WIRED: OpenAI Navier-Stokes discovery draws academic controversy

OpenAI says it has found an AI-made solution to Navier-Stokes, the 200-year-old equation used to describe how fluids like water and air move. That would be a big mathematical result on its own. But the announcement arrived wrapped in a dispute over timing, credit, and whether the company moved too fast after learning other researchers were close.

The equation sits among the Clay Millennium Prize problems, each carrying a $1 million prize. Sebastien Bubeck, a mathematician and AI researcher at OpenAI, said the company began training a new model with stronger math skills on August 28. After hearing rumors that Anthropic might be making progress on Navier-Stokes, OpenAI shifted more resources to the problem and eventually deployed more than 1,000 agents over more than 50 hours. At one point, Bubeck said, as many as 10,000 agents were working on it.

OpenAI says that effort finally produced a proof that was formalized in Lean, a programming language used to check mathematical arguments. Mark Chen, OpenAI’s head of research, said the work required computing power far beyond what the company had spent on previous math projects, at a cost in the millions of dollars. Bubeck said he initially assumed there had to be a mistake, then saw that the team had a final solution by Sunday morning.

The controversy comes from Tristan Buckmaster, a mathematician at NYU, and Anthropic researcher Levent Alpöge, who on Monday posted documents saying they had made key advances related to the same problem, using AI models including Claude and Codex. Buckmaster says he later learned OpenAI had become aware of their work and had turned significant resources toward the problem. He also says he asked whether OpenAI had accessed the pair’s Codex logs, and claims the company made proposals that included publishing a paper announcing the problem had been solved by an internal OpenAI model without Alpöge’s name.

OpenAI publicly denied that it had looked at their prompts before the work was released. Bubeck also pushed back on the accusation that OpenAI had tried to strip Alpöge’s name from any publication. Still, the company’s own blog post leaves one awkward opening: it says it cannot rule out that de-identified data from use of its products may have helped improve its models. That’s a very Silicon Valley sentence for “we’re not admitting anything, but please enjoy the drama.”