OpenAI Claims Navier-Stokes Breakthrough, Faces Data Access Dispute

OpenAI's announcement of a solution to the Navier-Stokes problem arrives with both fanfare and suspicion. The problem, unsolved for around 90 years, has been a towering challenge in mathematics, and a credible solution would be a landmark achievement. The claim was made in a blog post on Tuesday, with the company detailing that it used an internal AI model, which it says is more powerful than its recently released GPT-6 Astra, and ran it with 10,000 concurrent agents. Training reportedly began on August 28th, and OpenAI states the model exhibited unprecedented performance on its benchmarks, including mathematics.

The controversy stems from a pre-announcement publication by Tristan Buckmaster, a mathematics professor at NYU, who had been working on a related problem with Levent Alpöge. Buckmaster raised concerns that OpenAI might have used his private session data from Codex, where he stored his drafts, to gain an edge. OpenAI's response, that it did not access 'specific user data' but cannot rule out that 'de-identified data' helped improve its models, has done little to quell the dispute. This incident underscores the growing sensitivity around data usage in AI training, particularly when academic research is involved.

The resolution of this dispute hinges on whether the details of Buckmaster's work can be shown to have influenced OpenAI's model. OpenAI claims its proofs differ significantly from Buckmaster's, but the company's own acknowledgment of potential indirect data influence suggests a grey area. Given the high stakes—including the $1 million prize, which OpenAI says it won't claim—the veracity of the solution and the ethical conduct around it will likely be scrutinized by the mathematics community. The test is whether OpenAI can provide convincing evidence that its model's training data did not include any of Buckmaster's private research, a standard it has yet to fully meet.