AI Solves Millennium Math Problems Leaving Mathematicians Behind

At the Heidelberg Laureate Forum in September 2026, the corridors buzzed with a single topic. It was not the usual gossip about which famous researcher had arrived or which elegant proof someone had recently constructed. According to attendees, every overheard conversation, every accidental debate, centered on how OpenAI, Anthropic, and Google were steamrolling through mathematics. [1] The mathematicians had gathered to discuss their discipline. Instead, they found themselves discussing the machines that were rapidly making parts of their discipline unrecognizable.

The irony was not lost on anyone present. Mathematics, with its step-by-step logical reasoning and objectively verifiable answers, had long been considered the purest test of human intelligence. That purity made it the perfect target for AI companies seeking to demonstrate what their systems could do. What followed over the summer of 2026 was not a gradual encroachment but a rout. The tech giants moved from struggling with everyday research-level problems to solving a raft of teasers posed by the prolific Hungarian mathematician Paul Erdős, then verifying the proof of Fermat’s Last Theorem, and finally, most dramatically, OpenAI announcing it had solved the Navier-Stokes existence and smoothness problem.

Ailsa Robertson, a young researcher at the University of Amsterdam, captured the mood with precision. She observed that AI and LLMs set the math community on fire over summer. [1]

The Prize That Changed Everything

The Navier-Stokes existence and smoothness problem belongs to a select group of seven extremely difficult Millennium Prize Problems posed in 2000 by the Clay Mathematics Institute. Each carries a million-dollar reward. Only one, the Poincaré conjecture, had been solved by humans before this year. OpenAI’s claim to have cracked Navier-Stokes represents a watershed moment for automated reasoning, if the proof holds up to scrutiny.

Fields Medalist Jacob Tsimerman of the University of Toronto acknowledged the achievement during a press conference at the Forum. He noted that the community had seen rapid increases in capabilities, even faster than many people, including himself, expected. [1] When asked whether AI was doing something genuinely new, Tsimerman admitted he did not know the details, but added that solving Navier-Stokes felt pretty definitive.

The announcement did not emerge from a vacuum. The system made important progress on a related problem but failed to crack the main mission. With tech giants applying the full might of their most advanced unreleased models, many mathematicians now see it as inevitable that at least some of the remaining Millennium Problems will fall soon.

What struck me was not merely the speed of these advances but their purpose. The tech companies were not solving these problems because the solutions would advance human understanding of fluid dynamics or number theory. They were solving them as benchmarks, as demonstrations of raw capability. Fields Medalist Peter Scholze of the University of Bonn put it bluntly during a panel discussion at the Forum. He characterized the efforts as solving difficult mathematical problems as benchmarks, as a kind of PR stunt.

The distinction matters. A proof that emerges from a machine optimized for benchmark performance is not the same as a proof that emerges from a human mind grappling with a problem for years, developing intuition, building connections, and ultimately producing not just an answer but a new way of thinking. The AI systems produce answers. What they do not produce, at least not reliably, is understanding.

The Norms That Tech Companies Ignore

The controversy surrounding OpenAI’s Navier-Stokes announcement revealed a deeper tension. For mathematicians Michael Harris of Columbia University and Geordie Williamson of the University of Sydney, the problems facing the field stem from the way tech giants fail to adhere to the norms and values deeply instilled in the mathematics community.

Williamson did not mince words. He stated that OpenAI had behaved extremely poorly and that the community should acknowledge this. The allegation stemmed from an interaction between Tristan Buckmaster of New York University, who was making significant progress on Navier-Stokes with Levent Alpöge of Anthropic, and OpenAI representatives. Buckmaster claimed the correspondence was coercive, censorious, and even threatening. Harris, who had a front-row seat to the controversy, said he trusted Buckmaster’s account of the interaction and did his part in promoting the narrative. Harris also noted that on social media and in traditional media, most reports were consistent with his takeaway: they depicted OpenAI as bullying and disrupting disciplinary norms.

Beyond questions of sportsmanship, Williamson raised a more fundamental concern. The mathematical community values understanding above all else. But the field measures progress through unsolved problems, and according to Williamson, these two measurements are very, very quickly becoming uncorrelated. AI solutions might provide an answer, but they usually do not develop new methodology that is understandable or useful. This disconnect forces the community to re-evaluate things like how it assesses people, who gets jobs, and how it educates the next generation. Williamson called this a big challenge..

The problem is not that AI produces wrong answers. The problem is that it produces right answers without producing comprehension. A mathematical proof that no human can follow is, in a meaningful sense, not mathematics at all. It is a result. And results, divorced from understanding, cannot be built upon, cannot inspire new questions, cannot train the next generation of thinkers.

The Human Cost of Machine Speed

While the tech giants tear up the rule book in their battle for supremacy, ordinary human mathematicians bear the consequences. The unbridled access to powerful AI technologies is affecting how mathematicians work across the world, and not always for the better.

AI Solves Millennium Math Problems Leaving Mathematicians Behind (Image 1) — AI-generated image

Mita Ramabulana, a young mathematician at the University of Cape Town, found himself in an uncomfortable position. Two of the ten advances in mathematics that OpenAI announced in August 2026 overlapped with his own work. The experience left him disappointed. He explained that you spend time thinking about these things, and these days you do not know whether someone is just going to plug in a problem that you care about in some LLM and solve it. He worries that unscrupulous researchers are using LLMs to scoop others or gain professional advantage.

The anxiety is not paranoia. It is a rational response to a changed landscape. When a graduate student can feed a problem into a commercial AI system and receive a solution in minutes, the years of careful thought that traditionally went into mathematical research become a liability rather than an asset. The researcher who takes time to understand a problem deeply will be beaten to publication by the researcher who simply asks the machine.

For Ailsa Robertson, currently studying for a PhD in quantum-safe cryptography, the problem is even more acute. She has seen all of her mathematics colleagues turn to using LLMs intensively in their research. Many are maxing out their Pro subscriptions. Some are even spending thousands of Euros on additional tokens. Robertson noted with concern that these are PhD students who do not have thousands of Euros.

The financial pressure compounds the professional pressure. Robertson described colleagues who feel coerced by the tech giants, particularly after OpenAI announced in July 2026 that it was giving away 100,000 free licenses to its frontier models for researchers in academia. The giveaway looks generous on the surface. But it also functions as a trap. Once researchers become dependent on these tools, once their workflow is built around them, the free licenses will expire or the terms will change. And there are other colleagues who feel they simply have no choice. Robertson articulated their dilemma: if you do not work at the rate at which you could work with LLMs, then you will be behind your peers who will be applying for the same jobs as you.

This is part of the reason why Robertson’s PhD has become a lot less mathematics-heavy and more focused on the societal implications of transitioning to a quantum-safe ecosystem. Her reasoning is stark: she does not want to be in a career where you are verifying LLM output. The statement captures the transformation of mathematics from a creative discipline to a quality-control operation. The mathematician becomes a validator, checking the machine’s work, rather than a thinker, generating new ideas.

The Measurement Problem

The crisis in mathematics is not merely about jobs or publication pressure. It is about what the field values and how it measures those values. For centuries, mathematics has operated on an implicit bargain: the community rewards understanding, and understanding produces results. The two were intertwined. A mathematician who truly understood a problem would eventually solve it. A solution that emerged from genuine understanding would illuminate the surrounding territory, suggesting new questions and new approaches.

AI breaks this bargain. It produces results without understanding. It solves problems without illuminating them. And because the field has historically used problem-solving as a proxy for understanding, it now finds itself unable to distinguish between the two. Williamson’s observation that the measurements are becoming uncorrelated strikes at the heart of the matter. If the community cannot tell the difference between a solution that comes from comprehension and one that comes from computation, how can it evaluate its members? How can it decide who deserves a job, a grant, a promotion?

The problem extends beyond professional evaluation. It touches on education. If the purpose of mathematical training is to produce people who can solve problems, and machines can solve problems better, then what is the purpose of mathematical training? The question is not rhetorical. It demands an answer, and the community does not yet have one.

Some mathematicians have responded by retreating to areas where human understanding still matters. Robertson’s pivot to the societal implications of cryptography is one example. Others have doubled down on the collaborative, interpretive aspects of mathematics, the parts that cannot be reduced to problem-solving. But these responses are defensive. They concede that AI has won the battle for problem-solving and seek refuge in territories that machines have not yet conquered.

What the Machines Cannot Do

The story of AI’s summer conquest of mathematics is not, however, a simple tale of human obsolescence. The machines have demonstrated remarkable capabilities, but they have also revealed their limitations. The Riemann hypothesis remains unsolved. The Hodge conjecture remains unsolved. The other Millennium Problems remain unsolved. What the AI systems have done is solve some problems that humans had already solved, verify proofs that humans had already constructed, and make progress on problems that humans had already identified as important.

This is not nothing. The speed and scale of these achievements are genuinely impressive. But it is also not the same as doing mathematics. Mathematics is not a collection of problems to be solved. It is a living tradition of inquiry, a conversation across centuries, a way of understanding the world that is inseparable from the humans who practice it. The AI systems can participate in this conversation, but they cannot sustain it. They can answer questions, but they cannot ask them. They can solve problems, but they cannot wonder why the problems matter.

The mathematicians at Heidelberg understood this, even as they feared for their futures. The conversations in the corridors were not about whether AI would replace them. They were about what would be lost when it did. The understanding that emerges from struggle, the insight that comes from confusion, the satisfaction of finally seeing what was previously hidden — these are not incidental to mathematics. They are mathematics. And they are precisely what the machines cannot provide.

The Moment of Clarity

What the summer of 2026 revealed is not that AI can do mathematics. It is that mathematics, as practiced by humans, has always been about more than getting the right answer. The field has been forced to confront this truth by the arrival of machines that can get the right answer without any of the accompanying human experience. The confrontation is painful, but it is also clarifying.

The mathematicians who gathered in Heidelberg faced a choice. They could compete with the machines on the machines’ terms, racing to solve problems faster and more efficiently. Or they could redefine what mathematics is, reclaiming the aspects of the discipline that machines cannot replicate. The first path leads to obsolescence. The second path leads to an uncertain but potentially richer future.

AI Solves Millennium Math Problems Leaving Mathematicians Behind (Image 2) — AI-generated image

Robertson’s decision to focus on the societal implications of cryptography rather than the mathematics itself is one version of this redefinition. Ramabulana’s disappointment at having his work scooped by AI is another, more painful version. Williamson’s call to re-evaluate how the community assesses people and educates students is a third. None of these responses solves the problem. But they all point toward a recognition that the problem is not going away.

The machines will continue to solve problems. They will continue to improve. They will continue to make mathematicians redundant in the narrow sense of being able to produce correct answers. What they will not do is care about the answers. What they will not do is understand why the questions matter. What they will not do is build a community around the shared pursuit of knowledge.

The mathematicians at Heidelberg knew this. They also knew that knowing it is not enough. The challenge is to build a mathematics that survives the machines, not by competing with them but by being something they cannot be. The challenge is to remember that mathematics was never just about the answers. It was about the asking.

Sources

Spectrum — Quote source (original article)

AI Solves Millennium Math Problems Leaving Mathematicians Behind