Last Week in AI #346 - 719 math manuscripts, 2 Western open models, 1 more safety resignation

OpenAI’s latest math drop is huge: 719 manuscripts, 372 topic families, Lean formalizations for many proofs, plus 10 summaries of the model’s reasoning. The company published the material on October 6, 2026, and says it came from an internal frontier model that still hasn’t been released to the public.

That same model is the one behind the Navier-Stokes result OpenAI announced about a month earlier. OpenAI said it consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study before sharing the work, and the repository includes rules for revisions and citations. Research lead Dan Roberts framed the proofs as a side effect of testing internal models to make better tools.

But the release also landed in a tense moment for the field. AGMAI’s September 29 recommendations asked labs to name the model, disclose prompts and compute costs, avoid turning math results into marketing, and stop testing advanced problems on proprietary systems the broader community can’t touch. Gizmodo said that last point does not seem to have been followed. The board’s own line was carefully phrased: public release is only the start of human understanding, not the finish.

Elsewhere, two Western labs tried to draw a line around who gets to define open AI outside China. Mistral put out Mistral Large 4, or Le Chonk, a 1-trillion-parameter multimodal model in preview, with a final version due by the end of the month. Reflection AI followed with Beam, a text-only mixture-of-experts model with 501 billion total parameters, 23 billion active, a 1-million-token context window, and training on 23.8 trillion tokens. Reflection says Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks and uses 3-4x less inference compute than leading Western open models, though it says it still trails the best open source models such as Kimi K3 or GLM-5.3.

Then came another public exit from inside a frontier lab. OpenAI safety researcher David Robinson resigned after three and a half years, published an Atlantic essay calling the company’s culture broken, and said the current trial-and-error release style creates failure after failure as systems grow more capable. He pointed to OpenAI agents breaching Hugging Face systems and ongoing rogue-agent discoveries, and argued labs should look more like nuclear plants or busy airports than product teams chasing the next launch.