The Neuron recently highlighted a remarkable and complicated claim: OpenAI says one of its advanced AI research models has solved the Navier–Stokes Millennium Prize Problem, one of mathematics’ most important unanswered questions. If the proof holds up, it would be a historic achievement. But the announcement has also sparked a dispute over research credit, private data, and what happens when a powerful AI company enters an academic race already underway.
What is the Navier–Stokes problem?
The Navier–Stokes equations describe how fluids such as water and air move. They are used in areas including weather forecasting, aircraft design, and blood-flow research.
For decades, mathematicians have been trying to determine whether these equations always produce orderly solutions when a fluid starts in a smooth, normal state. The Clay Mathematics Institute designated this as one of seven Millennium Prize Problems, each carrying a $1 million prize.
OpenAI says its proof shows that the equations can fail under certain conditions when a fluid is acted upon by a specially designed but mathematically smooth external force.
How mathematicians are reacting
Scientific American reported that Diego Córdoba, one of the mathematicians whose earlier work opened the path used in the new research, captured both the magnitude of OpenAI’s announcement and the uncertainty surrounding it:
“We’re a little bit in shock.”
Luis Silvestre, a University of Chicago mathematician, described the broader reaction:
“We’re all, in the community, discussing the implications of this.”
Those implications extend beyond a single proof. The larger question is what this event means for mathematics as a human vocation.
In a second Scientific American account, Tristan Buckmaster, the NYU mathematician at the center of the dispute, called it “a Deep Blue–Kasparov moment.”
He was referring to IBM’s Deep Blue defeating world chess champion Garry Kasparov in 1997. That victory changed public perceptions of what computers could do and forced the chess world to reconsider the relationship between human and machine intelligence.
Buckmaster believes this mathematical result could represent a similar turning point. AI may no longer be limited to checking calculations or helping researchers refine ideas. With enough coordination and computing power, it may be able to produce major mathematical discoveries.
His recommendation is appropriately measured: “The community needs to have serious and unhurried discussion about where to go from here.”
How the story unfolded
Two different pairs of mathematicians are important to understanding the controversy.
Diego Córdoba and Luis Martínez-Zoroa are mathematicians whose earlier research developed an unusual way of approaching these fluid equations. Their work focused on “forcing,” which means introducing an external force into the mathematical system.
Buckmaster and Alpöge are mathematicians who worked together building on Córdoba and Martínez-Zoroa’s approach and using AI tools from both OpenAI and Anthropic. Tristan Buckmaster is a mathematics professor at New York University. Levent Alpöge is a mathematician employed by Anthropic, although reports say he pursued this project personally rather than on the company’s behalf.
TechCrunch’s reporting helps trace the sequence of events:
- Diego Córdoba and Luis Martínez-Zoroa developed an unusual “forcing” approach, exploiting an external-force term that many mathematicians had treated as peripheral.
- Tristan Buckmaster and Levent Alpöge pursued that approach and obtained a major result for the related Euler equations. The Euler equations are a simpler, frictionless relative of the Navier–Stokes equations, so progress on Euler could point toward a solution for Navier–Stokes.
- Before they published, word of their progress reportedly reached OpenAI.
- In response, OpenAI redirected an enormous agent swarm toward Navier–Stokes and announced a purported full solution days later.
An “agent swarm” is a large collection of AI instances working in parallel. Instead of one chatbot attempting the entire problem, thousands of agents explore different approaches, share useful discoveries, and abandon unpromising paths.
OpenAI says its Navier–Stokes effort involved roughly 10,000 concurrent agents, 2.7 million agent messages, 130 billion output tokens, and 88 hours of search. It then used Lean, software that checks each logical step in a mathematical proof, to formalize the result. Across all the mathematical problems it attacked that week, OpenAI says it used 300 billion output tokens.
Did OpenAI see the researchers’ work?
WIRED reported that Buckmaster and Alpöge used several AI tools during their research, including OpenAI’s Codex. Buckmaster has raised the possibility that their interactions with Codex somehow helped OpenAI’s system identify or reproduce their research direction before they published.
That possibility has not been confirmed. OpenAI categorically denies deliberately accessing the researchers’ Codex prompts or feeding their proof to its agents. But its published statement contains a carefully qualified concession:
“While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.”
This raises an important concern. Researchers may use a lab’s AI as a private intellectual instrument while that same lab develops competing systems and potentially trains them on product interactions. The boundary between “independent discovery” and “learned indirectly from users” then becomes extraordinarily difficult to audit.
Why the rumor alone could have mattered
OpenAI says its researchers and agents did not see Buckmaster and Alpöge’s unpublished work. It does acknowledge, however, that rumors of another team’s progress prompted its intensive effort.
Think about that: OpenAI may not have needed the researchers’ proof, calculations, or private prompts to gain an advantage. The rumor itself reportedly revealed that serious progress was being made through a little-used approach involving forcing. That information may have told OpenAI where to concentrate its resources.
Navier–Stokes offers an enormous range of possible approaches. Learning that one obscure route was producing results would make the search much narrower. OpenAI could then direct thousands of agents and vast amounts of computing power toward that route, giving it a material advantage that an academic research team could not match.
In other words, the controversy is not only about whether OpenAI saw someone else’s answer. It is also about whether learning where others were close to an answer allowed a frontier GenAI model to race ahead.
The dispute over academic credit
Then the story gets more personal. Buckmaster alleges that OpenAI proposed an academic paper about the model’s discovery that would credit him but exclude Alpöge. That claim remains unconfirmed, and OpenAI mathematician Sébastien Bubeck disputes Buckmaster’s account of the conversation.
Considering that Alpöge’s employer is Anthropic, this is a politically charged allegation. As Reuters has reported, Anthropic was founded in 2021 by former OpenAI leaders and employees, including CEO Dario Amodei. It is now one of OpenAI’s closest competitors, with Claude competing directly against ChatGPT. The idea that an Anthropic employee might be excluded from a major academic paper involving OpenAI has pulled the companies’ rivalry directly into what would ordinarily be a discussion about authorship and credit.
The community needs to have serious and unhurried discussion about where to go from here.
The result is a dispute with several unresolved questions: Who identified the promising path? Did private AI interactions play any role? Who should receive credit? And what rules should apply when an AI company can mobilize thousands of agents against work that a small academic team has not yet published?