NYU Mathematician: OpenAI Threatened Me Over Navier-Stokes Proof Dispute

Tristan Buckmaster, a mathematics professor at New York University, has revealed new details behind the dispute over authorship of a proof of one of the Millennium Prize Problems. He says that when he announced he would publicly disclose the course of negotiations with OpenAI, a company representative asked him why he would want to destroy his own career.

The case concerns the Navier-Stokes equations, one of seven Millennium Prize Problems designated by the Clay Mathematics Institute, carrying a $1 million prize for a solution. Nowosci.ai previously reported that OpenAI announced it had solved the problem using a swarm of agents, and mathematicians spoke of stolen credit. Now specific, previously undisclosed details of the entire dispute are coming to light, including the names of the researchers and the content of conversations with the company.

A Race for Priority

According to Buckmaster's account, he and Alpöge worked independently on a solution to a simplified version of the problem, the Euler equations, and completed formal verification of their results on August 22. A few days later, Buckmaster says, signals reached OpenAI that the two researchers were close to announcing their own breakthrough.

On September 1, OpenAI reportedly directed the full computing power of its agentic system at the Millennium Prize Problems, involving about 10,000 autonomous agents. The company first spent roughly 50 hours testing its approach on simpler equations before moving on to the actual Navier-Stokes equation. The entire process, including a 17-hour formal verification of the proof in the Lean 4 system, took 88 hours and reportedly cost several million dollars in computing power.

Allegations Over Private Data Access

Buckmaster suggests his team's working calculations may have been processed by OpenAI's Codex coding tool, raising suspicions that the company's model could have been informed of his team's not-yet-published results. The researcher himself notes he has no direct insight into his rival's methods and does not explicitly accuse the company of stealing data, but he does not hide his suspicions.

OpenAI researcher Sebastien Bubeck categorically denied any unauthorized access to private files on the servers or directly to the two researchers' work. The company officially acknowledged, however, that it cannot entirely rule out that anonymized data from users' activity in its tools helped in model development, which in practice means queries from competing researchers could have indirectly fed into the system's training.

Why would you want to destroy your career? - question posed to Tristan Buckmaster by an OpenAI representative when he announced he would disclose the course of the negotiations

A Settlement Offer, and a Co-Author Excluded

On September 6, OpenAI representatives contacted Buckmaster and Alpöge with a proposal for a joint announcement of the results. According to Buckmaster's account, the company presented two ways to resolve the dispute. In the first, the researchers' team would publish their results first, and OpenAI would present its own Navier-Stokes proof a day later. In the second, Buckmaster would independently prepare a paper describing OpenAI's result, but without Alpöge as a co-author.

The reason given for excluding Alpöge was his employment at Anthropic, a direct competitor of OpenAI in the large language model market. To Buckmaster and the mathematics community, the proposal looks like an attempt to break up the research team and remove from the history of the discovery a person tied to a rival company.

What Happens Next

OpenAI turned down the $1 million prize awarded by the Clay Mathematics Institute for solving the Millennium Prize Problem, even though the proof passed formal verification in the Lean 4 system. The mathematics community, however, is demanding an additional, traditional peer review, independent of the automated formal verification, before the result is considered fully confirmed.

The case is fueling a broader debate over the extent to which corporations developing large language models can indirectly benefit from the queries of researchers using their coding tools, even without a deliberate breach of confidentiality. Critics point out that even anonymized data from prompts can, in practice, give companies a time advantage over independent teams working on the same problems.

For Polish researchers and institutions using tools such as Codex or ChatGPT, the case is a warning sign when working on as-yet-unpublished research results. The dispute also shows how difficult it has become to separate genuine technological progress from the fight over discovery priority and reputation among AI labs.