OpenAI posts 719 math proofs, two Western open-weight models debut, safety lead quits
Last Week in AI · rss · 2026-10-09
- OpenAI publishes math from an unreleased frontier model: 719 manuscripts across 372 topic families, many with Lean formalizations; the same model produced the earlier Navier-Stokes result. AGMAI's recommendation to stop testing advanced problems on proprietary models appears unmet.
- Two Western open-weight launches: Mistral Large 4 ("Le Chonk"), a 1T-parameter multimodal model pitched as the strongest open-weight model outside China, focused on coding and cyberdefense. Reflection AI's Beam is a 501B/23B-active MoE pretrained on 23.8T tokens with 1M-token context, matching GLM-5.2 on reasoning while using 3-4x less inference compute; weights due this month.
- David Robinson resigns from OpenAI: after 3.5 years, having led the Preparedness Framework and safety reports for 12 frontier launches, he argues OpenAI's trial-and-error release model guarantees failures that scale with capability, and calls for nuclear-plant-style layered safety and stronger external incentives.
- Watermarking: Google opens SynthID Detector to the public (1M verifications/day; OpenAI, Nvidia, Kakao support it); OpenAI will watermark ChatGPT/Codex text in the EU to comply with the EU AI Act.
More from Companies & People
- Anthropic rolled out nearly its whole 5.5 family in 15 days, flipping the narrative on OpenAI — haider1 · 2026-10-09
- Even With Teams of Agents, Companies Still Run on 10x Humans, Founder Observes — annbordetsky · 2026-10-09
- Pamela Fox's slides: how AI is reshaping software engineering and CS teaching — lee_stott · 2026-10-09
- MOSS workshop at COLM explores how small-scale research can matter without big compute — AdtRaghunathan · 2026-10-09
- AI safety researcher: "I don't think working at OpenAI is moral" — BlancheMinerva · 2026-10-09
- LeCun recalls 1986 MIT visit, showed Minsky a neural net multiplying binary numbers — ylecun · 2026-10-09