OpenAI says it can't rule out researchers' Codex data aided Navier–Stokes breakthrough

OpenAI said Tuesday it cannot rule out that de-identified data from two mathematicians' use of its Codex tool helped improve the model behind its claimed solution to a $1 million math problem that has stumped researchers for nearly a century.

The ChatGPT developer announced it had cracked the Navier-Stokes existence and smoothness problem after deploying roughly 10,000 AI agents that worked in parallel and reached a resolution in about 88 hours.

But hours before the company paraded its milestone, NYU mathematics professor Tristan Buckmaster published a statement alleging the company rushed its work after learning his team was closing in on a breakthrough.

Buckmaster and Levent Alpöge, a mathematician at OpenAI rival Anthropic, had been quietly working on the same problem for nearly a year. The duo used OpenAI's Codex while storing drafts and research notes in the tool.

Questions over Codex data

When Buckmaster asked OpenAI whether its model had been trained on or had access to their Codex sessions, he says he got a murky response. "I was told the model did not look up user data. I asked again, about training, and I did not get an answer," he wrote.

OpenAI said no specific user data was accessed in its Navier-Stokes work, while acknowledging that "We cannot rule out that de-identified data derived from their usage of our products helped improve our models."

The dispute quickly turned personal.

Buckmaster alleged that OpenAI proposed arrangements that would have excluded Alpöge from credit, citing his Anthropic employment as the complication. He said OpenAI researcher Sébastien Bubeck asked him, "Why would you ruin your career?" and later warned, "If you don't want me to be nice, then I don't have to be nice."

OpenAI CEO Sam Altman defended Bubeck, saying he and others had acted with "integrity and generosity." Altman said OpenAI initially believed Buckmaster and Alpöge had solved Navier-Stokes and wanted to coordinate a joint release. After learning they'd solved a related Euler problem instead, OpenAI offered them first release, suggested they should get the prize, and proposed that Buckmaster lead a rewrite of OpenAI's proof, according to the CEO.

Bubeck separately disputed Buckmaster's account, saying his concern was whether an Anthropic employee should contribute to OpenAI's work. He acknowledged that telling Buckmaster, "Why would you risk your career?" was a poor choice of words, insisting it was intended as concern, not a threat.

Buckmaster has said he is not accusing OpenAI of anything and hasn't seen its full proof. OpenAI said it does not intend to claim the $1 million Clay prize.