Analysis
OpenAI announced September 8 that a multi-agent system -- an unreleased internal model coordinating up to 10,000 sub-agents -- had proved a blow-up condition in the Navier-Stokes equations, one of the seven Clay Mathematics Institute Millennium Prize Problems, each worth $1 million to solve, Axios reported. Within hours, the announcement turned into a credit dispute rather than a clean win.
- Tristan Buckmaster -- a mathematician at NYU's Courant Institute who says the approach OpenAI's system used was identical to one he had spent most of a year developing privately: "Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement."
- Levent Alpoge -- an Anthropic mathematician who worked with Buckmaster on the same problem using Claude and OpenAI's own Codex and GPT-5.6 Sol models; the pair's progress accelerated sharply in mid-August, shortly before OpenAI's own effort took off.
- Sebastien Bubeck -- the OpenAI researcher who led the project. Buckmaster says Bubeck pressured him during the dispute, asking "Why would you ruin your career?" and warning, "If you don't want me to be nice, then I don't have to be nice." Bubeck has called the allegations "false and inflammatory" and said, "We did not use their prompts or proofs to prompt our models or direct our agents."
- Terence Tao -- not a party to the dispute, but the Fields Medalist has separately warned that AI labs treating unsolved math problems as marketing opportunities risks "strip-mining" the field, since a model that outputs a proof without showing its failed attempts denies other mathematicians the intermediate insight that normally advances the field, TechCrunch reported.
“- Sebastien Bubeck -- the OpenAI researcher who led the project.”
OpenAI's most careful statement stopped short of a full denial: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." That line matters because Buckmaster and Alpoge did their work inside OpenAI's own Codex product, which by default can train on user data unless a user opts out -- meaning OpenAI's own terms of service leave room for exactly the scenario Buckmaster is describing, even without anyone at OpenAI directly reading his files.
