Expert scrutiny has caught up with OpenAI’s most ambitious mathematics claim. A 166-page potential solution to the Navier-Stokes problem was posted in early September, and the company said it followed the approach of Diego Cordoba and Luis Martinez-Zoroa, two Spanish mathematicians.
Research groups now disagree. Gonzalo Cao-Labora of EPFL told Scientific American that only surface similarities hold, and that the approaches diverge once the detail is examined. Every step is verified by Lean, the language the draft is written in, but readability for humans is not something that verification supplies.
Reviewers at the Clay Mathematics Institute have had the document since it was released. A separate charge came from NYU’s Tristan Buckmaster, who accused OpenAI of building on unpublished work he did with Anthropic’s Levent Alpoge. His data was never used, OpenAI says, and with closed models proving otherwise is close to impossible.
The machinery raises its own questions. Roughly 10,000 model instances were thrown at the problem by OpenAI for about 88 hours, consuming 130 billion tokens. What the agent swarm contributed remains unknown, by researcher Noam Brown’s own account, and he puts its share of the credit below 10 percent.
Base model capability did the work, in his argument, and flashy multi-agent setups collect outsized credit. Running in parallel and grinding through tasks over long periods is what a base model, not the swarm, brings. Pointed words, from inside the company that shipped the swarm.
