OpenAI says it cracked 90-year-old maths problem in 88 hours

OpenAI says one of its newest internal AI models helped crack a famous maths problem in just 88 hours, after the company sent roughly 10,000 AI agents at it. The target was the Navier-Stokes existence and smoothness problem, part of the equations that describe how fluids move. For 90 years, key parts of that problem have lacked proof.

The company is calling the result a milestone, but not a victory lap. Its proof has not been independently checked, and it has not been accepted by the Clay Mathematics Institute, which oversees the Millennium Prize. OpenAI also said it does not plan to claim the prize.

The timeline is tight. OpenAI said it started training the new model at the end of August, and that by 1 September it had heard rumours that two Millennium Prize problems had been solved. That pushed researchers to aim the model at the remaining problems. By 5 September, the company said the bots had found a solution to Navier-Stokes.

The scale of the effort is hard to ignore. OpenAI said the bots exchanged nearly 3 million messages and used 130 billion output tokens on Navier-Stokes alone. Based on the company’s own pricing for its most advanced models, that would have cost about $10m. And the result only resolves two of the four statements the prize requires.

There’s also already a dispute hanging over the announcement. New York University maths professor Tristan Buckmaster said he and Anthropic’s Levent Alpöge had been working on the same problem, and claimed OpenAI only started after learning of their progress. OpenAI pushed back, said it had not seen their work before public release, and added that it cannot rule out de-identified data from product use having helped its models. The proof itself may be mathematical, but the race around it is very human.