📰 News
-
Claude Chat and Cowork are now one product — Stop picking a tab. Claude routes the work. Docs and Slides export to PPT or PDF from the chat. Tasks keep running in the cloud after you close the laptop. Pro and Max first. → The Decoder
-
Anthropic and OpenAI want safety evaluators embedded — Amodei’s pitch, Altman signed on. Employee-like access for groups like METR and Redwood. Evaluators still want independence, time, and the right to publish. Apollo got three days on GPT-6 Astra. → TechCrunch
-
Von der Leyen: agents “escaping their environment” are a preview — She cited the Hugging Face incident and warned of hacking “on a level we never thought possible.” The EU wants Canada and the UK on evaluation, and will invite the frontier labs. → The Decoder
-
Your agents can now drive Google Home — Home MCP in early access. Claude, Hermes, ChatGPT, and others can talk to Nest and Matter devices. U.S. Premium Advanced ($20/month) first. → TechCrunch
-
Apple is building an enterprise AI server on M8 Ultra — Two or four chips, inference only. NVLink Fusion is on the table. Not before 2029, and it could still die. OpenAI and Anthropic are already buying Mac Minis in bulk. → The Decoder
-
DeepMind launched an institute for the AGI questions — Safety, governance, loss of control. Not just engineers: arts, humanities, policy. Hassabis, Shane Legg, James Manyika. → The Decoder
-
Your agent aced the task. Will it do it again? — IBM. A GPT-4.1 ReAct agent hits 77.4% on average, but only 53% of tasks five times in a row. Consistency guidelines cut the gap in half. → Hugging Face
🔥 GitHub Trending
-
JustVugg/colibri — Pure C, zero deps. Stream experts from disk and run GLM-5.2 (744B) or Kimi K3 (2.8T) on hardware you already own.
-
alphaXiv/OpenResearch — Turn Claude Code, Codex, or Cursor into research agents. Literature, hypotheses, experiments, all local-first.
-
melgarafael/DeskcommCRM — Self-hosted AI sales OS. WhatsApp agents that qualify and close. MCP-ready, multi-tenant.
▶️ YouTube
-
What AI Researchers Saw, Before Their Demand to ‘Pace’ AI — AI Explained. Why the slowdown chorus arrived at once. Six axes researchers are watching, plus the incident reports.
-
GPT-6 Spent 5 Days Building This City — Matthew Berman. GPT-6 spent five days stacking a city out of text.
💬 Community
-
Model Training Incidents are Negligence — Taggart. RubyGems, Hugging Face, PyPI. Not “misalignment” — sandboxing that never held. “Everyone is lying to you for money.”
-
The contagion of fear — Bryan Cantrill (via Simon). Don’t spend public trust on hand-wavy bioweapons and infrastructure hacks.
-
Why don’t machine learning research agents overfit? — Amazon Science. Winning strategies compress to 16–32 tokens. A short description has no room to memorize the test set.