "research" ], "collected_at": "2026-05-22T02:44:08.577+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:45:22.048+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该研究利用BERT大型语言模型分析Decentraland（MANA代币）的Discord社区情绪，并结合价格、交易量、市值等多模态金融数据，构建LSTM模型预测代币收益。结果显示社区情绪中性偏正面，多模态模型显著优于仅基于价格的基线模型，证明社区衍生信号对虚拟经济预测的价值。", "tags": [ "large language models", "sentiment analysis", "multi-modal", "cryptocurrency", "decentraland", "mana" ], "title": "Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token", "url": "https://arxiv.org/abs/2605.20192", "why_it_matters": "May add technical evidence for future radar tracking: Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token" }, { "id": "radar_54ff21cc34521714", "categories": [ "research" ], "collected_at": "2026-05-22T02:44:08.577+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:44:50.893+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "This paper investigates how LLMs represent disability by simulating social media posts from the perspective of individuals with disabilities, comparing them with posts by real disabled people. It finds that LLMs tend to idealize disability experiences with overly positive stereotypes, and exhibit negative bias by disproportionately associating topics like career and entertainment with non-disabled individuals.", "summary_zh": "本论文研究大语言模型（LLMs）如何通过模拟残疾人群体的视角生成社交媒体帖子，并与真实残疾人的帖子对比。发现LLMs倾向于理想化残疾经历，产生过度积极的刻板印象，同时存在负面偏见，将某些话题（如职业、娱乐）与健全人过度关联。", "tags": [ "llm", "disability", "bias", "stereotypes", "representation", "social media" ], "title": "Shiny Stories, Hidden Struggles: Investigating the Representation of Disability Through the Lens of LLMs", "url": "https://arxiv.org/abs/2605.20191", "why_it_matters": "May add technical evidence for future radar tracking: Shiny Stories, Hidden Struggles: Investigating the Representation of Disability Through the Lens of LLMs" }, { "id": "radar_e4bf0f7da04d7265", "categories": [ "model_release", "product_update", "research" ], "collected_at": "2026-05-22T02:44:08.561+00:00", "confidence": 0.8695, "evidence_notes": [ "source: https://www.anthropic.com/news, collected at 2026-05-22t02:44:08.561z", "link: introducing claude opus 4.7 (apr 16, 2026) - https://www.anthropic.com/news/claude-opus-4-7", "link: introducing claude design (apr 17, 2026) - https://www.anthropic.com/news/claude-design-anthropic-labs", "link: project glasswing (apr 7, 2026) - https://www.anthropic.com/glasswing", "link: what 81,000 people want from ai (mar 18, 2026) - https://www.anthropic.com/81k-interviews", "metadata indicates source_category: official_blog, source_tier: t1", "published_at missing; freshness uses collected_at", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:44:10.022+00:00", "scores": { "ai_relevance": 1, "credibility": 0.867, "freshness": 1, "importance": 0.937, "novelty": 0.813, "overall": 0.9298 }, "source_name": "Anthropic News", "source_tier": "T1", "status": "included", "summary_en": "Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "summary_zh": "Anthropic 新闻室页面（收集于2026年5月22日）展示了近期公告，包括 Claude Opus 4.7 的发布（2026年4月16日）、Claude Design（2026年4月17日）、Project Glasswing（2026年4月7日）以及来自81,000次用户访谈的见解（2026年3月18日）。", "tags": [ "anthropic", "claude", "company", "official", "product", "safety" ], "title": "Newsroom", "url": "https://www.anthropic.com/news", "why_it_matters": "May affect model capability tracking and product benchmarking: Newsroom" }, { "id": "radar_7acba60f65637c04", "categories": [ "media_interview", "opinion" ], "collected_at": "2026-05-22T02:37:33.653+00:00", "confidence": 0.1834, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:43:23.268+00:00", "published_at": "2026-03-23T16:28:42+00:00", "scores": { "ai_relevance": 1, "credibility": 0.65, "freshness": 0.45, "importance": 0.793, "novelty": 0.618, "overall": 0.7553 }, "source_name": "Lex Fridman", "source_tier": "T2", "status": "included", "summary_en": "Lex Fridman podcast #494 features Jensen Huang, co-founder and CEO of NVIDIA, discussing NVIDIA's rise to become the world's most valuable company at $4 trillion, the AI revolution, AI scaling laws, supply chain, power needs, TSMC and Taiwan, Jensen's engineering leadership philosophy, AGI timeline, consciousness, and more.", "summary_zh": "Lex Fridman播客第494期采访了NVIDIA联合创始人兼CEO黄仁勋，讨论了NVIDIA成为全球市值最高公司（4万亿美元）的历程、AI革命、AI缩放定律、供应链、电力需求、台积电与台湾、杰森的工程领导哲学、AGI时间线、意识等话题。", "tags": [ "nvidia", "jensen huang", "ai scaling laws", "agi", "data centers", "ai hardware" ], "title": "#494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution", "url": "https://lexfridman.com/jensen-huang", "why_it_matters": "Potentially relevant AI signal for review: #494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution" }, { "id": "radar_c2a3f9042e83816e", "categories": [ "media_interview" ], "collected_at": "2026-05-22T02:37:33.653+00:00", "confidence": 0.8166, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:41:55.489+00:00", "published_at": "2026-05-06T22:06:47+00:00", "scores": { "ai_relevance": 1, "credibility": 0.65, "freshness": 0.65, "importance": 0.793, "novelty": 0.618, "overall": 0.7753 }, "source_name": "Lex Fridman", "source_tier": "T2", "status": "included", "summary_en": "In this episode of Lex Fridman podcast, guests Jean-Baptiste Kempf (lead developer of VLC and president of VideoLAN) and Kieran Kunhya (longtime FFmpeg contributor and codec engineer) discuss the history and technology behind FFmpeg and VLC, video codecs, open-source community controversies (e.g., FFmpeg vs Google, Libav fork), reverse engineering codecs, assembly code, Rust programming, ultra-low latency streaming, AV2 codec, and video archiving.", "summary_zh": "本期Lex Fridman播客采访了VLC主要开发者兼VideoLAN主席Jean-Baptiste Kempf以及FFmpeg长期贡献者、编解码工程师Kieran Kunhya。他们讨论了FFmpeg和VLC的历史、视频编解码技术、开源社区争议（如FFmpeg与Google的冲突、Libav分叉）、反编译编解码器、汇编代码、Rust语言、低延迟流媒体、AV2编解码器以及视频归档等话题。", "tags": [ "ffmpeg", "vlc", "video codecs", "open source", "podcast", "interview", "lex fridman" ], "title": "#496 – FFmpeg: The Incredible Technology Behind Video on the Internet", "url": "https://lexfridman.com/ffmpeg", "why_it_matters": "Potentially relevant AI signal for review: #496 – FFmpeg: The Incredible Technology Behind Video on the Internet" }, { "id": "radar_dffd1aeebae810c0", "categories": [ "open_source" ], "collected_at": "2026-05-22T02:37:33.552+00:00", "confidence": 0.8641, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05", "source is the official github repository (https://github.com/openai/openai-cookbook), classified as t1 and official.", "metadata indicates 73,681 stars, 12,461 forks, and 185 open issues as of 2026-05-21t16:21:17z.", "no new release, breaking change, or novel content is reported; the item reflects existing repository status." ], "language": "en", "processed_at": "2026-05-22T02:41:24.37+00:00", "published_at": "2026-05-21T16:21:17+00:00", "scores": { "ai_relevance": 1, "credibility": 0.9373, "freshness": 1, "importance": 0.8258, "novelty": 0.586, "overall": 0.8855 }, "source_name": "OpenAI Cookbook", "source_tier": "T1", "status": "included", "summary_en": "The OpenAI Cookbook is a GitHub repository that provides examples and guides for using the OpenAI API. As of May 21, 2026, it has 73,681 stars, 12,461 forks, and 185 open issues.", "summary_zh": "OpenAI Cookbook 是一个 GitHub 仓库，提供使用 OpenAI API 的示例和指南。截至 2026年5月21日，该仓库拥有 73681 颗星、12461 个分支和 185 个开放问题。", "tags": [ "developer", "github", "official", "open-source" ], "title": "openai/openai-cookbook repository metadata", "url": "https://github.com/openai/openai-cookbook", "why_it_matters": "The OpenAI Cookbook is an official, high-engagement repository (73,681 stars) providing foundational API examples for developers." }, { "id": "radar_31ed662e4913e460", "categories": [ "research", "agent" ], "collected_at": "2026-05-22T02:37:33.542+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:40:38.537+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "This paper introduces OSCToM, an RL-guided approach for generating high-order Theory of Mind conflicts to improve LLMs' recursive reasoning in complex social settings. It achieves 76% accuracy on FANToM and is 6x more efficient in data synthesis.", "summary_zh": "本文提出OSCToM，一种利用强化学习生成高阶心智理论冲突的方法，旨在提升大语言模型在复杂社交场景中的递归推理能力。在FANToM基准上达到76%准确率，数据合成效率提高6倍。", "tags": [ "theory of mind", "llm", "reinforcement learning", "adversarial generation", "benchmark" ], "title": "OSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind", "url": "https://arxiv.org/abs/2605.20423", "why_it_matters": "May add technical evidence for future radar tracking: OSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind" }, { "id": "radar_2c3877a317d6de3c", "categories": [ "research" ], "collected_at": "2026-05-22T02:37:33.542+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:39:59.92+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "This paper proposes COSMO-Agent, a tool-augmented reinforcement learning framework that bridges the CAD-CAE semantic gap in 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. A multi-constraint reward and an industry-aligned dataset covering 25 component categories are introduced. Experiments show COSMO-Agent training substantially improves small open-source LLMs, exceeding larger models in feasibility, efficiency, and stability.", "summary_zh": "本文提出COSMO-Agent，一种工具增强的强化学习框架，通过将CAD生成、CAE求解、结果解析和几何修改整合为交互式RL环境，使大语言模型（LLM）学会编排外部工具并迭代修改参数化几何，以弥合工业设计中CAD-CAE的语义鸿沟。该框架采用多约束奖励，并贡献了包含25个组件类别的工业对齐数据集。实验表明，COSMO-Agent训练显著提升了小型开源LLM在可行性、效率和稳定性方面的表现，超越了大模型和强闭源模型。", "tags": [ "tool-augmented", "reinforcement learning", "llm", "cad-cae", "optimization" ], "title": "Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration", "url": "https://arxiv.org/abs/2605.20190", "why_it_matters": "May add technical evidence for future radar tracking: Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration" }, { "id": "radar_bb285e9034313ce1", "categories": [ "agent", "research" ], "collected_at": "2026-05-22T02:37:33.542+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:39:19.036+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "Proposes SOLAR, a self-optimizing lifelong autonomous reasoner that leverages parameter-level meta-learning and multi-level reinforcement learning for continual adaptation without gradient updates, outperforming strong baselines on commonsense, math, medical, coding, social, and logical reasoning tasks.", "summary_zh": "提出了一种名为 SOLAR 的自我优化终身自主推理器，通过参数级元学习和多级强化学习，在无需梯度更新的情况下实现持续适应，在常识、数学、医学、编程、社会与逻辑推理任务上超越强基线。", "tags": [ "lifelong learning", "continual learning", "meta-learning", "autonomous agent" ], "title": "SOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation", "url": "https://arxiv.org/abs/2605.20189", "why_it_matters": "May add technical evidence for future radar tracking: SOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation" }, { "id": "radar_f4ae49772d646fad", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:37:33.523+00:00", "confidence": 0.8671, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:38:47.977+00:00", "published_at": "2026-05-20T00:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.953, "freshness": 0.85, "importance": 0.898, "novelty": 0.722, "overall": 0.9115 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "OpenAI announces the next phase of its Education for Countries initiative, expanding AI adoption in schools with new partnerships, teacher training, and tools to improve global learning outcomes.", "summary_zh": "OpenAI宣布其“国家教育”计划进入新阶段，通过新合作伙伴关系、教师培训和工具，扩大AI在学校中的应用，以改善全球学习成果。", "tags": [ "education", "ai adoption", "teacher training", "global learning", "partnerships" ], "title": "The next phase of OpenAI’s Education for Countries", "url": "https://openai.com/index/the-next-phase-of-education-for-countries", "why_it_matters": "Potentially relevant AI signal for review: The next phase of OpenAI’s Education for Countries" }, { "id": "radar_e3e77416ac7655d1", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:37:33.523+00:00", "confidence": 0.2339, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:38:15.788+00:00", "published_at": "2026-05-20T00:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.953, "freshness": 0.85, "importance": 0.898, "novelty": 0.722, "overall": 0.9115 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "OpenAI News blog describes how Ramp engineers use Codex with GPT-5.5 to accelerate code review, reducing feedback time from hours to minutes.", "summary_zh": "OpenAI新闻发布博客，介绍Ramp工程师如何利用Codex（搭配GPT-5.5）加速代码审查，将反馈时间从数小时缩短至几分钟。", "tags": [ "codex", "gpt-5.5", "code review", "ramp", "engineering", "productivity" ], "title": "How Ramp engineers accelerate code review with Codex", "url": "https://openai.com/index/ramp", "why_it_matters": "Potentially relevant AI signal for review: How Ramp engineers accelerate code review with Codex" }, { "id": "radar_71898b8e7aeee243", "categories": [ "product_update" ], "collected_at": "2026-05-22T02:37:33.523+00:00", "confidence": 0.8338, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-22T02:37:44.337+00:00", "published_at": "2026-05-21T12:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.953, "freshness": 1, "importance": 0.843, "novelty": 0.702, "overall": 0.8825 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "AdventHealth is using OpenAI's ChatGPT for Healthcare to streamline workflows, reduce administrative burden, and return more time to patient care.", "summary_zh": "AdventHealth 使用 OpenAI 的 ChatGPT for Healthcare 来简化工作流程、减轻行政负担，从而将更多时间用于患者护理。", "tags": [ "chatgpt", "healthcare", "adventhealth", "workflow", "patient care" ], "title": "AdventHealth advances whole-person care with OpenAI", "url": "https://openai.com/index/adventhealth", "why_it_matters": "Potentially relevant AI signal for review: AdventHealth advances whole-person care with OpenAI" }, { "id": "radar_88267b869153e7d3", "categories": [ "product_update", "open_source" ], "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8799, "evidence_notes": [ "release url: https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.102.0", "published at: 2026-05-13t18:12:29z", "collected at: 2026-05-21t09:45:24.786z", "changelog: https://github.com/anthropics/anthropic-sdk-python/compare/v0.101.0...v0.102.0", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T09:56:41.267+00:00", "published_at": "2026-05-13T18:12:29+00:00", "scores": { "ai_relevance": 1, "credibility": 0.929, "freshness": 0.65, "importance": 0.898, "novelty": 0.748, "overall": 0.8866 }, "source_name": "Anthropic Python SDK", "source_tier": "T1", "status": "included", "summary_en": "Anthropic Python SDK v0.102.0 released, adding BetaManagedAgentsSearchResultBlock types, cache diagnostics support, and eager validation for Pydantic iterators.", "summary_zh": "Anthropic Python SDK v0.102.0 发布，新增 Beta 版 ManagedAgentsSearchResultBlock 类型、缓存诊断支持，并优化了 Pydantic 迭代器验证。", "tags": [ "anthropic", "sdk", "python", "api", "managed_agents", "cache_diagnostics", "github_release" ], "title": "v0.102.0", "url": "https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.102.0", "why_it_matters": "May change available building blocks for teams evaluating open implementations: v0.102.0" }, { "id": "radar_26255f33ab09d8c3", "categories": [ "research" ], "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T09:48:36.762+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文评估了大型语言模型（Gemini 3.0 Flash）在利用个人健康记录（PHR）回答用户健康查询方面的潜力。从三种来源（网络搜索、聊天机器人模板、患者呼叫）收集了2,257个查询，并与1,945个去识别化PHR匹配。Gemini在无PHR上下文、基础摘要或完整临床笔记条件下生成回答。使用了SHARP框架和新的PHR特定错误框架进行评估。结果表明，使用PHR数据后回答有用性显著提升（p<0.001），并在安全性、准确性、相关性和个性化方面有潜在改进。同时发现了时间定向障碍和罕见但重要的虚构等缺陷。研究支持PHR数据的潜力，并提供了监控LLM回答缺陷的框架。", "tags": [ "llm", "evaluation", "health", "personalized health", "safety", "benchmark" ], "title": "Evaluating the Utility of Personal Health Records in Personalized Health AI", "url": "https://arxiv.org/abs/2605.18937", "why_it_matters": "May add technical evidence for future radar tracking: Evaluating the Utility of Personal Health Records in Personalized Health AI" }, { "id": "radar_8d5fd97b76b9d62a", "categories": [ "research" ], "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8748, "evidence_notes": [ "source: arxiv cs.ai, paper id 2605.18818, published 2026-05-21. url: https://arxiv.org/abs/2605.18818", "abstract states: 'ocr, not language-model parsing, dominates end-to-end latency' and 'system saturates at a concurrency determined by shared gpu-inference capacity rather than worker count.'", "title and abstract confirm the microservice architecture and design decisions.", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T09:47:58.567+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文提出了一种用于文档AI的微服务架构，将分类、OCR和LLM结构化字段提取等多个模型集成到生产级流水线中。关键设计包括混合分类、GPU推理与CPU编排分离、异步IO处理及独立水平扩展。通过批量剖析发现两个意外结果：OCR端到端延迟占比最大，且系统并发受限于共享GPU推理容量而非工作进程数。该工作旨在为生产环境中文档理解系统的构建提供实践架构模式。", "tags": [ "document ai", "microservice architecture", "ocr", "llm", "production" ], "title": "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production", "url": "https://arxiv.org/abs/2605.18818", "why_it_matters": "May add technical evidence for future radar tracking: Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production" }, { "id": "radar_2a64bcb1fa50406f", "categories": [ "research" ], "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8249, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T09:47:21.409+00:00", "published_at": "2026-05-21T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文提出了一种系统方法，通过从定义良好的随机过程中生成合成序列（称为数据探针），来研究数据特征如何影响大语言模型（LLM）的性能、泛化性和鲁棒性。作者认为当前依赖大规模实验和启发式方法缺乏原则性理解，而数据探针方法结合典型集等理论概念，可为理解数据在LLM训练和推理中的作用提供基础性见解。", "tags": [ "llm", "data", "probes", "understanding", "performance" ], "title": "Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance", "url": "https://arxiv.org/abs/2605.18801", "why_it_matters": "May add technical evidence for future radar tracking: Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance" }, { "id": "radar_51c3a2bb06ffd0ba", "categories": [ "research", "model_release" ], "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8671, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T09:46:12.648+00:00", "published_at": "2026-05-20T00:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.953, "freshness": 0.85, "importance": 0.94, "novelty": 0.813, "overall": 0.9335 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "OpenAI 的一个模型解决了已有80年历史的单位距离问题，推翻了几何学中的一个核心猜想，标志着人工智能在数学领域的里程碑。", "tags": [ "ai", "mathematics", "geometry", "conjecture", "reasoning" ], "title": "An OpenAI model has disproved a central conjecture in discrete geometry", "url": "https://openai.com/index/model-disproves-discrete-geometry-conjecture", "why_it_matters": "May affect model capability tracking and product benchmarking: An OpenAI model has disproved a central conjecture in discrete geometry" }, { "id": "radar_2237852952a864df", "categories": [ "product_update" ], "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8631, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T08:14:09.85+00:00", "published_at": "2026-05-15T22:30:20+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.929, "freshness": 0.85, "importance": 0.831, "novelty": 0.702, "overall": 0.8563 }, "source_name": "OpenAI Python SDK", "source_tier": "T1", "status": "included", "summary_en": "OpenAI Python SDK released v2.37.0 with new features: added service_tier parameter to responses compact method, support for eagerly validating pydantic iterators, removed unnecessary client_id when using workload identity provider for auth; fixed missing f-string prefix in file type error message.", "summary_zh": "OpenAI Python SDK 发布 v2.37.0 版本，主要新增功能包括：在 responses compact 方法中添加 service_tier 参数、支持即时验证 pydantic 迭代器、移除使用 workload identity provider 认证时不必要的 client_id；同时修复了文件类型错误消息中缺少 f-string 前缀的问题。", "tags": [ "openai", "python-sdk", "release", "v2.37.0" ], "title": "v2.37.0", "url": "https://github.com/openai/openai-python/releases/tag/v2.37.0", "why_it_matters": "Potentially relevant AI signal for review: v2.37.0" }, { "id": "radar_2d9d801ed264c310", "categories": [ "other" ], "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.2278, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T04:20:20.352+00:00", "published_at": "2025-06-27T08:46:37+00:00", "scores": { "ai_relevance": 1, "credibility": 0.917, "freshness": 0.25, "importance": 0.781, "novelty": 0.5075, "overall": 0.7827 }, "source_name": "DeepSeek-V3", "source_tier": "T1", "status": "included", "summary_en": "DeepSeek-V3 released v1.0.0, which is solely for archival purposes and DOI generation, with no substantive content.", "summary_zh": "DeepSeek-V3 发布 v1.0.0 版本，仅用于归档和生成 DOI，无其他实质性内容。", "tags": [ "deepseek", "github", "release", "archival", "doi" ], "title": "v1.0.0", "url": "https://github.com/deepseek-ai/DeepSeek-V3/releases/tag/v1.0.0", "why_it_matters": "Potentially relevant AI signal for review: v1.0.0" }, { "id": "radar_eaacfd054a1f2b19", "categories": [ "research" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:59:50.527+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "HELLoRA 是一种针对混合专家（MoE）模型的参数高效微调方法，通过仅将 LoRA 模块附加到每层最常激活的专家上，减少了可训练参数和适配器 FLOPs，同时提升下游性能。在 OlMoE、Mixtral 和 DeepSeekMoE 等模型上，HELLoRA 相比标准 LoRA 显著降低了参数量并提高了准确率和训练吞吐量。", "tags": [ "lora", "mixture-of-experts", "parameter-efficient fine-tuning", "arxiv" ], "title": "HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models", "url": "https://arxiv.org/abs/2605.18795", "why_it_matters": "May add technical evidence for future radar tracking: HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models" }, { "id": "radar_6b75310f17f6510b", "categories": [ "research" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:59:11.747+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文提出了一种基于B样条的分解框架，用于变压器模型压缩。通过约束耦合矩阵-张量分解和鲁棒交替最小二乘算法R-CMTF-BSD，实现了参数大幅减少同时保持竞争精度。在Vision和Swin Transformer架构上的实验验证了其有效性。", "tags": [ "transformer compression", "decoupling", "b-spline", "neural network compression", "tensor factorization", "academic", "arxiv", "machine-learning", "papers" ], "title": "Robust Basis Spline Decoupling for the Compression of Transformer Models", "url": "https://arxiv.org/abs/2605.18794", "why_it_matters": "May add technical evidence for future radar tracking: Robust Basis Spline Decoupling for the Compression of Transformer Models" }, { "id": "radar_1488431b28a1ca16", "categories": [ "research" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:58:32.503+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文提出一种基于空间和时间熵诊断的维度平衡框架，通过低秩矩阵压缩空间维度和扩展时间视野，以改善大规模时空预测性能。在交通、气象和流行病数据集上实验显示显著准确率提升。", "tags": [ "academic", "arxiv", "machine-learning", "papers", "research", "spatiotemporal", "prediction" ], "title": "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance", "url": "https://arxiv.org/abs/2605.18793", "why_it_matters": "May add technical evidence for future radar tracking: Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance" }, { "id": "radar_47d31ae62f736dc6", "categories": [ "research", "benchmark" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:57:55.57+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "Artifact-Bench是一个用于评估多模态大语言模型（MLLM）检测和分析AI生成视频伪影的综合基准。它建立了涵盖照片级真实、动画和CG风格视频的三级分层伪影分类法，并定义了三个互补任务：真实与AI生成视频分类、成对真实感比较、细粒度伪影识别。对19个主流MLLM的实验表明，它们在伪影感知和推理方面存在显著局限，许多模型在挑战性场景下表现接近甚至低于随机水平，且MLLM判断与人类感知偏好严重不一致。", "tags": [ "arxiv", "benchmark", "mllm", "ai-generated video", "artifact detection" ], "title": "Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos", "url": "https://arxiv.org/abs/2605.18984", "why_it_matters": "May add technical evidence for future radar tracking: Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos" }, { "id": "radar_0bed1c17d6ef1231", "categories": [ "research" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:57:26.093+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文研究了自监督特征在艺术作品分类和检索中的有效性，使用DINO和CLIP模型进行实验，结果表明自监督骨干网络能持续提升分类性能，并讨论了在虚拟现实博物馆导航等实际应用中的潜力。", "tags": [ "art classification", "self-supervised learning", "computer vision", "dino", "clip", "fine-grained classification" ], "title": "Harnessing Self-Supervised Features for Art Classification", "url": "https://arxiv.org/abs/2605.18974", "why_it_matters": "May add technical evidence for future radar tracking: Harnessing Self-Supervised Features for Art Classification" }, { "id": "radar_9d2bda13a994f350", "categories": [ "research", "benchmark" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:56:46.552+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "MotionMERGE 是一个统一框架，通过在大语言模型中显式建模局部和时序运动，实现细粒度的人体运动编辑、推理和生成。它提出了 ReasoningAware Granularity-Synergy 预训练策略，并构建了包含 981K 三元组的大规模数据集 MotionFineEdit，支持细粒度时空校正和运动链式推理。实验表明，该方法在运动生成、理解和编辑任务上表现更精确，且具有零样本泛化能力。", "tags": [ "motion generation", "motion editing", "fine-grained control", "llm", "dataset", "benchmark" ], "title": "MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation", "url": "https://arxiv.org/abs/2605.18956", "why_it_matters": "May add technical evidence for future radar tracking: MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation" }, { "id": "radar_1ca98fbb9b413a4e", "categories": [ "research", "agent" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:56:09.772+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "ReacTOD是一种用于零样本对话状态跟踪的有界神经符号架构。它将NLU重构为自校正ReAct循环中的离散工具调用，并配合确定性验证。在MultiWOZ 2.1上，gpt-oss-20B达到52.71%的联合目标准确率（提升14个百分点），Qwen3-8B达到47.34%。在SGD基准上，Claude-Opus-4.6达到80.68%的JGA。该架构相比单次推理准确率提升最高9.3%，错误自校正率达93.1%。", "tags": [ "neuro-symbolic", "dialogue state tracking", "zero-shot", "llm", "react", "multiwoz", "sgd" ], "title": "ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking", "url": "https://arxiv.org/abs/2605.19077", "why_it_matters": "May add technical evidence for future radar tracking: ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking" }, { "id": "radar_cb96e49c0763bb7c", "categories": [ "research", "benchmark" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8748, "evidence_notes": [ "abstract from arxiv:2605.19069v1, published 2026-05-20, accessed via https://arxiv.org/abs/2605.19069", "dataset available at https://huggingface.co/datasets/perle-ai/asr_code_switch", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:55:24.416+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文提出一个基准，评估五个商业ASR系统在代码切换语音上的表现，涵盖四个语言对（埃及阿拉伯语-英语、沙特阿拉伯语-英语、波斯语-英语、德语-英语）。每个数据集包含300个样本，通过两阶段流水线选择。ElevenLabs Scribe v2在总体WER（13.2%）和BERTScore（0.936）上表现最佳。作者认为对于阿拉伯语和波斯语，BERTScore因音译差异更可靠。数据集已公开。", "tags": [ "asr", "code-switching", "commercial asr", "speech recognition", "multilingual" ], "title": "Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German", "url": "https://arxiv.org/abs/2605.19069", "why_it_matters": "May add technical evidence for future radar tracking: Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German" }, { "id": "radar_aed4721e852e3137", "categories": [ "research" ], "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-21T01:54:47.787+00:00", "published_at": "2026-05-20T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "这篇论文提出了低资源NLP评估中的“注释稀缺悖论”，指出模型规模化能力远超所需的人类评估基础设施。文章回顾了2014年至今的三个阶段，并探讨了数据增强、模型评估等应对措施，呼吁转向社区嵌入的评估范式。", "tags": [ "low-resource nlp", "evaluation", "annotation", "survey", "linguistics" ], "title": "The Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints", "url": "https://arxiv.org/abs/2605.19066", "why_it_matters": "May add technical evidence for future radar tracking: The Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints" }, { "id": "radar_d76af0556f232088", "categories": [ "research" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:29:58.738+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文系统性研究了Apple M3 Ultra（60核GPU，512GB统一内存）上扩散模型的实时推理优化。通过10个阶段的实验，探索了CoreML转换、量化、Token Merging、神经引擎利用等技术。最终，结合CoreML转换的蒸馏模型SDXS-512与三线程相机流水线，在512x512分辨率下实现了22.7 FPS的实时图像转换。研究表明，针对CUDA的优化（如量化加速、并行推理）在Apple Silicon上无效，揭示了不同的优化特性，并提供了实践指南。", "tags": [ "diffusion models", "apple silicon", "inference optimization", "real-time image generation", "coreml" ], "title": "Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra", "url": "https://arxiv.org/abs/2605.16259", "why_it_matters": "May add technical evidence for future radar tracking: Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra" }, { "id": "radar_2bef32d76ec9b113", "categories": [ "research" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:29:23.266+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "这篇论文提出 F^3A，一种无需训练的视觉令牌剪枝路由器，用于多模态语言模型，通过任务条件证据搜索在固定视觉令牌预算下高效分配令牌，避免额外LLM前向传播。", "tags": [ "visual token pruning", "vision-language models", "multimodal", "model efficiency", "f3a" ], "title": "How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A", "url": "https://arxiv.org/abs/2605.16359", "why_it_matters": "May add technical evidence for future radar tracking: How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A" }, { "id": "radar_50c5d288e25844a8", "categories": [ "research", "open_source" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:28:31.751+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文提出StrLoRA框架，用于多模态大语言模型在流式连续视觉指令调优（Streaming CVIT）中学习。Streaming CVIT是一个新的、更真实的设定，其中数据以动态混合任务的连续块形式到达。StrLoRA采用正则化的两阶段专家路由，首先通过文本指令进行任务感知的专家选择，然后通过跨模态注意力进行词元级专家加权，并引入路由稳定性正则化。在StrCVIT基准上，StrLoRA显著优于现有方法。", "tags": [ "continual learning", "visual instruction tuning", "multimodal llm", "lora", "streaming" ], "title": "StrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs", "url": "https://arxiv.org/abs/2605.16353", "why_it_matters": "May change available building blocks for teams evaluating open implementations: StrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs" }, { "id": "radar_d15b00276fe9c755", "categories": [ "research" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8249, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:27:54.445+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文提出一个基于光子统计的事件相机概率模型，统一了静态场景噪声事件与S曲线的描述。基于该模型，提出Noise2Params方法，通过最小化观测噪声事件分布误差来确定相机参数（B、α、θ），仅需静态均匀场景记录。实验表明，利用模型生成的合成噪声数据训练CNN，在静态场景重建中优于仅用实验数据训练的模型。", "tags": [ "event camera", "probabilistic model", "calibration", "noise", "cnn", "synthetic data", "photon statistics" ], "title": "Noise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model", "url": "https://arxiv.org/abs/2605.16317", "why_it_matters": "May add technical evidence for future radar tracking: Noise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model" }, { "id": "radar_cd4b1cce0eefb05c", "categories": [ "infrastructure" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8915, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:27:13.238+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.865, "novelty": 0.709, "overall": 0.8982 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "arXiv报告其HTML论文项目（自2023年起可用）的进展，重点包括社区驱动的改进、大规模转换达到75%无错误HTML（目标90%）、用于无障碍语音输出的初版MathML 4 Intent注释，以及正在进行的LaTeXML的Rust移植以降低计算成本。", "tags": [ "arxiv", "html conversion", "mathml", "accessibility", "latex", "rust", "open access" ], "title": "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4", "url": "https://arxiv.org/abs/2605.16562", "why_it_matters": "Potentially relevant AI signal for review: Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4" }, { "id": "radar_9d267e1ad2ccf835", "categories": [ "research", "agent" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:26:28.172+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "论文介绍PQR框架，用于自动生成多样且真实的用户查询，以触发基于LLM的QA代理的失败（如不帮助、不安全行为）。该框架通过查询改写和提示改进两个模块迭代交互，生成更接近真实用户意图的故障触发查询。在电商QA代理测试中，相比此前方法，PQR能多发现23%-78%的不帮助响应，且生成的查询更多样化、更真实。", "tags": [ "academic", "arxiv", "language-models", "research", "llm-agents", "qa-agents", "query-generation" ], "title": "PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures", "url": "https://arxiv.org/abs/2605.16551", "why_it_matters": "May add technical evidence for future radar tracking: PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures" }, { "id": "radar_9a57f4127aadeb8d", "categories": [ "research", "agent" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:25:53.783+00:00", "published_at": "2026-05-19T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该研究通过对15个前沿LLM、1141个真实世界技能和超过300万次路由或执行决策的分析，发现了LLM Agent系统中技能库规模与性能之间的两种耦合缩放定律：路由定律（单步路由精度随库大小对数衰减）和执行定律（正确执行可将困难下游决策改进约4倍）。通过参数b耦合两条定律，并基于定律的优化将路由准确率从71.3%提升至91.7%，减少劫持从22.4%至4.1%，并在下游基准测试中提升通过率。结果表明，Agent性能不仅取决于模型能力，还取决于技能库的结构、粒度和暴露策略。", "tags": [ "scaling laws", "llm agents", "skill library", "routing", "execution" ], "title": "The Scaling Laws of Skills in LLM Agent Systems", "url": "https://arxiv.org/abs/2605.16508", "why_it_matters": "May add technical evidence for future radar tracking: The Scaling Laws of Skills in LLM Agent Systems" }, { "id": "radar_66fe731da0cf1113", "categories": [ "business", "product_update" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:25:23.859+00:00", "published_at": "2026-05-16T00:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.953, "freshness": 0.85, "importance": 0.898, "novelty": 0.722, "overall": 0.9115 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "summary_zh": "OpenAI与马耳他合作，向所有公民提供ChatGPT Plus和人工智能培训。", "tags": [ "partnership", "chatgpt plus", "malta", "government", "ai access", "training" ], "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership", "why_it_matters": "Potentially relevant AI signal for review: OpenAI and Malta partner to bring ChatGPT Plus to all citizens" }, { "id": "radar_45988b1714b66da1", "categories": [ "product_update", "business" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "evidence_notes": [ "source: openai news official blog, url: https://openai.com/index/dell-codex-enterprise-partnership", "published: 2026-05-18t10:00:00.000z", "raw text states: 'openai and dell partner to bring codex to hybrid and on-premise environments, helping enterprises deploy ai coding agents securely across data and workflows.'", "metadata tags include: product, research, safety", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:24:54.006+00:00", "published_at": "2026-05-18T10:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.953, "freshness": 0.85, "importance": 0.898, "novelty": 0.722, "overall": 0.9115 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "summary_zh": "OpenAI与戴尔合作，将Codex引入混合云和本地企业环境，帮助企业安全地在数据和流程中部署AI编程代理。", "tags": [ "enterprise", "partnership", "codex", "dell", "on-premise", "hybrid" ], "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership", "why_it_matters": "Potentially relevant AI signal for review: OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments" }, { "id": "radar_6e0f17d9d665540d", "categories": [ "safety", "product_update", "research" ], "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-20T02:24:22.408+00:00", "published_at": "2026-05-19T10:45:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.953, "freshness": 1, "importance": 0.8885, "novelty": 0.751, "overall": 0.914 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "summary_zh": "OpenAI 宣布推进内容溯源技术，采用 Content Credentials、SynthID 和验证工具，帮助用户识别和信任 AI 生成的媒体内容。", "tags": [ "company", "official", "product", "research", "safety", "content provenance", "synthid", "verification tool", "transparency" ], "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance", "why_it_matters": "May affect AI deployment risk, governance, or compliance planning: Advancing content provenance for a safer, more transparent AI ecosystem" }, { "id": "radar_5b1af1f62cebeac2", "categories": [ "research", "safety", "benchmark" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:45:24.199+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该研究通过控制实验，测试了三个指令微调模型（Qwen2.5-7B、Mistral-7B、Phi-3.5-mini）在五种精度（BF16到3比特）下的偏见表现，使用了12,148个BBQ基准项和5个随机种子，总计911,100次推理。结果发现，3比特量化导致6-21%的先前无偏项目出现新的刻板行为，模型选择“未知”答案的意愿下降17.4%。标准质量指标（如困惑度）在8比特时增加不到0.5%，在4比特时低于3%，但4比特时已有2.5-5.6%的项目出现新偏见，表明聚合指标忽略了公平性关键退化。", "tags": [ "quantization", "bias", "llm compression", "fairness", "bbq benchmark", "safety" ], "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208", "why_it_matters": "May affect AI deployment risk, governance, or compliance planning: Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels" }, { "id": "radar_b1d8dd4a445d5849", "categories": [ "research", "open_source", "agent" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8249, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:44:42.601+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文指出多智能体LLM系统中顺序微调存在“复合占有偏移”问题，并提出了信任域框架TeamTR，通过重采样轨迹和每个智能体的散度控制来保证改进下界，实验显示平均性能提升7.1%。", "tags": [ "multi-agent", "llm", "fine-tuning", "trust-region", "coordination", "arxiv" ], "title": "TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination", "url": "https://arxiv.org/abs/2605.15207", "why_it_matters": "May change available building blocks for teams evaluating open implementations: TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination" }, { "id": "radar_a369622fa45fb443", "categories": [ "research", "agent" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:43:55.845+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.LG", "source_tier": "T1", "status": "included", "summary_en": "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).", "summary_zh": "AgentStop是一种轻量级效率监督器，用于本地部署的LLM代理，通过预测并提前终止可能失败的任务轨迹，减少能源浪费15-20%，同时性能损失小于5%。", "tags": [ "local ai agents", "energy efficiency", "llm", "early termination", "consumer devices", "privacy-preserving" ], "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206", "why_it_matters": "May add technical evidence for future radar tracking: AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices" }, { "id": "radar_d4ce2b3966a2e234", "categories": [ "research" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:43:02.286+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该论文提出RTM方法，通过递归潜在细化替代单次潜在映射，同时提升图像生成的质量和多样性。研究表明，传统FID指标已饱和且混淆模式覆盖，而RTM结合隐式最大似然估计（IMLE）在CIFAR-10、CelebA-HQ等数据集上实现了最高的精度和召回率，同时保持竞争性FID，且不依赖特定生成器架构。", "tags": [ "academic", "arxiv", "computer-vision", "papers", "research" ], "title": "One Pass Is Not Enough: Recursive Latent Refinement for Generative Models", "url": "https://arxiv.org/abs/2605.15309", "why_it_matters": "May add technical evidence for future radar tracking: One Pass Is Not Enough: Recursive Latent Refinement for Generative Models" }, { "id": "radar_a4546429147dce6a", "categories": [ "research" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8748, "evidence_notes": [ "problem: 'visual features remain distant from the text space in the initial layers of the llm, forcing the model to waste critical depth on superficial modality alignment'", "solution: 'replaces the standard vit encoder with a small vlm as perceiver, ensuring visual features are deeply aligned with the text space'", "results: 'on the 4b parameter scale, dpa outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32b scale'", "language forgetting: 'achieves a 32.9% reduction in language capability forgetting over 3 text benchmarks'", "generality: 'these gains are consistent across different llm families including qwen3 and llama 3.2'", "practicality: 'requires only a modular replacement for the visual encoder with marginal computation overhead'", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:42:18.079+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文提出深度预对齐（DPA）架构，通过用小规模视觉语言模型（VLM）作为感知器替换标准ViT编码器，使视觉特征与目标大语言模型文本空间深度融合。在4B参数规模下，DPA在8个多模态基准上平均提升1.9个百分点，32B规模下提升3.0个百分点；同时语言能力遗忘减少32.9%。该方法在不同LLM家族（Qwen3、LLaMA 3.2）上表现一致。", "tags": [ "vision-language-models", "alignment", "arxiv", "multimodal", "deep-learning" ], "title": "Deep Pre-Alignment for VLMs", "url": "https://arxiv.org/abs/2605.15300", "why_it_matters": "May add technical evidence for future radar tracking: Deep Pre-Alignment for VLMs" }, { "id": "radar_4010b14c3118eeb7", "categories": [ "research" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:41:35.871+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CV", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "ReactiveGWM 是一种反应式游戏世界模型，通过解耦玩家控制与 NPC 行为，利用扩散骨干网络中的附加偏置和交叉注意力模块，实现玩家与 NPC 的动态交互，并支持零样本策略迁移。在《街头霸王》游戏中验证，能保持玩家可控性并实现 NPC 策略对齐。", "tags": [ "academic", "arxiv", "computer-vision", "game-world-models", "npc-intelligence", "diffusion-models", "cross-attention" ], "title": "ReactiveGWM: Steering NPC in Reactive Game World Models", "url": "https://arxiv.org/abs/2605.15256", "why_it_matters": "May add technical evidence for future radar tracking: ReactiveGWM: Steering NPC in Reactive Game World Models" }, { "id": "radar_fe8b36ce629c83c5", "categories": [ "research", "open_source" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:40:53.761+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文提出 DiscoExplorer，一个开放源码的 Web 界面，用于研究多语言话语关系。它基于 DISRPT 共享任务的数据集，覆盖 16 种语言，提供查询、搜索和可视化功能，支持对话语关系及信号装置（如连接词）的分析。", "tags": [ "discourse relations", "multilingual", "open source", "computational linguistics", "disrpt" ], "title": "DiscoExplorer: An Open Interface for the Study of Multilingual Discourse Relations", "url": "https://arxiv.org/abs/2605.15304", "why_it_matters": "May change available building blocks for teams evaluating open implementations: DiscoExplorer: An Open Interface for the Study of Multilingual Discourse Relations" }, { "id": "radar_79d12ca4aa91147e", "categories": [ "research" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:40:11.281+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该研究分析了130,486段来自106部小说、16种源语言的文学翻译（包括人类、Google Translate和TranslateGemma），发现流畅性和忠实度之间存在一致的负相关，但TranslateGemma的相关性较弱且不显著，表明在文学翻译中流畅性和忠实度存在权衡，且段落长度影响自动评估。", "tags": [ "translation", "literary translation", "fluency", "faithfulness", "llm", "machine translation", "evaluation" ], "title": "Fluency and Faithfulness in Human and Machine Literary Translation", "url": "https://arxiv.org/abs/2605.15282", "why_it_matters": "May add technical evidence for future radar tracking: Fluency and Faithfulness in Human and Machine Literary Translation" }, { "id": "radar_791ce9578c0fdb1b", "categories": [ "research" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:39:33.528+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.CL", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "本文介绍OP-Mix，一种跨语言模型训练全生命周期的数据混合算法。它通过在当前模型上训练的低秩适配器插值模拟候选数据混合，无需单独代理模型。在预训练中，OP-Mix将平均困惑度提升6.3%；在持续学习中，匹配重训练和策略蒸馏性能的同时，节省66%和95%的计算量。", "tags": [ "data mixing", "language model training", "continual learning", "op-mix", "efficient training", "arxiv", "research paper" ], "title": "Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time", "url": "https://arxiv.org/abs/2605.15220", "why_it_matters": "May add technical evidence for future radar tracking: Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time" }, { "id": "radar_18701c2f78a3692d", "categories": [ "research" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8415, "evidence_notes": [ "source: arxiv cs.ai, url: https://arxiv.org/abs/2605.15205, published_at: 2026-05-18t04:00:00.000z, collected_at: 2026-05-18t09:35:07.557z", "abstract states: 'our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic hai interactions.'", "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:38:51.675+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 0.95, "credibility": 0.899, "freshness": 1, "importance": 0.8615, "novelty": 0.751, "overall": 0.8888 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "该研究探讨了改进大语言模型心理理论能力对动态人机交互的实际影响。通过提出互动式评估范式并系统研究四种ToM增强技术，发现静态基准测试上的提升并不总能转化为动态交互中的更好表现，强调了交互评估的必要性。", "tags": [ "theory of mind", "llm", "human-ai interaction", "evaluation", "benchmark" ], "title": "Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations", "url": "https://arxiv.org/abs/2605.15205", "why_it_matters": "May add technical evidence for future radar tracking: Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations" }, { "id": "radar_63afdc4d5a09f75b", "categories": [ "research", "agent" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8249, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:37:52.203+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.899, "freshness": 1, "importance": 0.889, "novelty": 0.761, "overall": 0.9107 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "arXiv cs.AI 新论文提出 SDOF 框架，将多智能体编排建模为约束状态机，通过在线 RLHF 意图路由器（基于 GRPO 训练）和状态感知调度器实现业务阶段约束。在招聘系统（Beisen iTalent，6000+企业）评估中，7B 模型在有限状态机约束下的联合准确率达 80.9%（GPT-4o 为 48.9%），端到端任务完成率 86.5%，并完全阻止 22 种注入/非法操作。消息级拦截精确率 100%，召回率 88%。", "tags": [ "multi-agent", "orchestration", "alignment", "research" ], "title": "SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch", "url": "https://arxiv.org/abs/2605.15204", "why_it_matters": "May add technical evidence for future radar tracking: SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch" }, { "id": "radar_3f20880877462dde", "categories": [ "research", "agent" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:37:05.118+00:00", "published_at": "2026-05-18T04:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.899, "freshness": 1, "importance": 0.834, "novelty": 0.741, "overall": 0.8668 }, "source_name": "arXiv cs.AI", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "DeepSlide是一个人机协同的多智能体系统，支持从需求收集、时间预算叙述规划、基于证据的幻灯片与脚本生成、注意力增强到排练支持的完整演示准备流程。它集成了可控逻辑链规划器、轻量级内容树检索器、马尔可夫式顺序渲染和沙盒执行，并引入了双记分牌基准来分离静态工件质量和动态交付质量。在20个领域和多样化的受众中，DeepSlide在工件质量上与强基线匹配，但在叙述流、节奏精度、幻灯片-脚本协同和注意力引导等交付指标上取得更大改进。", "tags": [ "deepslide", "presentation generation", "multi-agent system", "narrative planning", "delivery optimization" ], "title": "DeepSlide: From Artifacts to Presentation Delivery", "url": "https://arxiv.org/abs/2605.15202", "why_it_matters": "May add technical evidence for future radar tracking: DeepSlide: From Artifacts to Presentation Delivery" }, { "id": "radar_4959502399d31ee9", "categories": [ "product_update" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:36:30.015+00:00", "published_at": "2026-05-15T00:00:00+00:00", "scores": { "ai_relevance": 1, "credibility": 0.953, "freshness": 0.85, "importance": 0.898, "novelty": 0.722, "overall": 0.9115 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "OpenAI published an article explaining how data science teams can use Codex to automate tasks such as creating root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.", "summary_zh": "OpenAI 发布了一篇文章，解释数据科学团队如何使用 Codex 来自动化创建根本原因简报、影响报告、KPI 备忘录、范围分析和仪表板规范等任务，这些任务基于实际工作输入。", "tags": [ "codex", "data science", "use case" ], "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex", "why_it_matters": "Potentially relevant AI signal for review: How data science teams use Codex" }, { "id": "radar_25c14f78c707362c", "categories": [ "product_update" ], "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T09:35:53.929+00:00", "published_at": "2026-05-15T00:00:00+00:00", "scores": { "ai_relevance": 0.9, "credibility": 0.9525, "freshness": 0.85, "importance": 0.8073, "novelty": 0.714, "overall": 0.8621 }, "source_name": "OpenAI News", "source_tier": "T1", "status": "included", "summary_en": "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.", "summary_zh": "OpenAI 宣布为美国 Pro 用户推出 ChatGPT 个人理财体验预览版，可安全连接金融账户，基于用户财务背景和目标提供 AI 洞察与指导。", "tags": [ "chatgpt", "personal finance", "pro users", "ai insights", "financial accounts" ], "title": "A new personal finance experience in ChatGPT", "url": "https://openai.com/index/personal-finance-chatgpt", "why_it_matters": "ChatGPT 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." }, { "id": "radar_b7af6e6b43138147", "categories": [ "media_interview" ], "collected_at": "2026-05-18T07:19:47.088+00:00", "confidence": 0.8166, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T07:27:56.411+00:00", "published_at": "2026-03-01T04:33:22+00:00", "scores": { "ai_relevance": 1, "credibility": 0.65, "freshness": 0.45, "importance": 0.793, "novelty": 0.538, "overall": 0.7433 }, "source_name": "Lex Fridman", "source_tier": "T2", "status": "included", "summary_en": "Lex Fridman Podcast #492 with Rick Beato, a music educator and multi-instrumentalist. Topics include greatest guitarists of all time, history and future of music, guitar solos, jazz, perfect pitch, learning guitar, AI in music, YouTube copyright strikes, Spotify, and more.", "summary_zh": "莱克斯·弗里德曼播客第492期，嘉宾里克·比托（音乐教育家、制作人、多乐器演奏家）。主题涵盖史上最伟大吉他手、音乐历史与未来，包括吉他独奏、爵士、完美音高、学习吉他、AI音乐、YouTube版权问题、Spotify等。", "tags": [ "music", "guitar", "podcast", "interview", "rick beato", "lex fridman", "ai in music" ], "title": "#492 – Rick Beato: Greatest Guitarists of All Time, History & Future of Music", "url": "https://lexfridman.com/rick-beato", "why_it_matters": "Potentially relevant AI signal for review: #492 – Rick Beato: Greatest Guitarists of All Time, History & Future of Music" }, { "id": "radar_035c35ec143301f5", "categories": [ "media_interview" ], "collected_at": "2026-05-18T07:19:47.088+00:00", "confidence": 0.8166, "evidence_notes": [ "formula weights: relevance 0.3, importance 0.2, credibility 0.2, novelty 0.15, freshness 0.1, source_weight 0.05" ], "language": "en", "processed_at": "2026-05-18T07:27:03.616+00:00", "published_at": "2026-03-11T20:37:33+00:00", "scores": { "ai_relevance": 1, "credibility": 0.65, "freshness": 0.45, "importance": 0.793, "novelty": 0.618, "overall": 0.7553 }, "source_name": "Lex Fridman", "source_tier": "T2", "status": "included", "summary_en": "Lex Fridman Podcast episode #493 features Jeff Kaplan, legendary Blizzard game designer of World of Warcraft and Overwatch, who is preparing to launch a new game 'The Legend of California' from his new studio Kintsugiyama, now available to wishlist on Steam with alpha in March.", "summary_zh": "Lex Fridman 播客第493期采访了暴雪传奇游戏设计师 Jeff Kaplan（《魔兽世界》《守望先锋》），他正筹备从新工作室 Kintsugiyama 发布新游戏《The Legend of California》，该游戏已在 Steam 上架并开放愿望单，预计三月进行 alpha 测试。", "tags": [ "gaming", "interview", "lex fridman", "jeff kaplan", "blizzard", "overwatch", "world of warcraft", "game design" ], "title": "#493 – Jeff Kaplan: World of Warcraft, Overwatch, Blizzard, and Future of Gaming", "url": "https://lexfridman.com/jeff-kaplan", "why_it_matters": "Potentially relevant AI signal for review: #493 – Jeff Kaplan: World of Warcraft, Overwatch, Blizzard, and Future of Gaming" } ], "reports": [ { "id": "8ed70cc5-8d42-4502-b78b-ba762e56210f", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "5 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_51c3a2bb06ffd0ba", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8671, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "An OpenAI model has disproved a central conjecture in discrete geometry", "url": "https://openai.com/index/model-disproves-discrete-geometry-conjecture" }, { "id": "radar_e4bf0f7da04d7265", "collected_at": "2026-05-22T02:44:08.561+00:00", "confidence": 0.8695, "source_name": "Anthropic News", "status": "included", "title": "Newsroom", "url": "https://www.anthropic.com/news" }, { "id": "radar_e7b25d6c012d3d58", "collected_at": "2026-05-22T02:49:38.351+00:00", "confidence": 0.2465, "published_at": "2026-05-20T14:12:54+00:00", "source_name": "Hugging Face Transformers", "status": "included", "title": "Release v5.9.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.9.0" }, { "id": "radar_cadc705c71d8e5d7", "collected_at": "2026-05-22T02:44:08.587+00:00", "confidence": 0.2298, "published_at": "2026-05-21T20:01:49+00:00", "source_name": "Anthropic Python SDK", "status": "included", "title": "v0.104.0", "url": "https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.104.0" }, { "id": "radar_5e6cf9ddee209b91", "collected_at": "2026-05-22T03:10:25.69+00:00", "confidence": 0.8488, "source_name": "Alibaba Cloud Model Studio Release Notes", "status": "included", "title": "Changelog - Alibaba Cloud", "url": "https://www.alibabacloud.com/help/en/model-studio/release-notes" }, { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_6ca572668c630cd9", "collected_at": "2026-05-22T03:13:49.337+00:00", "confidence": 0.8468, "source_name": "Kimi Platform Docs", "status": "included", "title": "Welcome to Kimi API Docs - Kimi API Platform", "url": "https://platform.kimi.ai/docs/overview" }, { "id": "radar_45988b1714b66da1", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "published_at": "2026-05-18T10:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_e3e77416ac7655d1", "collected_at": "2026-05-22T02:37:33.523+00:00", "confidence": 0.2339, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How Ramp engineers accelerate code review with Codex", "url": "https://openai.com/index/ramp" }, { "id": "radar_79d879c0eafe1136", "collected_at": "2026-05-22T02:49:38.39+00:00", "confidence": 0.2248, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation", "url": "https://arxiv.org/abs/2605.20188" } ], "confidence": 0.588, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 94 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 94 usable radar item(s). 89 included and 5 needs_review item(s). Top visible signal: \"An OpenAI model has disproved a central conjecture in discrete geometry\" from OpenAI News. Visible categories: research, model_release, product_update, open_source, safety. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-22T04:17:16.941Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-22T04:17:19.003171+00:00", "sections": [ { "bullets": [ "An OpenAI model has disproved a central conjecture in discrete geometry (2026-05-20): 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.", "Newsroom (2026-05-22): Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "Release v5.9.0 (2026-05-20): Hugging Face Transformers released v5.9.0, adding three new models: Cohere2Moe (Command A+, a Mixture-of-Experts with hybrid attention and large context), Parakeet tdt, and HRM-Text (a hierarchical recurrent autoregressive model with dual transformer stacks and PrefixLM attention).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming." ], "caveats": [], "citations": [ "radar_51c3a2bb06ffd0ba", "radar_e4bf0f7da04d7265", "radar_e7b25d6c012d3d58", "radar_cadc705c71d8e5d7", "radar_5e6cf9ddee209b91" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "An OpenAI model has disproved a central conjecture in discrete geometry (2026-05-20): 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.", "Newsroom (2026-05-22): Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "Release v5.9.0 (2026-05-20): Hugging Face Transformers released v5.9.0, adding three new models: Cohere2Moe (Command A+, a Mixture-of-Experts with hybrid attention and large context), Parakeet tdt, and HRM-Text (a hierarchical recurrent autoregressive model with dual transformer stacks and PrefixLM attention).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming." ], "caveats": [], "citations": [ "radar_51c3a2bb06ffd0ba", "radar_e4bf0f7da04d7265", "radar_e7b25d6c012d3d58", "radar_cadc705c71d8e5d7", "radar_6e0f17d9d665540d" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Newsroom (2026-05-22): Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "Changelog - Alibaba Cloud (2026-05-22): Alibaba Cloud Model Studio release notes cover Qwen model updates, OpenAI-compatible endpoint changes, and LLM capability deprecation timelines. Consult them to avoid deprecated API call failures.", "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265", "radar_cadc705c71d8e5d7", "radar_5e6cf9ddee209b91", "radar_6e0f17d9d665540d", "radar_6ca572668c630cd9" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "Welcome to Kimi API Docs - Kimi API Platform (2026-05-22): Kimi API Platform launches the K2.6 Open Platform, providing a trillion-parameter K2.5 large language model API, supporting 256K long context and Tool Calling, with professional code generation, intelligent dialogue, and visual reasoning capabilities to help developers build AI applications.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_6ca572668c630cd9", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_5b1af1f62cebeac2" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers released v5.9.0, adding three new models: Cohere2Moe (Command A+, a Mixture-of-Experts with hybrid attention and large context), Parakeet tdt, and HRM-Text (a hierarchical recurrent autoregressive model with dual transformer stacks and PrefixLM attention).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "How Ramp engineers accelerate code review with Codex (2026-05-20): OpenAI News blog describes how Ramp engineers use Codex with GPT-5.5 to accelerate code review, reducing feedback time from hours to minutes.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_cadc705c71d8e5d7", "radar_e3e77416ac7655d1", "radar_66fe731da0cf1113", "radar_79d879c0eafe1136" ], "summary": "6 radar item(s) matched this section.", "title": "Weak signals / needs_review" } ], "source_item_count": 13, "status": "needs_review", "summary": "Deterministic weekly preview from 94 usable radar item(s). 89 included and 5 needs_review item(s). Top visible signal: \"An OpenAI model has disproved a central conjecture in discrete geometry\" from OpenAI News. Visible categories: research, model_release, product_update, open_source, safety.", "time_window": { "end": "2026-05-22T04:13:38.288+00:00", "start": "2026-05-15T04:13:38.288+00:00" }, "title": "Weekly AI Radar preview - ending May 22, 2026" }, { "id": "ec51a8ff-f700-47a9-9f1c-4b6bd647f9e5", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "5 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_e4bf0f7da04d7265", "collected_at": "2026-05-22T02:44:08.561+00:00", "confidence": 0.8695, "source_name": "Anthropic News", "status": "included", "title": "Newsroom", "url": "https://www.anthropic.com/news" }, { "id": "radar_cadc705c71d8e5d7", "collected_at": "2026-05-22T02:44:08.587+00:00", "confidence": 0.2298, "published_at": "2026-05-21T20:01:49+00:00", "source_name": "Anthropic Python SDK", "status": "included", "title": "v0.104.0", "url": "https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.104.0" }, { "id": "radar_5e6cf9ddee209b91", "collected_at": "2026-05-22T03:10:25.69+00:00", "confidence": 0.8488, "source_name": "Alibaba Cloud Model Studio Release Notes", "status": "included", "title": "Changelog - Alibaba Cloud", "url": "https://www.alibabacloud.com/help/en/model-studio/release-notes" }, { "id": "radar_6ca572668c630cd9", "collected_at": "2026-05-22T03:13:49.337+00:00", "confidence": 0.8468, "source_name": "Kimi Platform Docs", "status": "included", "title": "Welcome to Kimi API Docs - Kimi API Platform", "url": "https://platform.kimi.ai/docs/overview" }, { "id": "radar_57c2241ccf3b26cc", "collected_at": "2026-05-22T03:03:02.499+00:00", "confidence": 0.8508, "source_name": "Google AI for Developers", "status": "included", "title": "Gemini generateContent API  |  Google AI for Developers", "url": "https://ai.google.dev/gemini-api/docs" }, { "id": "radar_dffd1aeebae810c0", "collected_at": "2026-05-22T02:37:33.552+00:00", "confidence": 0.8641, "published_at": "2026-05-21T16:21:17+00:00", "source_name": "OpenAI Cookbook", "status": "included", "title": "openai/openai-cookbook repository metadata", "url": "https://github.com/openai/openai-cookbook" }, { "id": "radar_f84d222af7af5acf", "collected_at": "2026-05-22T02:56:56.872+00:00", "confidence": 0.8631, "published_at": "2026-05-21T21:23:37+00:00", "source_name": "OpenAI Python SDK", "status": "included", "title": "openai/openai-python repository metadata", "url": "https://github.com/openai/openai-python" }, { "id": "radar_52aed8fac227edc4", "collected_at": "2026-05-22T03:10:25.741+00:00", "confidence": 0.7964, "source_name": "Yarin Gal", "status": "included", "title": "Yarin Gal - Home Page | Oxford Machine Learning", "url": "http://yarin.co/" }, { "id": "radar_f13e235515c7869a", "collected_at": "2026-05-22T03:03:02.482+00:00", "confidence": 0.8463, "source_name": "Latent Space", "status": "included", "title": "Latent.Space | Substack", "url": "https://www.latent.space/" }, { "id": "radar_43125b0e2fc1ee24", "collected_at": "2026-05-22T02:49:38.361+00:00", "confidence": 0.7297, "source_name": "Every", "status": "included", "title": "Every", "url": "https://every.to/" }, { "id": "radar_057cc49ab17c43e0", "collected_at": "2026-05-22T02:56:56.887+00:00", "confidence": 0.763, "source_name": "Fabricated Knowledge", "status": "included", "title": "Fabricated Knowledge | Doug OLaughlin | Substack", "url": "https://www.fabricatedknowledge.com/" }, { "id": "radar_59875977de6cf2b4", "collected_at": "2026-05-22T02:59:32.444+00:00", "confidence": 0.2185, "source_name": "Google Gemini Blog", "status": "included", "title": "Gemini", "url": "https://blog.google/products-and-platforms/products/gemini" } ], "confidence": 0.638, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 41 条可用雷达条目和 12 条引用生成。 Deterministic daily preview from 41 usable radar item(s). 36 included and 5 needs_review item(s). Top visible signal: \"Newsroom\" from Anthropic News. Visible categories: model_release, product_update, research, open_source, infrastructure. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-22T04:17:17.018Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-22T04:17:18.992613+00:00", "sections": [ { "bullets": [ "Newsroom (2026-05-22): Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "Changelog - Alibaba Cloud (2026-05-22): Alibaba Cloud Model Studio release notes cover Qwen model updates, OpenAI-compatible endpoint changes, and LLM capability deprecation timelines. Consult them to avoid deprecated API call failures.", "Welcome to Kimi API Docs - Kimi API Platform (2026-05-22): Kimi API Platform launches the K2.6 Open Platform, providing a trillion-parameter K2.5 large language model API, supporting 256K long context and Tool Calling, with professional code generation, intelligent dialogue, and visual reasoning capabilities to help developers build AI applications." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265", "radar_cadc705c71d8e5d7", "radar_5e6cf9ddee209b91", "radar_6ca572668c630cd9", "radar_57c2241ccf3b26cc" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Newsroom (2026-05-22): Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "openai/openai-cookbook repository metadata (2026-05-21): The OpenAI Cookbook is a GitHub repository that provides examples and guides for using the OpenAI API. As of May 21, 2026, it has 73,681 stars, 12,461 forks, and 185 open issues.", "openai/openai-python repository metadata (2026-05-21): OpenAI Python SDK official repository updated metadata on May 21, 2026, with 30810 stars, 4796 forks, and 537 open issues." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265", "radar_cadc705c71d8e5d7", "radar_dffd1aeebae810c0", "radar_f84d222af7af5acf", "radar_52aed8fac227edc4" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Newsroom (2026-05-22): Anthropic's newsroom page, collected on May 22, 2026, features recent announcements including the launch of Claude Opus 4.7 (April 16, 2026), Claude Design (April 17, 2026), Project Glasswing (April 7, 2026), and insights from 81,000 user interviews (March 18, 2026).", "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "Changelog - Alibaba Cloud (2026-05-22): Alibaba Cloud Model Studio release notes cover Qwen model updates, OpenAI-compatible endpoint changes, and LLM capability deprecation timelines. Consult them to avoid deprecated API call failures.", "Welcome to Kimi API Docs - Kimi API Platform (2026-05-22): Kimi API Platform launches the K2.6 Open Platform, providing a trillion-parameter K2.5 large language model API, supporting 256K long context and Tool Calling, with professional code generation, intelligent dialogue, and visual reasoning capabilities to help developers build AI applications." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265", "radar_cadc705c71d8e5d7", "radar_5e6cf9ddee209b91", "radar_6ca572668c630cd9", "radar_57c2241ccf3b26cc" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Welcome to Kimi API Docs - Kimi API Platform (2026-05-22): Kimi API Platform launches the K2.6 Open Platform, providing a trillion-parameter K2.5 large language model API, supporting 256K long context and Tool Calling, with professional code generation, intelligent dialogue, and visual reasoning capabilities to help developers build AI applications.", "openai/openai-python repository metadata (2026-05-21): OpenAI Python SDK official repository updated metadata on May 21, 2026, with 30810 stars, 4796 forks, and 537 open issues.", "Latent.Space | Substack (2026-05-22): Latent Space is an AI Engineer newsletter and top technical AI podcast covering how leading labs build Agents, Models, Infra, & AI for Science. It features highlights from Greg Brockman, Andrej Karpathy, and others. The Substack publication has hundreds of thousands of subscribers.", "Every (2026-05-22): Every is a subscription service focused on AI, offering ideas, apps, and training from practitioners, including a newsletter, podcast, events, and more." ], "caveats": [], "citations": [ "radar_6ca572668c630cd9", "radar_f84d222af7af5acf", "radar_f13e235515c7869a", "radar_43125b0e2fc1ee24", "radar_057cc49ab17c43e0" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "v0.104.0 (2026-05-21): Anthropic Python SDK v0.104.0 released, adding support for thinking-token-count beta for estimated tokens in thinking block deltas when streaming.", "Gemini (2026-05-22): The product page on the Google Gemini Blog serves as a central hub for news and updates about Gemini AI, including official information on writing, planning, learning, and other features.", "SemiAnalysis (2026-05-22): SemiAnalysis is a tech media outlet bridging semiconductors and business, offering in-depth research and models on accelerators, HBM, AI cloud TCO, networking, datacenter, energy, and more.", "Tailwinds | Apoorv Agrawal | Substack (2026-05-22): This is the homepage of \"Tailwinds,\" a Substack publication by venture investor Apoorv Agrawal, focusing on the business of technology and the tailwinds that power it." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_cadc705c71d8e5d7", "radar_59875977de6cf2b4" ], "summary": "6 radar item(s) matched this section. 1 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 16, "status": "needs_review", "summary": "Deterministic daily preview from 41 usable radar item(s). 36 included and 5 needs_review item(s). Top visible signal: \"Newsroom\" from Anthropic News. Visible categories: model_release, product_update, research, open_source, infrastructure.", "time_window": { "end": "2026-05-22T04:13:38.288+00:00", "start": "2026-05-21T04:13:38.288+00:00" }, "title": "Daily AI Radar preview - May 22, 2026" }, { "id": "2a28689e-9bd8-498d-9b51-c4e96c3dfea0", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_e7b25d6c012d3d58", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8631, "published_at": "2026-05-20T14:12:54+00:00", "source_name": "Hugging Face Transformers", "status": "included", "title": "Release v5.9.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.9.0" }, { "id": "radar_51c3a2bb06ffd0ba", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8671, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "An OpenAI model has disproved a central conjecture in discrete geometry", "url": "https://openai.com/index/model-disproves-discrete-geometry-conjecture" }, { "id": "radar_dffd1aeebae810c0", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8641, "published_at": "2026-05-20T21:53:11+00:00", "source_name": "OpenAI Cookbook", "status": "included", "title": "openai/openai-cookbook repository metadata", "url": "https://github.com/openai/openai-cookbook" }, { "id": "radar_2a46a79464ecf9bb", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.6862, "source_name": "Anthropic Research", "status": "included", "title": "Research", "url": "https://www.anthropic.com/research" }, { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_f4ae49772d646fad", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8338, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "The next phase of OpenAI’s Education for Countries", "url": "https://openai.com/index/the-next-phase-of-education-for-countries" }, { "id": "radar_45988b1714b66da1", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "published_at": "2026-05-18T10:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_4959502399d31ee9", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_cd4b1cce0eefb05c", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8915, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4", "url": "https://arxiv.org/abs/2605.16562" }, { "id": "radar_a369622fa45fb443", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206" } ], "confidence": 0.605, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 66 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 66 usable radar item(s). 64 included and 2 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, safety, product_update. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-21T10:09:14.737Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-21T10:09:16.780297+00:00", "sections": [ { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three new models: Cohere's MoE model Command A+ (Cohere2Moe), Parakeet tdt, and hierarchical reasoning model HRM-Text.", "An OpenAI model has disproved a central conjecture in discrete geometry (2026-05-20): 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.", "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The next phase of OpenAI’s Education for Countries (2026-05-20): OpenAI advances the next phase of Education for Countries, expanding AI adoption in schools with new partnerships, teacher training, and tools to improve global learning outcomes." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_51c3a2bb06ffd0ba", "radar_6e0f17d9d665540d", "radar_f4ae49772d646fad", "radar_45988b1714b66da1" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three new models: Cohere's MoE model Command A+ (Cohere2Moe), Parakeet tdt, and hierarchical reasoning model HRM-Text.", "An OpenAI model has disproved a central conjecture in discrete geometry (2026-05-20): 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.", "openai/openai-cookbook repository metadata (2026-05-20): The OpenAI Cookbook is an open-source GitHub repository providing examples and guides for using the OpenAI API. As of May 20, 2026, it has 73,671 stars, 12,459 forks, and 184 open issues.", "Research (2026-05-21): Anthropic's research page provides an overview of their AI safety and research efforts, with links to teams (Alignment, Interpretability, etc.) and recent publications (e.g., Natural Language Autoencoders, Teaching Claude why). No specific research results are detailed in this ingestion." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_51c3a2bb06ffd0ba", "radar_dffd1aeebae810c0", "radar_2a46a79464ecf9bb", "radar_6e0f17d9d665540d" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The next phase of OpenAI’s Education for Countries (2026-05-20): OpenAI advances the next phase of Education for Countries, expanding AI adoption in schools with new partnerships, teacher training, and tools to improve global learning outcomes.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_f4ae49772d646fad", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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" ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_5b1af1f62cebeac2", "radar_cd4b1cce0eefb05c" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "ReactiveGWM: Steering NPC in Reactive Game World Models (2026-05-18): 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." ], "caveats": [], "citations": [ "radar_66fe731da0cf1113", "radar_a369622fa45fb443", "radar_5b1af1f62cebeac2" ], "summary": "6 radar item(s) matched this section.", "title": "Weak signals / needs_review" } ], "source_item_count": 15, "status": "needs_review", "summary": "Deterministic weekly preview from 66 usable radar item(s). 64 included and 2 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, safety, product_update.", "time_window": { "end": "2026-05-21T10:05:00.655+00:00", "start": "2026-05-14T10:05:00.655+00:00" }, "title": "Weekly AI Radar preview - ending May 21, 2026" }, { "id": "3c439b35-ba3a-4873-aceb-f2b675ed854c", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_e7b25d6c012d3d58", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8631, "published_at": "2026-05-20T14:12:54+00:00", "source_name": "Hugging Face Transformers", "status": "included", "title": "Release v5.9.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.9.0" }, { "id": "radar_dffd1aeebae810c0", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8641, "published_at": "2026-05-20T21:53:11+00:00", "source_name": "OpenAI Cookbook", "status": "included", "title": "openai/openai-cookbook repository metadata", "url": "https://github.com/openai/openai-cookbook" }, { "id": "radar_2a46a79464ecf9bb", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.6862, "source_name": "Anthropic Research", "status": "included", "title": "Research", "url": "https://www.anthropic.com/research" }, { "id": "radar_d9de0c996dcddce8", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token", "url": "https://arxiv.org/abs/2605.20192" }, { "id": "radar_1eea04c5503e2f54", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Why Latent Actions Fail, and How to Prevent It", "url": "https://arxiv.org/abs/2605.20223" }, { "id": "radar_e4bf0f7da04d7265", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8528, "source_name": "Anthropic News", "status": "included", "title": "Newsroom", "url": "https://www.anthropic.com/news" }, { "id": "radar_bbdaaeb7903db9de", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.8528, "source_name": "Google DeepMind Blog", "status": "included", "title": "News — Google DeepMind", "url": "https://deepmind.google/blog" }, { "id": "radar_43125b0e2fc1ee24", "collected_at": "2026-05-21T09:45:24.786+00:00", "confidence": 0.5963, "source_name": "Every", "status": "included", "title": "Every", "url": "https://every.to/" } ], "confidence": 0.683, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 18 条可用雷达条目和 8 条引用生成。 Deterministic daily preview from 18 usable radar item(s). 18 included and 0 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, product_update, other. Model / product / company updates: 2 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 1 radar item(s) matched this section.", "generated_at": "2026-05-21T10:09:14.738Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-21T10:09:16.76191+00:00", "sections": [ { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three new models: Cohere's MoE model Command A+ (Cohere2Moe), Parakeet tdt, and hierarchical reasoning model HRM-Text.", "Newsroom (2026-05-21): Anthropic's Newsroom page aggregating recent announcements including Claude Opus 4.7 model, Claude Design, Project Glasswing, and a user survey on AI expectations." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_e4bf0f7da04d7265" ], "summary": "2 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three new models: Cohere's MoE model Command A+ (Cohere2Moe), Parakeet tdt, and hierarchical reasoning model HRM-Text.", "openai/openai-cookbook repository metadata (2026-05-20): The OpenAI Cookbook is an open-source GitHub repository providing examples and guides for using the OpenAI API. As of May 20, 2026, it has 73,671 stars, 12,459 forks, and 184 open issues.", "Research (2026-05-21): Anthropic's research page provides an overview of their AI safety and research efforts, with links to teams (Alignment, Interpretability, etc.) and recent publications (e.g., Natural Language Autoencoders, Teaching Claude why). No specific research results are detailed in this ingestion.", "Leveraging Large Language Models for Sentiment Analysis: Multi-Modal Analysis of Decentraland's MANA Token (2026-05-21): This study uses a BERT-based large language model to analyze sentiment from Decentraland's Discord community and integrates it with multi-modal financial data (historical prices, trading volume, market capitalization) in LSTM architectures for MANA token return prediction. Results show predominantly neutral sentiment with a positive skew, and the multi-modal model significantly outperforms the price-only baseline, demonstrating the predictive value of community signals for virtual economy forecasting." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_dffd1aeebae810c0", "radar_2a46a79464ecf9bb", "radar_d9de0c996dcddce8", "radar_1eea04c5503e2f54" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Newsroom (2026-05-21): Anthropic's Newsroom page aggregating recent announcements including Claude Opus 4.7 model, Claude Design, Project Glasswing, and a user survey on AI expectations." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265" ], "summary": "1 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "News — Google DeepMind (2026-05-21): This entry is a summary of the Google DeepMind official blog homepage, featuring navigation links such as Gemini for Science, Experimental Tools, etc., but no specific article content.", "Every (2026-05-21): A subscription newsletter focused on AI, offering ideas, apps, and training from practitioners who build with AI daily." ], "caveats": [], "citations": [ "radar_bbdaaeb7903db9de", "radar_43125b0e2fc1ee24" ], "summary": "2 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "No usable radar evidence currently supports this section." ], "caveats": [], "citations": [], "summary": "No retrieved radar items in this window support this section.", "title": "Weak signals / needs_review" } ], "source_item_count": 8, "status": "needs_review", "summary": "Deterministic daily preview from 18 usable radar item(s). 18 included and 0 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, product_update, other.", "time_window": { "end": "2026-05-21T10:05:00.655+00:00", "start": "2026-05-20T10:05:00.655+00:00" }, "title": "Daily AI Radar preview - May 21, 2026" }, { "id": "11245d8b-5abb-4943-8a88-d4c0fee81c94", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "3 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_e7b25d6c012d3d58", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8631, "published_at": "2026-05-20T14:12:54+00:00", "source_name": "Hugging Face Transformers", "status": "included", "title": "Release v5.9.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.9.0" }, { "id": "radar_dffd1aeebae810c0", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8641, "published_at": "2026-05-20T21:53:11+00:00", "source_name": "OpenAI Cookbook", "status": "included", "title": "openai/openai-cookbook repository metadata", "url": "https://github.com/openai/openai-cookbook" }, { "id": "radar_2a46a79464ecf9bb", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8528, "source_name": "Anthropic Research", "status": "included", "title": "Research", "url": "https://www.anthropic.com/research" }, { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_45988b1714b66da1", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "published_at": "2026-05-18T10:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_4959502399d31ee9", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex" }, { "id": "radar_26255f33ab09d8c3", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Evaluating the Utility of Personal Health Records in Personalized Health AI", "url": "https://arxiv.org/abs/2605.18937" }, { "id": "radar_9a57f4127aadeb8d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "The Scaling Laws of Skills in LLM Agent Systems", "url": "https://arxiv.org/abs/2605.16508" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_cd4b1cce0eefb05c", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8915, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4", "url": "https://arxiv.org/abs/2605.16562" }, { "id": "radar_a369622fa45fb443", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206" } ], "confidence": 0.605, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 66 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 66 usable radar item(s). 63 included and 3 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, safety, product_update. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-21T08:15:51.351Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-21T08:15:53.371015+00:00", "sections": [ { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three models: Cohere2Moe (Command A+, a Mixture-of-Experts model with hybrid attention and large context window), Parakeet tdt, and HRM-Text (hierarchical recurrent autoregressive model with dual transformer stacks for slow planning and fast computation).", "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three models: Cohere2Moe (Command A+, a Mixture-of-Experts model with hybrid attention and large context window), Parakeet tdt, and HRM-Text (hierarchical recurrent autoregressive model with dual transformer stacks for slow planning and fast computation).", "openai/openai-cookbook repository metadata (2026-05-20): The OpenAI Cookbook repository on GitHub provides examples and guides for using the OpenAI API. As of collection time, it has 73,669 stars and 12,457 forks.", "Research (2026-05-21): Anthropic is an AI safety and research company working to build reliable, interpretable, and steerable AI systems.", "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_dffd1aeebae810c0", "radar_2a46a79464ecf9bb", "radar_6e0f17d9d665540d", "radar_26255f33ab09d8c3" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "How data science teams use Codex (2026-05-15): OpenAI published an article explaining how data science teams can use Codex to automate tasks such as creating root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9", "radar_9a57f4127aadeb8d" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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" ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_5b1af1f62cebeac2", "radar_cd4b1cce0eefb05c" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "ReactiveGWM: Steering NPC in Reactive Game World Models (2026-05-18): 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." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_66fe731da0cf1113", "radar_a369622fa45fb443", "radar_5b1af1f62cebeac2" ], "summary": "6 radar item(s) matched this section. 1 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 15, "status": "needs_review", "summary": "Deterministic weekly preview from 66 usable radar item(s). 63 included and 3 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, safety, product_update.", "time_window": { "end": "2026-05-21T08:14:09.85+00:00", "start": "2026-05-14T08:14:09.85+00:00" }, "title": "Weekly AI Radar preview - ending May 21, 2026" }, { "id": "a9552a33-552a-4fc1-b6f8-12514de9efd5", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "3 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_e7b25d6c012d3d58", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8631, "published_at": "2026-05-20T14:12:54+00:00", "source_name": "Hugging Face Transformers", "status": "included", "title": "Release v5.9.0", "url": "https://github.com/huggingface/transformers/releases/tag/v5.9.0" }, { "id": "radar_dffd1aeebae810c0", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8641, "published_at": "2026-05-20T21:53:11+00:00", "source_name": "OpenAI Cookbook", "status": "included", "title": "openai/openai-cookbook repository metadata", "url": "https://github.com/openai/openai-cookbook" }, { "id": "radar_2a46a79464ecf9bb", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8528, "source_name": "Anthropic Research", "status": "included", "title": "Research", "url": "https://www.anthropic.com/research" }, { "id": "radar_26255f33ab09d8c3", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Evaluating the Utility of Personal Health Records in Personalized Health AI", "url": "https://arxiv.org/abs/2605.18937" }, { "id": "radar_79d879c0eafe1136", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation", "url": "https://arxiv.org/abs/2605.20188" }, { "id": "radar_2a64bcb1fa50406f", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8249, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance", "url": "https://arxiv.org/abs/2605.18801" }, { "id": "radar_bbdaaeb7903db9de", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8528, "source_name": "Google DeepMind Blog", "status": "included", "title": "News — Google DeepMind", "url": "https://deepmind.google/blog" }, { "id": "radar_e4bf0f7da04d7265", "collected_at": "2026-05-21T07:52:35.749+00:00", "confidence": 0.8528, "source_name": "Anthropic News", "status": "included", "title": "Newsroom", "url": "https://www.anthropic.com/news" }, { "id": "radar_43125b0e2fc1ee24", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.3963, "source_name": "Every", "status": "needs_review", "title": "Every", "url": "https://every.to/" } ], "confidence": 0.7, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 18 条可用雷达条目和 9 条引用生成。 Deterministic daily preview from 18 usable radar item(s). 17 included and 1 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, opinion, other. Model / product / company updates: 1 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Business / ecosystem: 4 radar item(s) matched this section. 1 still need review.", "generated_at": "2026-05-21T08:15:38.921Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-21T08:15:40.888374+00:00", "sections": [ { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three models: Cohere2Moe (Command A+, a Mixture-of-Experts model with hybrid attention and large context window), Parakeet tdt, and HRM-Text (hierarchical recurrent autoregressive model with dual transformer stacks for slow planning and fast computation)." ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58" ], "summary": "1 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Release v5.9.0 (2026-05-20): Hugging Face Transformers releases v5.9.0, adding three models: Cohere2Moe (Command A+, a Mixture-of-Experts model with hybrid attention and large context window), Parakeet tdt, and HRM-Text (hierarchical recurrent autoregressive model with dual transformer stacks for slow planning and fast computation).", "openai/openai-cookbook repository metadata (2026-05-20): The OpenAI Cookbook repository on GitHub provides examples and guides for using the OpenAI API. As of collection time, it has 73,669 stars and 12,457 forks.", "Research (2026-05-21): Anthropic is an AI safety and research company working to build reliable, interpretable, and steerable AI systems.", "Evaluating the Utility of Personal Health Records in Personalized Health AI (2026-05-21): This study evaluates the utility of Personal Health Records (PHRs) in personalized health AI by using Gemini 3.0 Flash LLM to answer 2,257 user queries (from three distributions) paired with 1,945 de-identified PHRs under three conditions: no PHR context, basic summary (demographics, conditions, medications), and full clinical notes. Evaluation used the SHARP framework and a novel PHR error mode framework, with auto-raters and clinician raters (n=95). Results show significant improvements in helpfulness for all query types with PHR context (p<0.001), and potential gains in safety, accuracy, relevance, a" ], "caveats": [], "citations": [ "radar_e7b25d6c012d3d58", "radar_dffd1aeebae810c0", "radar_2a46a79464ecf9bb", "radar_26255f33ab09d8c3", "radar_79d879c0eafe1136" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "No usable radar evidence currently supports this section." ], "caveats": [], "citations": [], "summary": "No retrieved radar items in this window support this section.", "title": "Agents / products" }, { "bullets": [ "Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance (2026-05-21): A position paper proposing the development of 'data probes'—synthetic sequences from random processes—to systematically study how data characteristics influence LLM performance, generalization, and robustness, beyond current empirical heuristics.", "News — Google DeepMind (2026-05-21): This is the homepage of the Google DeepMind official blog, listing links such as Gemini for Science, Experimental Tools, Build with Gemini, etc., but no specific article content.", "Newsroom (2026-05-21): Anthropic official newsroom page listing recent announcements including Claude Opus 4.7, Claude Design, Project Glasswing, and user AI needs survey.", "Every (2026-05-21): Every is a subscription-based newsletter platform focused on AI, offering ideas, apps, and training from AI practitioners. This page is its homepage, with navigation links for sign-in, subscribe, columnists, podcast, etc." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_2a64bcb1fa50406f", "radar_bbdaaeb7903db9de", "radar_e4bf0f7da04d7265", "radar_43125b0e2fc1ee24" ], "summary": "4 radar item(s) matched this section. 1 still need review.", "title": "Business / ecosystem" }, { "bullets": [ "Every (2026-05-21): Every is a subscription-based newsletter platform focused on AI, offering ideas, apps, and training from AI practitioners. This page is its homepage, with navigation links for sign-in, subscribe, columnists, podcast, etc." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_43125b0e2fc1ee24" ], "summary": "1 radar item(s) matched this section. 1 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 9, "status": "needs_review", "summary": "Deterministic daily preview from 18 usable radar item(s). 17 included and 1 needs_review item(s). Top visible signal: \"Release v5.9.0\" from Hugging Face Transformers. Visible categories: model_release, open_source, research, opinion, other.", "time_window": { "end": "2026-05-21T08:14:09.85+00:00", "start": "2026-05-20T08:14:09.85+00:00" }, "title": "Daily AI Radar preview - May 21, 2026" }, { "id": "61216c20-620f-4064-8296-5cc9162fdd10", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "3 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_8d5fd97b76b9d62a", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production", "url": "https://arxiv.org/abs/2605.18818" }, { "id": "radar_1eea04c5503e2f54", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Why Latent Actions Fail, and How to Prevent It", "url": "https://arxiv.org/abs/2605.20223" }, { "id": "radar_2a46a79464ecf9bb", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8195, "source_name": "Anthropic Research", "status": "included", "title": "Research", "url": "https://www.anthropic.com/research" }, { "id": "radar_2a64bcb1fa50406f", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8415, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance", "url": "https://arxiv.org/abs/2605.18801" }, { "id": "radar_26255f33ab09d8c3", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Evaluating the Utility of Personal Health Records in Personalized Health AI", "url": "https://arxiv.org/abs/2605.18937" }, { "id": "radar_e4bf0f7da04d7265", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8695, "source_name": "Anthropic News", "status": "included", "title": "Newsroom", "url": "https://www.anthropic.com/news" }, { "id": "radar_54ff21cc34521714", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Shiny Stories, Hidden Struggles: Investigating the Representation of Disability Through the Lens of LLMs", "url": "https://arxiv.org/abs/2605.20191" }, { "id": "radar_43125b0e2fc1ee24", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.3963, "source_name": "Every", "status": "needs_review", "title": "Every", "url": "https://every.to/" } ], "confidence": 0.768, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 11 条可用雷达条目和 8 条引用生成。 Deterministic daily preview from 11 usable radar item(s). 10 included and 1 needs_review item(s). Top visible signal: \"Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production\" from arXiv cs.AI. Visible categories: research, infrastructure, safety, product_update, other. Model / product / company updates: 1 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 1 radar item(s) matched this section.", "generated_at": "2026-05-21T04:24:32.511Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-21T04:24:35.214014+00:00", "sections": [ { "bullets": [ "Newsroom (2026-05-21): Anthropic's Newsroom page listing recent announcements and product updates, including Claude Opus 4.7, Claude Design, Project Glasswing, and a user survey report." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265" ], "summary": "1 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production (2026-05-21): This paper presents a microservice architecture for document AI that encapsulates pipelines for classification, OCR, and LLM-based structured field extraction, capable of processing thousands of multi-page documents per hour. Key findings include: OCR dominates end-to-end latency over LLM parsing, and system concurrency is limited by shared GPU inference capacity rather than worker count. The goal is to provide practitioners with concrete architectural patterns for production-scale document understanding systems.", "Why Latent Actions Fail, and How to Prevent It (2026-05-21): This paper investigates why latent action models (LAMs) fail due to exogenous state (e.g., background clutter) when learning action representations from unlabeled videos. Extending a linear LAM framework, the authors find that minimizing the standard reconstruction objective causes latent actions to encode exogenous information from future observations, and that focusing on endogenous components in the representation space mitigates noise. Furthermore, auxiliary objectives like action-supervision provably encourage consistency of latent actions across exogenous states. Experiments on linear and nonlinear LAMs validate the theoretic", "Research (2026-05-21): Anthropic Research homepage overview, listing research areas (alignment, interpretability, economic research, societal impacts) and recent blog posts (e.g., Natural Language Autoencoders, Teaching Claude why), emphasizing building reliable, interpretable, and steerable AI systems.", "Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance (2026-05-21): This position paper advocates for developing 'data probes'—synthetic sequences generated from well-defined random processes—to systematically study how data characteristics influence LLM performance, generalization, and robustness across training, tuning, alignment, and in-context learning, moving beyond compute-intensive empirical heuristics." ], "caveats": [], "citations": [ "radar_8d5fd97b76b9d62a", "radar_1eea04c5503e2f54", "radar_2a46a79464ecf9bb", "radar_2a64bcb1fa50406f", "radar_26255f33ab09d8c3" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Newsroom (2026-05-21): Anthropic's Newsroom page listing recent announcements and product updates, including Claude Opus 4.7, Claude Design, Project Glasswing, and a user survey report." ], "caveats": [], "citations": [ "radar_e4bf0f7da04d7265" ], "summary": "1 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production (2026-05-21): This paper presents a microservice architecture for document AI that encapsulates pipelines for classification, OCR, and LLM-based structured field extraction, capable of processing thousands of multi-page documents per hour. Key findings include: OCR dominates end-to-end latency over LLM parsing, and system concurrency is limited by shared GPU inference capacity rather than worker count. The goal is to provide practitioners with concrete architectural patterns for production-scale document understanding systems.", "Shiny Stories, Hidden Struggles: Investigating the Representation of Disability Through the Lens of LLMs (2026-05-21): This study investigates how LLMs represent disability by simulating social media posts from the perspective of individuals with disabilities and comparing them to real posts. Findings reveal that LLMs tend to idealize disability experiences (overly positive stereotypes) and exhibit a negative bias in topic association (e.g., disproportionately linking career and entertainment to non-disabled individuals), reinforcing exclusionary narratives and misrepresenting the actual challenges faced by the community.", "Every (2026-05-21): Every is a subscription-based newsletter platform focused on AI, offering ideas, apps, and training from AI practitioners. This page is its homepage, with navigation links for sign-in, subscribe, columnists, podcast, etc." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_8d5fd97b76b9d62a", "radar_54ff21cc34521714", "radar_43125b0e2fc1ee24" ], "summary": "3 radar item(s) matched this section. 1 still need review.", "title": "Business / ecosystem" }, { "bullets": [ "Every (2026-05-21): Every is a subscription-based newsletter platform focused on AI, offering ideas, apps, and training from AI practitioners. This page is its homepage, with navigation links for sign-in, subscribe, columnists, podcast, etc." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_43125b0e2fc1ee24" ], "summary": "1 radar item(s) matched this section. 1 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 8, "status": "needs_review", "summary": "Deterministic daily preview from 11 usable radar item(s). 10 included and 1 needs_review item(s). Top visible signal: \"Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production\" from arXiv cs.AI. Visible categories: research, infrastructure, safety, product_update, other.", "time_window": { "end": "2026-05-21T04:22:24.667+00:00", "start": "2026-05-20T04:22:24.667+00:00" }, "title": "Daily AI Radar preview - May 21, 2026" }, { "id": "d577110b-27a3-4765-898b-8df9c4a697aa", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "3 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_f4ae49772d646fad", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8671, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "The next phase of OpenAI’s Education for Countries", "url": "https://openai.com/index/the-next-phase-of-education-for-countries" }, { "id": "radar_45988b1714b66da1", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "published_at": "2026-05-18T10:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_4959502399d31ee9", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex" }, { "id": "radar_8d5fd97b76b9d62a", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production", "url": "https://arxiv.org/abs/2605.18818" }, { "id": "radar_1eea04c5503e2f54", "collected_at": "2026-05-21T04:10:32.377+00:00", "confidence": 0.8581, "published_at": "2026-05-21T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Why Latent Actions Fail, and How to Prevent It", "url": "https://arxiv.org/abs/2605.20223" }, { "id": "radar_1488431b28a1ca16", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance", "url": "https://arxiv.org/abs/2605.18793" }, { "id": "radar_0bed1c17d6ef1231", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Harnessing Self-Supervised Features for Art Classification", "url": "https://arxiv.org/abs/2605.18974" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_a369622fa45fb443", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206" }, { "id": "radar_4010b14c3118eeb7", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "ReactiveGWM: Steering NPC in Reactive Game World Models", "url": "https://arxiv.org/abs/2605.15256" } ], "confidence": 0.614, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 56 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 56 usable radar item(s). 53 included and 3 needs_review item(s). Top visible signal: \"Advancing content provenance for a safer, more transparent AI ecosystem\" from OpenAI News. Visible categories: safety, product_update, research, business, infrastructure. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-21T04:24:32.451Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-21T04:24:35.189038+00:00", "sections": [ { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The next phase of OpenAI’s Education for Countries (2026-05-20): OpenAI advances Education for Countries, expanding AI adoption in schools with new partnerships, teacher training, and tools to improve global learning outcomes.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_f4ae49772d646fad", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production (2026-05-21): This paper presents a microservice architecture for document AI that encapsulates pipelines for classification, OCR, and LLM-based structured field extraction, capable of processing thousands of multi-page documents per hour. Key findings include: OCR dominates end-to-end latency over LLM parsing, and system concurrency is limited by shared GPU inference capacity rather than worker count. The goal is to provide practitioners with concrete architectural patterns for production-scale document understanding systems.", "Why Latent Actions Fail, and How to Prevent It (2026-05-21): This paper investigates why latent action models (LAMs) fail due to exogenous state (e.g., background clutter) when learning action representations from unlabeled videos. Extending a linear LAM framework, the authors find that minimizing the standard reconstruction objective causes latent actions to encode exogenous information from future observations, and that focusing on endogenous components in the representation space mitigates noise. Furthermore, auxiliary objectives like action-supervision provably encourage consistency of latent actions across exogenous states. Experiments on linear and nonlinear LAMs validate the theoretic", "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance (2026-05-20): 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." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_8d5fd97b76b9d62a", "radar_1eea04c5503e2f54", "radar_1488431b28a1ca16", "radar_0bed1c17d6ef1231" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The next phase of OpenAI’s Education for Countries (2026-05-20): OpenAI advances Education for Countries, expanding AI adoption in schools with new partnerships, teacher training, and tools to improve global learning outcomes.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_f4ae49772d646fad", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production (2026-05-21): This paper presents a microservice architecture for document AI that encapsulates pipelines for classification, OCR, and LLM-based structured field extraction, capable of processing thousands of multi-page documents per hour. Key findings include: OCR dominates end-to-end latency over LLM parsing, and system concurrency is limited by shared GPU inference capacity rather than worker count. The goal is to provide practitioners with concrete architectural patterns for production-scale document understanding systems." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_8d5fd97b76b9d62a", "radar_5b1af1f62cebeac2" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "ReactiveGWM: Steering NPC in Reactive Game World Models (2026-05-18): 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." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_66fe731da0cf1113", "radar_a369622fa45fb443", "radar_5b1af1f62cebeac2", "radar_4010b14c3118eeb7" ], "summary": "6 radar item(s) matched this section. 1 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 14, "status": "needs_review", "summary": "Deterministic weekly preview from 56 usable radar item(s). 53 included and 3 needs_review item(s). Top visible signal: \"Advancing content provenance for a safer, more transparent AI ecosystem\" from OpenAI News. Visible categories: safety, product_update, research, business, infrastructure.", "time_window": { "end": "2026-05-21T04:22:24.667+00:00", "start": "2026-05-14T04:22:24.667+00:00" }, "title": "Weekly AI Radar preview - ending May 21, 2026" }, { "id": "9fc18556-9a4e-4071-8b70-b6f8e8192109", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_51c3a2bb06ffd0ba", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8671, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "An OpenAI model has disproved a central conjecture in discrete geometry", "url": "https://openai.com/index/model-disproves-discrete-geometry-conjecture" }, { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_f4ae49772d646fad", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8671, "published_at": "2026-05-20T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "The next phase of OpenAI’s Education for Countries", "url": "https://openai.com/index/the-next-phase-of-education-for-countries" }, { "id": "radar_45988b1714b66da1", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "published_at": "2026-05-18T10:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_4959502399d31ee9", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex" }, { "id": "radar_1488431b28a1ca16", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance", "url": "https://arxiv.org/abs/2605.18793" }, { "id": "radar_0bed1c17d6ef1231", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Harnessing Self-Supervised Features for Art Classification", "url": "https://arxiv.org/abs/2605.18974" }, { "id": "radar_eaacfd054a1f2b19", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models", "url": "https://arxiv.org/abs/2605.18795" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_cd4b1cce0eefb05c", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8915, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4", "url": "https://arxiv.org/abs/2605.16562" }, { "id": "radar_8d5fd97b76b9d62a", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.2248, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production", "url": "https://arxiv.org/abs/2605.18818" } ], "confidence": 0.625, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 48 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 48 usable radar item(s). 46 included and 2 needs_review item(s). Top visible signal: \"An OpenAI model has disproved a central conjecture in discrete geometry\" from OpenAI News. Visible categories: research, safety, product_update, business, benchmark. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-21T02:05:48.920Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-21T02:05:51.02891+00:00", "sections": [ { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The next phase of OpenAI’s Education for Countries (2026-05-20): OpenAI announced the next phase of its Education for Countries initiative, expanding AI adoption in schools through new partnerships, teacher training, and tools to improve global learning outcomes.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_f4ae49772d646fad", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "An OpenAI model has disproved a central conjecture in discrete geometry (2026-05-20): An OpenAI model solved the 80-year-old unit distance problem, disproving a major conjecture in discrete geometry and marking a milestone in AI-driven mathematics.", "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance (2026-05-20): 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.", "Harnessing Self-Supervised Features for Art Classification (2026-05-20): 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." ], "caveats": [], "citations": [ "radar_51c3a2bb06ffd0ba", "radar_6e0f17d9d665540d", "radar_1488431b28a1ca16", "radar_0bed1c17d6ef1231", "radar_eaacfd054a1f2b19" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The next phase of OpenAI’s Education for Countries (2026-05-20): OpenAI announced the next phase of its Education for Countries initiative, expanding AI adoption in schools through new partnerships, teacher training, and tools to improve global learning outcomes.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_f4ae49772d646fad", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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" ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_5b1af1f62cebeac2", "radar_cd4b1cce0eefb05c" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production (2026-05-20): This paper presents a microservice architecture for production document AI, encapsulating pipelines for classification, OCR, and LLM-based structured field extraction, based on experience processing thousands of multi-page documents per hour. 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 over language model parsing, and system saturation is determined by shared GPU capa", "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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" ], "caveats": [], "citations": [ "radar_66fe731da0cf1113", "radar_8d5fd97b76b9d62a", "radar_5b1af1f62cebeac2" ], "summary": "6 radar item(s) matched this section.", "title": "Weak signals / needs_review" } ], "source_item_count": 15, "status": "needs_review", "summary": "Deterministic weekly preview from 48 usable radar item(s). 46 included and 2 needs_review item(s). Top visible signal: \"An OpenAI model has disproved a central conjecture in discrete geometry\" from OpenAI News. Visible categories: research, safety, product_update, business, benchmark.", "time_window": { "end": "2026-05-21T01:59:50.527+00:00", "start": "2026-05-14T01:59:50.527+00:00" }, "title": "Weekly AI Radar preview - ending May 21, 2026" }, { "id": "def910a5-d895-4001-b825-935791bf3d23", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_1488431b28a1ca16", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance", "url": "https://arxiv.org/abs/2605.18793" }, { "id": "radar_0bed1c17d6ef1231", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Harnessing Self-Supervised Features for Art Classification", "url": "https://arxiv.org/abs/2605.18974" }, { "id": "radar_eaacfd054a1f2b19", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models", "url": "https://arxiv.org/abs/2605.18795" }, { "id": "radar_9d2bda13a994f350", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation", "url": "https://arxiv.org/abs/2605.18956" }, { "id": "radar_8d5fd97b76b9d62a", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.2248, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production", "url": "https://arxiv.org/abs/2605.18818" }, { "id": "radar_47d31ae62f736dc6", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos", "url": "https://arxiv.org/abs/2605.18984" }, { "id": "radar_cb96e49c0763bb7c", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8748, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German", "url": "https://arxiv.org/abs/2605.19069" }, { "id": "radar_1ca98fbb9b413a4e", "collected_at": "2026-05-21T01:51:10.614+00:00", "confidence": 0.8581, "published_at": "2026-05-20T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking", "url": "https://arxiv.org/abs/2605.19077" } ], "confidence": 0.75, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 12 条可用雷达条目和 8 条引用生成。 Deterministic daily preview from 12 usable radar item(s). 12 included and 0 needs_review item(s). Top visible signal: \"Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance\" from arXiv cs.LG. Visible categories: research, benchmark, agent. Model / product / company updates: 3 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 1 radar item(s) matched this section.", "generated_at": "2026-05-21T02:05:42.956Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-21T02:05:45.508431+00:00", "sections": [ { "bullets": [ "MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation (2026-05-20): 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 ", "Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos (2026-05-20): 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-", "Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German (2026-05-20): 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." ], "caveats": [], "citations": [ "radar_9d2bda13a994f350", "radar_47d31ae62f736dc6", "radar_cb96e49c0763bb7c" ], "summary": "3 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance (2026-05-20): 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.", "Harnessing Self-Supervised Features for Art Classification (2026-05-20): 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.", "HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models (2026-05-20): 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.", "MotionMERGE: A Multi-granular Framework for Human Motion Editing, Reasoning, Generation, and Explanation (2026-05-20): 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 " ], "caveats": [], "citations": [ "radar_1488431b28a1ca16", "radar_0bed1c17d6ef1231", "radar_eaacfd054a1f2b19", "radar_9d2bda13a994f350", "radar_8d5fd97b76b9d62a" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking (2026-05-20): 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." ], "caveats": [], "citations": [ "radar_1ca98fbb9b413a4e" ], "summary": "1 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "No usable radar evidence currently supports this section." ], "caveats": [], "citations": [], "summary": "No retrieved radar items in this window support this section.", "title": "Business / ecosystem" }, { "bullets": [ "Operationalizing Document AI: A Microservice Architecture for OCR and LLM Pipelines in Production (2026-05-20): This paper presents a microservice architecture for production document AI, encapsulating pipelines for classification, OCR, and LLM-based structured field extraction, based on experience processing thousands of multi-page documents per hour. 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 over language model parsing, and system saturation is determined by shared GPU capa" ], "caveats": [], "citations": [ "radar_8d5fd97b76b9d62a" ], "summary": "1 radar item(s) matched this section.", "title": "Weak signals / needs_review" } ], "source_item_count": 8, "status": "needs_review", "summary": "Deterministic daily preview from 12 usable radar item(s). 12 included and 0 needs_review item(s). Top visible signal: \"Dimensional Balance Improves Large Scale Spatiotemporal Prediction Performance\" from arXiv cs.LG. Visible categories: research, benchmark, agent.", "time_window": { "end": "2026-05-21T01:59:50.527+00:00", "start": "2026-05-20T01:59:50.527+00:00" }, "title": "Daily AI Radar preview - May 21, 2026" }, { "id": "683cd568-9ee9-4ff3-bb5d-4e644c17b0e2", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "Supabase coverage depends on rows already persisted into the public retrieval view.", "Live DeepSeek synthesis failed; deterministic report draft is shown instead." ], "citations": [ { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_45988b1714b66da1", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8839, "published_at": "2026-05-18T10:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments", "url": "https://openai.com/index/dell-codex-enterprise-partnership" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.2339, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_4959502399d31ee9", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex" }, { "id": "radar_2bef32d76ec9b113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A", "url": "https://arxiv.org/abs/2605.16359" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_d76af0556f232088", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra", "url": "https://arxiv.org/abs/2605.16259" }, { "id": "radar_9a57f4127aadeb8d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "The Scaling Laws of Skills in LLM Agent Systems", "url": "https://arxiv.org/abs/2605.16508" }, { "id": "radar_a369622fa45fb443", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206" }, { "id": "radar_cd4b1cce0eefb05c", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8915, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4", "url": "https://arxiv.org/abs/2605.16562" }, { "id": "radar_4010b14c3118eeb7", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "ReactiveGWM: Steering NPC in Reactive Game World Models", "url": "https://arxiv.org/abs/2605.15256" }, { "id": "radar_8f9b517bb2f4740f", "collected_at": "2026-05-18T08:37:47.631+00:00", "confidence": 0.17, "source_name": "Christopher Olah", "status": "included", "title": "Home - colah's blog", "url": "http://colah.github.io/" } ], "confidence": 0.659, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 33 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 33 usable radar item(s). 31 included and 2 needs_review item(s). Top visible signal: \"Advancing content provenance for a safer, more transparent AI ecosystem\" from OpenAI News. Visible categories: safety, product_update, research, business, agent. Model / product / company updates: 5 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-20T02:39:05.980Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-20T02:39:08.347729+00:00", "sections": [ { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "How data science teams use Codex (2026-05-15): OpenAI published an article explaining how data science teams can use Codex to automate tasks such as creating root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9", "radar_5b1af1f62cebeac2" ], "summary": "5 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A (2026-05-19): 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.", "Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra (2026-05-19): 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 gui", "The Scaling Laws of Skills in LLM Agent Systems (2026-05-19): 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" ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_2bef32d76ec9b113", "radar_d76af0556f232088", "radar_9a57f4127aadeb8d", "radar_a369622fa45fb443" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "How data science teams use Codex (2026-05-15): OpenAI published an article explaining how data science teams can use Codex to automate tasks such as creating root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_4959502399d31ee9", "radar_9a57f4127aadeb8d" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments (2026-05-18): OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments, helping enterprises securely deploy AI coding agents across data and workflows.", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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" ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_45988b1714b66da1", "radar_66fe731da0cf1113", "radar_5b1af1f62cebeac2", "radar_cd4b1cce0eefb05c" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to provide ChatGPT Plus and AI training to all citizens.", "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "ReactiveGWM: Steering NPC in Reactive Game World Models (2026-05-18): 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." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_66fe731da0cf1113", "radar_a369622fa45fb443", "radar_5b1af1f62cebeac2", "radar_4010b14c3118eeb7", "radar_8f9b517bb2f4740f" ], "summary": "6 radar item(s) matched this section. 1 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 13, "status": "needs_review", "summary": "Deterministic weekly preview from 33 usable radar item(s). 31 included and 2 needs_review item(s). Top visible signal: \"Advancing content provenance for a safer, more transparent AI ecosystem\" from OpenAI News. Visible categories: safety, product_update, research, business, agent.", "time_window": { "end": "2026-05-20T02:29:58.738+00:00", "start": "2026-05-13T02:29:58.738+00:00" }, "title": "Weekly AI Radar preview - ending May 20, 2026" }, { "id": "aff83d70-aa8b-4282-b218-ccf94849b292", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "Supabase coverage depends on rows already persisted into the public retrieval view.", "Live DeepSeek synthesis failed; deterministic report draft is shown instead." ], "citations": [ { "id": "radar_6e0f17d9d665540d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8671, "published_at": "2026-05-19T10:45:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "Advancing content provenance for a safer, more transparent AI ecosystem", "url": "https://openai.com/index/advancing-content-provenance" }, { "id": "radar_2bef32d76ec9b113", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A", "url": "https://arxiv.org/abs/2605.16359" }, { "id": "radar_d76af0556f232088", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra", "url": "https://arxiv.org/abs/2605.16259" }, { "id": "radar_9a57f4127aadeb8d", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "The Scaling Laws of Skills in LLM Agent Systems", "url": "https://arxiv.org/abs/2605.16508" }, { "id": "radar_cd4b1cce0eefb05c", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8915, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4", "url": "https://arxiv.org/abs/2605.16562" }, { "id": "radar_d15b00276fe9c755", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8249, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Noise2Params: Unification and Parameter Determination from Noise via a Probabilistic Event Camera Model", "url": "https://arxiv.org/abs/2605.16317" }, { "id": "radar_9d267e1ad2ccf835", "collected_at": "2026-05-20T02:24:00.683+00:00", "confidence": 0.8581, "published_at": "2026-05-19T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures", "url": "https://arxiv.org/abs/2605.16551" } ], "confidence": 0.813, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 8 条可用雷达条目和 7 条引用生成。 Deterministic daily preview from 8 usable radar item(s). 8 included and 0 needs_review item(s). Top visible signal: \"Advancing content provenance for a safer, more transparent AI ecosystem\" from OpenAI News. Visible categories: safety, product_update, research, agent, infrastructure. Model / product / company updates: 1 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 3 radar item(s) matched this section.", "generated_at": "2026-05-20T02:34:59.675Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-20T02:35:02.683602+00:00", "sections": [ { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d" ], "summary": "1 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "How Many Visual Tokens Do Multimodal Language Models Need? Scaling Visual Token Pruning with F^3A (2026-05-19): 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.", "Systematic Optimization of Real-Time Diffusion Model Inference on Apple M3 Ultra (2026-05-19): 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 gui", "The Scaling Laws of Skills in LLM Agent Systems (2026-05-19): 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" ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_2bef32d76ec9b113", "radar_d76af0556f232088", "radar_9a57f4127aadeb8d", "radar_d15b00276fe9c755" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "The Scaling Laws of Skills in LLM Agent Systems (2026-05-19): 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", "PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures (2026-05-19): 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." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_9a57f4127aadeb8d", "radar_9d267e1ad2ccf835" ], "summary": "3 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Advancing content provenance for a safer, more transparent AI ecosystem (2026-05-19): OpenAI advances AI content provenance with Content Credentials, SynthID, and a verification tool to help people identify and trust AI-generated media.", "Scaling Accessible Mathematics on arXiv: HTML Conversion and MathML 4 (2026-05-19): 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." ], "caveats": [], "citations": [ "radar_6e0f17d9d665540d", "radar_cd4b1cce0eefb05c" ], "summary": "2 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "No usable radar evidence currently supports this section." ], "caveats": [], "citations": [], "summary": "No retrieved radar items in this window support this section.", "title": "Weak signals / needs_review" } ], "source_item_count": 7, "status": "needs_review", "summary": "Deterministic daily preview from 8 usable radar item(s). 8 included and 0 needs_review item(s). Top visible signal: \"Advancing content provenance for a safer, more transparent AI ecosystem\" from OpenAI News. Visible categories: safety, product_update, research, agent, infrastructure.", "time_window": { "end": "2026-05-20T02:29:58.738+00:00", "start": "2026-05-19T02:29:58.738+00:00" }, "title": "Daily AI Radar preview - May 20, 2026" }, { "id": "71e96d51-c942-48b9-a677-632ccfbd8d30", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_4959502399d31ee9", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "How data science teams use Codex", "url": "https://openai.com/academy/codex-for-work/how-data-science-teams-use-codex" }, { "id": "radar_a369622fa45fb443", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206" }, { "id": "radar_a4546429147dce6a", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8748, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Deep Pre-Alignment for VLMs", "url": "https://arxiv.org/abs/2605.15300" }, { "id": "radar_79d12ca4aa91147e", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Fluency and Faithfulness in Human and Machine Literary Translation", "url": "https://arxiv.org/abs/2605.15282" }, { "id": "radar_d4ce2b3966a2e234", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "One Pass Is Not Enough: Recursive Latent Refinement for Generative Models", "url": "https://arxiv.org/abs/2605.15309" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_66fe731da0cf1113", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8338, "published_at": "2026-05-16T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "OpenAI and Malta partner to bring ChatGPT Plus to all citizens", "url": "https://openai.com/index/malta-chatgpt-plus-partnership" }, { "id": "radar_25c14f78c707362c", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8671, "published_at": "2026-05-15T00:00:00+00:00", "source_name": "OpenAI News", "status": "included", "title": "A new personal finance experience in ChatGPT", "url": "https://openai.com/index/personal-finance-chatgpt" }, { "id": "radar_63afdc4d5a09f75b", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8249, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch", "url": "https://arxiv.org/abs/2605.15204" }, { "id": "radar_b1d8dd4a445d5849", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8249, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination", "url": "https://arxiv.org/abs/2605.15207" }, { "id": "radar_3f20880877462dde", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "DeepSlide: From Artifacts to Presentation Delivery", "url": "https://arxiv.org/abs/2605.15202" }, { "id": "radar_886fae3dd2af9c8a", "collected_at": "2026-05-18T08:37:47.631+00:00", "confidence": 0.7034, "source_name": "Andrej Karpathy", "status": "included", "title": "Andrej Karpathy", "url": "https://karpathy.ai/" } ], "confidence": 0.7, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 24 条可用雷达条目和 12 条引用生成。 Deterministic weekly preview from 24 usable radar item(s). 22 included and 2 needs_review item(s). Top visible signal: \"How data science teams use Codex\" from OpenAI News. Visible categories: product_update, research, agent, safety, benchmark. Model / product / company updates: 4 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 5 radar item(s) matched this section.", "generated_at": "2026-05-19T01:53:20.905Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "weekly", "saved_at": "2026-05-19T01:53:21.333424+00:00", "sections": [ { "bullets": [ "How data science teams use Codex (2026-05-15): OpenAI published an article explaining how data science teams can use Codex to automate tasks such as creating root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to expand AI access, offering ChatGPT Plus and training to help citizens build practical AI skills and use AI responsibly.", "A new personal finance experience in ChatGPT (2026-05-15): 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." ], "caveats": [], "citations": [ "radar_4959502399d31ee9", "radar_5b1af1f62cebeac2", "radar_66fe731da0cf1113", "radar_25c14f78c707362c" ], "summary": "4 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Deep Pre-Alignment for VLMs (2026-05-18): 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.", "Fluency and Faithfulness in Human and Machine Literary Translation (2026-05-18): 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.", "One Pass Is Not Enough: Recursive Latent Refinement for Generative Models (2026-05-18): 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." ], "caveats": [], "citations": [ "radar_a369622fa45fb443", "radar_a4546429147dce6a", "radar_79d12ca4aa91147e", "radar_d4ce2b3966a2e234", "radar_5b1af1f62cebeac2" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "How data science teams use Codex (2026-05-15): OpenAI published an article explaining how data science teams can use Codex to automate tasks such as creating root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.", "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch (2026-05-18): 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.", "TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (2026-05-18): 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." ], "caveats": [], "citations": [ "radar_4959502399d31ee9", "radar_a369622fa45fb443", "radar_63afdc4d5a09f75b", "radar_b1d8dd4a445d5849", "radar_3f20880877462dde" ], "summary": "5 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "OpenAI and Malta partner to bring ChatGPT Plus to all citizens (2026-05-16): OpenAI partners with Malta to expand AI access, offering ChatGPT Plus and training to help citizens build practical AI skills and use AI responsibly.", "Andrej Karpathy (2026-05-18): The personal technical blog of Andrej Karpathy, former OpenAI and Tesla AI leader, featuring many tutorial articles on large language models that are accessible to non-technical audiences.", "Yarin Gal - Home Page | Oxford Machine Learning (2026-05-18): Personal homepage of Yarin Gal at Oxford Machine Learning. The page provides navigation links to bio, publications, talks, software, blog, and contact. As of collection time, only basic metadata with no substantive text content." ], "caveats": [], "citations": [ "radar_5b1af1f62cebeac2", "radar_66fe731da0cf1113", "radar_886fae3dd2af9c8a" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "ReactiveGWM: Steering NPC in Reactive Game World Models (2026-05-18): 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.", "Home - colah's blog (2026-05-18): Personal blog of Christopher Olah, featuring high-quality original content on deep learning, neural networks, and topology. It is frequently referenced by Chinese AI media and includes collaborations like Transformer Circuits and Distill." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_a369622fa45fb443", "radar_5b1af1f62cebeac2" ], "summary": "6 radar item(s) matched this section. 2 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 18, "status": "needs_review", "summary": "Deterministic weekly preview from 24 usable radar item(s). 22 included and 2 needs_review item(s). Top visible signal: \"How data science teams use Codex\" from OpenAI News. Visible categories: product_update, research, agent, safety, benchmark.", "time_window": { "end": "2026-05-18T09:45:24.199+00:00", "start": "2026-05-11T09:45:24.199+00:00" }, "title": "Weekly AI Radar preview - ending May 18, 2026" }, { "id": "c2ea6cb1-324c-4f20-9ae2-92d26b7f0fa5", "caveats": [ "Read-only Supabase public radar retrieval was used; no Supabase write path ran.", "2 item(s) are marked needs_review and require human confirmation before confident synthesis.", "This surface shows available AI Radar evidence only; it is not a claim of complete current AI industry coverage.", "This is a deterministic preview, not a published report.", "No live DeepSeek call, Supabase write, or scheduled persistence job was run.", "Supabase coverage depends on rows already persisted into the public retrieval view." ], "citations": [ { "id": "radar_a369622fa45fb443", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices", "url": "https://arxiv.org/abs/2605.15206" }, { "id": "radar_a4546429147dce6a", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8748, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "Deep Pre-Alignment for VLMs", "url": "https://arxiv.org/abs/2605.15300" }, { "id": "radar_79d12ca4aa91147e", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CL", "status": "included", "title": "Fluency and Faithfulness in Human and Machine Literary Translation", "url": "https://arxiv.org/abs/2605.15282" }, { "id": "radar_d4ce2b3966a2e234", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.CV", "status": "included", "title": "One Pass Is Not Enough: Recursive Latent Refinement for Generative Models", "url": "https://arxiv.org/abs/2605.15309" }, { "id": "radar_5b1af1f62cebeac2", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.2248, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "url": "https://arxiv.org/abs/2605.15208" }, { "id": "radar_63afdc4d5a09f75b", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8249, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch", "url": "https://arxiv.org/abs/2605.15204" }, { "id": "radar_b1d8dd4a445d5849", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8249, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.LG", "status": "included", "title": "TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination", "url": "https://arxiv.org/abs/2605.15207" }, { "id": "radar_3f20880877462dde", "collected_at": "2026-05-18T09:35:07.557+00:00", "confidence": 0.8581, "published_at": "2026-05-18T04:00:00+00:00", "source_name": "arXiv cs.AI", "status": "included", "title": "DeepSlide: From Artifacts to Presentation Delivery", "url": "https://arxiv.org/abs/2605.15202" }, { "id": "radar_886fae3dd2af9c8a", "collected_at": "2026-05-18T08:37:47.631+00:00", "confidence": 0.7034, "source_name": "Andrej Karpathy", "status": "included", "title": "Andrej Karpathy", "url": "https://karpathy.ai/" }, { "id": "radar_52aed8fac227edc4", "collected_at": "2026-05-18T08:37:47.631+00:00", "confidence": 0.7297, "source_name": "Yarin Gal", "status": "included", "title": "Yarin Gal - Home Page | Oxford Machine Learning", "url": "http://yarin.co/" }, { "id": "radar_4370138e50c44321", "collected_at": "2026-05-18T07:19:47.088+00:00", "confidence": 0.8296, "source_name": "Implications", "status": "included", "title": "Archive - Implications, by Scott Belsky", "url": "https://www.implications.com/archive" }, { "id": "radar_7fe3bd6d4f17fc49", "collected_at": "2026-05-18T07:19:47.088+00:00", "confidence": 0.77, "source_name": "Elad Gil", "status": "included", "title": "Archive - Elad Blog", "url": "https://blog.eladgil.com/archive" } ], "confidence": 0.721, "data_source": "supabase_radar_items", "executive_summary": "本报告草稿基于 21 条可用雷达条目和 12 条引用生成。 Deterministic daily preview from 21 usable radar item(s). 19 included and 2 needs_review item(s). Top visible signal: \"AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices\" from arXiv cs.LG. Visible categories: research, agent, safety, benchmark, open_source. Model / product / company updates: 1 radar item(s) matched this section. Research / open-source: 5 radar item(s) matched this section. Agents / products: 4 radar item(s) matched this section.", "generated_at": "2026-05-19T01:53:15.170Z", "missing_evidence": [], "mode": "saved_candidate", "report_type": "daily", "saved_at": "2026-05-19T01:53:15.87307+00:00", "sections": [ { "bullets": [ "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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" ], "caveats": [], "citations": [ "radar_5b1af1f62cebeac2" ], "summary": "1 radar item(s) matched this section.", "title": "Model / product / company updates" }, { "bullets": [ "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Deep Pre-Alignment for VLMs (2026-05-18): 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.", "Fluency and Faithfulness in Human and Machine Literary Translation (2026-05-18): 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.", "One Pass Is Not Enough: Recursive Latent Refinement for Generative Models (2026-05-18): 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." ], "caveats": [], "citations": [ "radar_a369622fa45fb443", "radar_a4546429147dce6a", "radar_79d12ca4aa91147e", "radar_d4ce2b3966a2e234", "radar_5b1af1f62cebeac2" ], "summary": "5 radar item(s) matched this section.", "title": "Research / open-source" }, { "bullets": [ "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch (2026-05-18): 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.", "TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (2026-05-18): 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.", "DeepSlide: From Artifacts to Presentation Delivery (2026-05-18): 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 qu" ], "caveats": [], "citations": [ "radar_a369622fa45fb443", "radar_63afdc4d5a09f75b", "radar_b1d8dd4a445d5849", "radar_3f20880877462dde" ], "summary": "4 radar item(s) matched this section.", "title": "Agents / products" }, { "bullets": [ "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "Andrej Karpathy (2026-05-18): The personal technical blog of Andrej Karpathy, former OpenAI and Tesla AI leader, featuring many tutorial articles on large language models that are accessible to non-technical audiences.", "Yarin Gal - Home Page | Oxford Machine Learning (2026-05-18): Personal homepage of Yarin Gal at Oxford Machine Learning. The page provides navigation links to bio, publications, talks, software, blog, and contact. As of collection time, only basic metadata with no substantive text content.", "Archive - Implications, by Scott Belsky (2026-05-18): This is the archive page of Implications newsletter by Scott Belsky, listing multiple past articles but without providing article content." ], "caveats": [], "citations": [ "radar_5b1af1f62cebeac2", "radar_886fae3dd2af9c8a", "radar_52aed8fac227edc4", "radar_4370138e50c44321", "radar_7fe3bd6d4f17fc49" ], "summary": "5 radar item(s) matched this section.", "title": "Business / ecosystem" }, { "bullets": [ "AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices (2026-05-18): 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).", "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels (2026-05-18): 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", "ReactiveGWM: Steering NPC in Reactive Game World Models (2026-05-18): 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.", "Home - colah's blog (2026-05-18): Personal blog of Christopher Olah, featuring high-quality original content on deep learning, neural networks, and topology. It is frequently referenced by Chinese AI media and includes collaborations like Transformer Circuits and Distill." ], "caveats": [ "needs_review items should use cautious language and require confirmation." ], "citations": [ "radar_a369622fa45fb443", "radar_5b1af1f62cebeac2" ], "summary": "6 radar item(s) matched this section. 2 still need review.", "title": "Weak signals / needs_review" } ], "source_item_count": 16, "status": "needs_review", "summary": "Deterministic daily preview from 21 usable radar item(s). 19 included and 2 needs_review item(s). Top visible signal: \"AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices\" from arXiv cs.LG. Visible categories: research, agent, safety, benchmark, open_source.", "time_window": { "end": "2026-05-18T09:45:24.199+00:00", "start": "2026-05-17T09:45:24.199+00:00" }, "title": "Daily AI Radar preview - May 18, 2026" } ], "caveats": [ "Cloudflare Pages is the primary public read surface. Auth, Admin, server actions, and write workflows remain outside this public Cloudflare surface.", "Only public-safe radar and report fields are included. Private raw content, provider metadata, internal notes, service-role access, and secrets are excluded.", "Snapshot data came from Supabase public-safe read views using anon read access." ] }