Mathematicians question OpenAI's understanding of its Navier-Stokes proof

In early September OpenAI published a 166-page potential solution to the Navier-Stokes problem, claiming it followed the approach of Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. But EPFL mathematician Gonzalo Cao-Labora says a closer look shows a different method, with the similarity only superficial. The proof is written in Lean, so each logical step can be checked, but it is hard for humans to read; the Clay Mathematics Institute has been reviewing it since release. The dispute also concerns provenance. Before publication, NYU mathematician Tristan Buckmaster accused OpenAI of building on work he did with Anthropic's Levent Alpöge using public models; OpenAI denies using their data. Camillo De Lellis of the Institute for Advanced Study thinks a definitive answer may be impossible because language models are black boxes. In early October OpenAI loosely posted about 400 new results on GitHub, which the math community struggled to keep up with; many feel it devalues the field. Compute is another question: OpenAI deployed about 10,000 model instances running roughly 88 hours and consuming 130 billion tokens. Researcher Noam Brown admits no one really knows what the agent swarm contributed, attributing under 10% of the success to the multi-agent setup and saying the key is the base model's ability to work on tasks for long stretches and run in parallel. For researchers in formal verification and long-horizon reasoning, this means evaluating methods should measure baseline models and swarm scale separately.