OpenAI's 166-page Navier-Stokes proof uses a different approach, mathematicians say

OpenAI's Navier-Stokes effort came with unusually heavy machinery: about 10,000 model instances running for roughly 88 hours and 130 billion tokens. Yet according to OpenAI researcher Noam Brown, nobody actually knows what that massive agent swarm contributed, because behavior at that scale hasn't been studied and running systematic comparison tests with thousands of agents costs too much. Brown himself puts less than 10 percent of the success down to the multi-agent setup, saying these systems are new and flashy and get a disproportionate share of the credit; what really matters, he says, is the base model's ability to grind through tasks over long periods and run in parallel.

The provenance question is separate but related. Before publication, NYU mathematician Tristan Buckmaster accused OpenAI of building on work that he and Anthropic's Levent Alpöge had done with publicly available models, a claim OpenAI denied. Camillo De Lellis of the Institute for Advanced Study says a definitive answer is probably impossible because language models are black boxes — the same opacity that makes it hard to say where the proof's ideas came from.

Meanwhile, the volume of output is straining the field. In early October, OpenAI released roughly 400 more results as a loose collection on GitHub, which angered many mathematicians who felt it cheapened their field. The Clay Mathematics Institute has been reviewing the original proof since its release, but the dispute over method suggests the review will have to settle not just whether the proof is correct but what it actually is.