OpenAI Solved a 90-Year Math Problem in 88 Hours. The Backlash Is Just Beginning.

The proof is out. Nobody knows yet if it belongs to OpenAI.

On September 8, OpenAI announced that an unreleased internal model had solved the Navier-Stokes existence and smoothness problem in 88 hours. The claim landed on a news cycle already saturated with AI demonstrations and quietly upended one of the most visible frontiers in pure mathematics.

The problem is 90 years old. It asks whether the equations that describe how fluids move — water, air, plasma, blood — always produce smooth solutions or whether singularities can form, no matter how gently you start. It is one of the seven Millennium Prize Problems, each carrying a $1 million award from the Clay Mathematics Institute. Since the list was published in 2000, only one problem has been solved: Grigori Perelman’s proof of the Poincaré conjecture, delivered between 2002 and 2003. The other six remain open.

OpenAI’s answer came fast. The company said a team of roughly 10,000 AI agents worked in parallel, communicating with each other, consuming millions of dollars in compute. It arrived at a result by September 5, working from an internal model more capable than GPT-6 Astra, which the company publicly unveiled just two days earlier. The headline finding, as reported by South Korea’s JoongAng Ilbo: the model produced a proof that smooth fluid velocity can grow without bound over time — a direction of solution the mathematical community had debated for decades.

Speed alone is not the story here. The story is what happens next.

Who wrote this proof, exactly?

Within hours of OpenAI’s announcement, Tristan Buckmaster, a professor of mathematics at New York University, filed a complaint that has not yet been dismissed. Buckmaster said he and a researcher at Anthropic were actively working on the Navier-Stokes problem before OpenAI’s result appeared. He said OpenAI had used unpublished work from his group to reach its conclusion.

“Almost no one else was researching our approach,” Buckmaster said in a statement, according to the Korean report. “It is not the kind of direction you feed into a model and find in a few days.”

OpenAI has not publicly responded to the allegation. The company has also not released the full proof or the codebase behind the 10,000-agent system. What exists today is a press announcement and a set of claims about what the model found — not a peer-reviewed paper, not a verified repository, not the kind of transparent trail that mathematical research normally demands.

This gap is where the real controversy lives.

The proof may be unreadable by design

Even if OpenAI’s result is correct — and that word is doing a lot of heavy lifting right now — the proof may be useless to the very people who need to validate it. A Navier-Stokes proof produced by thousands of autonomous agents does not look like a human-written argument. It does not follow the chain of lemmas, the elegant reductions, the “let us introduce a clever test function” moments that make a paper legible to other mathematicians.

Terence Tao, the UCLA professor and Fields Medalist, put it plainly in a social media post on September 8. “When new techniques or tools appear, it becomes easier to solve problems, but it also becomes harder to see where promising discoveries are hiding,” Tao wrote. “AI has already made it easy to solve problems across many areas of mathematics, which makes it hard to filter for the truly promising research directions.”

Tao’s concern cuts both ways. AI may be getting better at generating candidate proofs than humans can evaluate them. The field of formal verification — using systems like Lean to mechanically check a mathematical argument line by line — is one path toward resolution, but Lean requires a human to first translate the proof into a formal language. If the raw output from OpenAI’s system cannot be parsed into that format, the verification bottleneck simply shifts rather than disappears.

The alpha-go echo, louder this time

The Korean press’s comparison to the 2016 AlphaGo match is not entirely metaphorical. In Go, the boundary between human mastery and machine dominance was crossed visibly, in a single match broadcast to a global audience. The mathematics story is playing out differently. There is no match to watch. There is only an announcement and a growing number of unresolved questions about authorship, method, and access.

But the structural parallel is real. AlphaGo taught the world that a domain once believed to require human intuition could be cracked by a system trained on data and computation. The Navier-Stokes result, if validated, extends that lesson into an even older and more abstract territory — not a game with rules, but a branch of pure mathematics that has resisted formalization for nearly a century.

Kim Kyu-dong, a professor at UNIST who specializes in applied mathematics, told the JoongAng Ilbo that the shock this time exceeds AlphaGo’s. “The problem seemed unsolvable,” Kim said. “About ten years ago, people began to say it might be solvable this way. In the last two to three years, mathematicians had been experimenting with AI models on various possibilities.”

Kim also flagged the attribution problem that will now dominate the field: “Future papers will likely be assumed to involve AI assistance, and the question of how to divide credit between human researchers and AI systems will become central.”

Who funds what, and why it matters for the West

OpenAI’s result did not emerge from a publicly funded university lab. It came from a private company with access to what the Korean report describes as “millions of dollars in computing resources.” The cost of the computation was not disclosed, but the scale suggests a number that most national research councils could not replicate in a single grant cycle.

This has direct implications for how Western institutions evaluate AI research. For decades, the standard model of mathematical breakthrough ran through universities, peer review, and public funding. A proof appears in a journal. Colleagues check it. The community decides. OpenAI’s approach skips the first several steps and goes straight from computation to announcement.

The Fields Medal — often described as the mathematics equivalent of the Nobel Prize — may face a structural problem within a decade. If prizes are awarded for results that no individual human authored, or for proofs no human can read from start to finish, the award’s underlying logic breaks. The Korean report raised this explicitly; Tao’s comment reinforced it.

Terence Tao’s observation carries particular weight here. He has spent his career building bridges between analysis, combinatorics, and the emerging computational methods that now appear capable of competing with human insight. His frustration is not with AI itself but with the opacity of the companies producing it.

“The technology is advancing too quickly, and the fact that AI companies do not reveal how they find answers makes the problem worse,” Tao wrote.

What comes next

Three outcomes are plausible. The first is that OpenAI eventually releases enough material for the community to verify the result, possibly through a Lean-certified formalization. The second is that the proof contains an error hidden inside the scale of its own computation — a possibility every mathematician in the field is quietly considering. The third is that the community refuses to recognize a result it cannot trace back to a human author, regardless of its truth value.

None of these scenarios serve the field well if they play out without coordination. The Korean report captures the mood accurately: this moment feels larger than AlphaGo because AlphaGo operated inside a closed system with clear rules. Mathematics is not a closed system. It is a communal enterprise built on shared standards of evidence, attribution, and citation. An AI can solve a problem. It cannot yet participate in the republic of letters.

What is already clear is that OpenAI has redrawn the map of what machines can attempt — and that the next question will not be whether the proof is correct, but who gets to say it is.