Mathematicians are challenging OpenAI’s account of how it developed a potential solution to the Navier-Stokes problem. The company presented the 166-page proof in early September and said it followed an approach by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa.
Several groups of experts disagree, Scientific American reports. “When you go to the details, it is a different approach,” said Gonzalo Cao-Labora, a mathematician at EPFL. He said the similarities are only surface-level.
The potential proof is written in Lean, a formal verification language that checks each logical step. That does not, by itself, make the work readable to humans.
A mathematician accused OpenAI of using prior work
Before OpenAI published the proof, NYU mathematician Tristan Buckmaster accused the company of building on work he had done with Anthropic’s Levent Alpöge using publicly available models. OpenAI denied using Buckmaster’s data. Camillo De Lellis of the Institute for Advanced Study said a definitive answer is probably impossible because language models are black boxes.
In early October, OpenAI released roughly 400 more results. The math community has struggled to keep up with them. Many mathematicians were angered by their publication as a loose collection on GitHub and felt it cheapened their field.
Read nextFTC probes OpenAI and Anthropic over AI safety incidentsOpenAI used 10,000 agents to develop the solution
OpenAI used about 10,000 model instances over roughly 88 hours and 130 billion tokens to develop the potential Navier-Stokes solution. OpenAI researcher Noam Brown said nobody knows what that large group of agents contributed to the result.
Behavior at that scale has not been studied because systematic comparisons involving thousands of agents cost too much. No one has tested how long a single agent would take to solve the same problem. It is also nearly impossible to identify a concrete advantage of 10,000 agents over 1,000.
Brown attributes less than 10 percent of the success to the multi-agent setup. He said these new systems get a disproportionate share of the credit. In his view, the base model’s ability to work on tasks for long periods and run in parallel matters more.
The Clay Mathematics Institute has been reviewing the potential proof since its release.
