In early September, OpenAI released a 166-page potential solution to the Navier-Stokes problem, claiming the proof followed the approach of Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa.
Several groups of experts now disagree, Scientific American reports. Gonzalo Cao-Labora of EPFL says that when you go to the details, it is a different approach. He notes the similarities are only surface-level.
The potential proof is written in Lean, a formal verification language that checks every logical step but does not make it readable to humans on its own. The Clay Mathematics Institute (CMI) has been reviewing it since its release.
Before publication, NYU mathematician Tristan Buckmaster accused OpenAI of building on work that he and Levent Alpöge from Anthropic had done with publicly available models. OpenAI denied the claim, saying it never used Buckmaster’s data. Camillo De Lellis of the Institute for Advanced Study suggests a definitive answer is probably impossible because language models are black boxes.
In early October, OpenAI dumped roughly 400 more results that the math community can barely keep up with. Publishing them as a loose collection on GitHub angered many mathematicians, who felt it cheapened their field.
Nobody knows if 10,000 AI agents actually helped
There is another thing OpenAI does not seem to understand about its own work. The company threw a lot of compute at the effort, deploying about 10,000 model instances for roughly 88 hours and 130 billion tokens to develop the potential Navier-Stokes solution. According to OpenAI researcher Noam Brown, nobody actually knows what that massive agent swarm contributed to the result.
Behaviour at that scale has not been studied because running systematic comparison tests with thousands of agents costs too much. No one has tested how long a single agent would take to solve the same problem, and it is nearly impossible to tell what concrete advantage 10,000 agents offer over 1,000.
Brown puts less than 10 percent of the success down to the multi-agent setup. These systems are new and flashy, he says, so they get a disproportionate share of the credit. What really matters is the base model itself, which can grind through tasks over long periods and run in parallel.
