NYU Mathematician Alleges OpenAI Used Inside Information to Rush Solution to Million-Dollar Math Problem

NYU mathematics professor Tristan Buckmaster has lodged serious allegations against OpenAI, claiming the company obtained confidential information about his research into one of mathematics' most significant unsolved problems and then used its computational advantage to race him to a solution.

Buckmaster, working alongside Anthropic mathematician Levent Alpöge, announced preliminary findings on the Navier-Stokes existence and smoothness problem—one of the Clay Mathematics Institute's seven Millennium Prize problems, each worth $1 million to whoever solves it. Their results represent a genuine breakthrough in theoretical mathematics, but Buckmaster's accompanying statement reveals a dispute with OpenAI that goes well beyond typical academic rivalry.

How OpenAI Allegedly Obtained the Information#

According to Buckmaster's detailed public statement, word of his team's progress somehow reached OpenAI while they were still completing their work. When he contacted OpenAI to inquire, the company initially indicated it had already achieved a full proof. However, when pressed for specifics—such as when research began and how much human guidance was required—OpenAI's explanations grew contradictory. Buckmaster learned that an entire team had been dedicated to the problem and that an unusually large amount of computational resources had been deployed. OpenAI eventually acknowledged that the initial attempt had started only days before, after learning of Buckmaster's approach.

If accurate, this timeline suggests OpenAI recognized the validity of Buckmaster and Alpöge's strategy and mobilized its superior computing resources to reach the proof first.

The Suspicious Convergence#

Buckmaster emphasizes the improbability of the coincidence. The specific path they chose to attack the Navier-Stokes problem—through what he describes as "a smooth force" using options c and d from mathematician Fefferman's formulation—was a relatively obscure avenue pursued by very few researchers. "Almost nobody else I know of was working on it," Buckmaster notes. He argues that OpenAI's AI models, given only the problem statement, would be extremely unlikely to independently arrive at this particular approach within days.

This suspicion is further complicated by the fact that Buckmaster relied heavily on OpenAI's Codex tool during his research. OpenAI maintains the right to train its models on user interactions with Codex, though researchers can opt out. The possibility that Buckmaster's own Codex sessions—which may have contained his thinking process and intermediate work—could have influenced OpenAI's model when solving a similar problem adds another layer to his concerns.