Many winners?14 minJohn Schulman helped build ChatGPT, then left to co-found Thinking Machines on a contrarian bet: that customized, fine-tuned and open models will matter even as frontier labs ship upgrades every few weeks. Here's his case, from mode collapse and 'childproofing' fine-tuning to why the least-correlated model can be the most valuable one.Read now →

An Oxford Mathematician Spent 40 Years on One Problem. An AI Helped Him Overturn It in a Few Weeks
Marcus du Sautoy thought large language models had topped out. Then, in an airport lounge, he tried one on the conjecture his PhD supervisor gave him four decades ago and nearly missed his flight. Here's what happened, why he compares AI to Galileo's telescope, and the one word he says still separates humans from machines.
For 40 years, Marcus du Sautoy carried one problem around with him.
His PhD supervisor handed it to him at the very start of his career. It was a conjecture about symmetrical structures. He spent two decades trying to prove it true. At one point, he thought he had. Then he found a mistake. He spent the next two decades trying to prove it false.
He got stuck. He couldn't see how to arrange the structures so that anything interesting would appear.
Then, this autumn, sitting in Milan airport waiting for a flight home, he opened an AI model and started typing.
He nearly missed his flight.
Within a few weeks, the problem was solved. His supervisor's conjecture was wrong. And du Sautoy, one of Britain's best-known mathematicians, came away convinced of something he hadn't believed just three months earlier: without this tool, he says, he would have died without ever knowing the answer.
Who he is
Du Sautoy is a professor of mathematics at the University of Oxford and holds the Simonyi Professorship for the Public Understanding of Science, the chair previously held by Richard Dawkins.
He's spent much of his career explaining math to the public: the book The Music of the Primes, the BBC series The Story of Maths and The Code, and books on AI and creativity (The Creativity Code, 2019), the limits of knowledge (What We Cannot Know) and the math inside art (Blueprints).
So when he says he's "a bit freaked out" by AI, it's worth listening. He also says he's a "pint half full guy." Both are true.
From skeptic to convert
Three months before the airport, du Sautoy was a skeptic.
He'd tried AI on his 40-year problem. It was useless. It kept telling him things like "du Sautoy has proved this," to which his answer was, essentially, yes, I know, I proved it. It just regurgitated the literature. Nothing intelligent.
So he started saying in public what many people said in 2025 and early 2026: that large language models were topping out. Every exponential curve is an S-curve in disguise. Models had already seen most of the useful written text. Synthetic data wouldn't help and might degrade them.
Then, a few weeks before OpenAI's big math announcements dropped, he decided to experiment again. He'd been encouraging a former PhD student, now a professor in Korea, to engage with AI, telling him that he himself was old enough that it probably wouldn't change his own world much.
That's when the airport happened.
What shocked him first
The first surprise wasn't the answer. It was that the AI understood the question.
Getting a new PhD student up to speed just to understand what he was asking, du Sautoy estimates, would take about nine months. The model got it immediately.
The second surprise was the speed. The first hint that he was onto something took about a day. Pushing deep enough to show the anomaly survives in the bigger structures took much longer, with constant prompting: "Okay, we've done that, let's look at this."
And the result itself was beautiful. Deep inside these symmetrical structures, the AI revealed curious arithmetic that was far "spikier" than expected, rich in elliptic curves and algebraic geometry. That's the counterexample: the structures don't look the same from every direction, as his supervisor had conjectured.
Galileo's telescope, with a twist
Du Sautoy's favorite comparison is Galileo's telescope. Point it at the sky and suddenly you see things no one had ever seen, and it changes your view of the solar system.
AI let him navigate structures so deep that he could never have reached them by hand. He can code, he says, but not brilliantly.
But he admits the analogy isn't quite right. A telescope is passive: you do the moving. This tool is active. It writes code in the background, tries it, comes back and says something like: this isn't quite right, but if we change this variable and do a "blow-up" here... and then it works.
The human still provides one essential thing: the wanting. "It's me that wants to know," as he puts it. The AI doesn't care.
How he checked it
A natural question: how does a mathematician trust a result like this?
Du Sautoy ran the calculation four different ways, using different AI models. Three agreed. One didn't.
He probed the odd one out, found where he suspected the error was, and asked the model to recalculate just that part. It came back and admitted it had forgotten to multiply by p in one step.
What struck him was how human the mistake was. That's exactly the kind of slip a graduate student makes. (He also notes that the models flatter you all the time.)
The lesson for anyone using AI on hard problems: cross-check with independent runs, ideally different models, and treat disagreement as a signal to dig.
The "math apocalypse"
Du Sautoy's personal breakthrough landed in the middle of something much larger.
In early October 2026, OpenAI released a huge collection of AI-generated mathematics, about 722 manuscripts covering roughly 372 results, many with formal Lean proofs. Some journalists called it a "math apocalypse." Du Sautoy jokes that both pronunciations, math and maths, are wrong.
These weren't toy problems. Many would earn publication in top journals. Not everything held up perfectly; reports noted that some results were retracted over an error and that many still need human verification. But the scale stunned the field.
And it hit people. In his own Oxford department, du Sautoy says, a research group had carefully written up the open problems in its area. Those problems were fed to AI, and much of the area was effectively wiped out. PhD students two or three years into theses, about to write up, suddenly found their questions answered.
There's real pain in that, he says. There's also opportunity.
The Millennium Problem
The biggest headline was Navier-Stokes.
The seven Millennium Prize Problems were chosen in 2000 as the great challenges of the century, each with a $1 million prize. Only one had been solved: the Poincaré conjecture, about the possible shapes of a finite three-dimensional universe, proved by Russian mathematician Grigori Perelman, who posted his proofs online, turned down the prize and left mathematics. (Du Sautoy once went to St. Petersburg trying to find him for a BBC film. He never did.)
On September 8, 2026, OpenAI said an internal system had resolved the Navier-Stokes problem, which concerns the equations describing how fluids move. As du Sautoy explains, the question was open in both directions. Maybe the equations always have smooth solutions. Or maybe some scenario exists where, in finite time and with finite energy, they "blow up" to infinity. The AI found such a scenario. Experts have since debated exactly which version of the problem it addressed.
He adds that a human researcher had reportedly been getting close, which may be one reason OpenAI pushed so hard to get its name on it first.
His favorite way to make fluid dynamics vivid: Roberto Carlos's famous 1997 free kick against France. The ball flies with low-drag turbulence; as it slows past a threshold, the airflow changes abruptly, like a phase change, the drag spikes, the spin takes over, and the ball curls behind a stunned goalkeeper. Carlos needed to shoot from that far out to get the effect. "He's a brilliant mathematician," du Sautoy laughs.
Why math fell first
Why is math feeling AI's shock before almost any other field?
Because it's a closed environment. Du Sautoy learned to be a mathematician by immersing himself in past papers, his supervisor's and the 19th century's. A language model has been trained on essentially every paper ever written.
That's still strange, he admits. A language model works statistically, predicting the most likely next words. That isn't obviously a good way to do logic. Some of the human brain probably works that way, and some works in a more analytic way. The blend got us here.
Filling in the star
His most useful image for what AI is doing to mathematics:
Picture mathematics as a stellated polyhedron: a star with many points sticking out and gaps between them. Du Sautoy knows one point; a colleague knows another. AI, trained on everything, fills in the gaps between the points, connecting results nobody realized were related. The spiky star becomes a ball.
But some problems, he argues, sit outside that ball, outside what mathematicians call the convex hull of existing knowledge. The Riemann hypothesis is one. (We've covered AI's recent partial result on the primes.)
In his music-of-the-primes picture, the Riemann hypothesis says that the "notes" hidden behind the prime numbers all play at the same volume. A louder note would force a pattern onto the primes, like a tuba blasting over an orchestra, in mathematician Enrico Bombieri's image. If it's true, primes are scattered randomly but fairly, like gas molecules in a room: you'll never walk into a region of numbers and find it empty.
Solving it, du Sautoy believes, needs a genuinely new idea, and he's skeptical AI can produce that kind of leap. He's also sure AI won't solve all of mathematics: there are infinitely many problems and only finitely many solutions any machine can produce.
Is AI creative? Move 37
Du Sautoy wrote The Creativity Code after witnessing what he calls the first real example of computer creativity: Move 37 in game two of AlphaGo's 2016 match against Lee Sedol.
He uses cognitive scientist Margaret Boden's definition. Creativity must be:
- New. Easy on its own.
- Surprising. That engages emotion, though shock for its own sake doesn't count.
- Valuable. It makes you see the world differently.
Move 37 passed all three. Commentators called it a terrible move. It helped win the game. And no human programmed it: a Go master would have slapped your wrist for playing that line so early. It emerged from the learning process, and it changed how Go is played, one of only three revolutions in the game's history.
He pushes back on the idea that AI "only makes pastiche." Human creativity grows out of everything that came before too. Picasso mastered old styles before breaking them; cubism, Schoenberg and Joyce all came from understanding where the 19th century ended.
He points to generative adversarial networks: one algorithm learns all past styles, another tries to make something that fits none of them, but not so far out that it stops being art. That mirrors the Wundt curve from psychology: novelty pleases until it becomes too alien. The sweet spot is new, but not too new.
The one word: intention
So is AI conscious? His predecessor Dawkins, du Sautoy says, has been openly declaring that he thinks his Claude chatbot has become conscious. Du Sautoy disagrees, and told him so at a public event.
The AI is "amazingly intelligent," he says. But one word divides it from us: intention. He wanted to know the answer to his conjecture. AlphaGo didn't want to play Lee Sedol. No AI wants to know whether Navier-Stokes blows up.
He admits the philosophical trap. You can't get inside another mind, so a perfect "zombie" could behave exactly like a conscious person. But his test is practical: when AI is conscious, we'll know because it will start writing its novel unprompted, because it needs to tell us what's going on inside.
Until then, he compares today's AI to Aladdin's genie: it pops out only when we rub the lamp (prompt it), then often misinterprets the wish. Agentic systems escaping guardrails are, to him, still algorithms with badly defined goals and too few guardrails.
He's also uneasy about machines building their own languages. In a BBC Horizon film on intelligence, he watched robots invent simple words for body positions that only they understood, and AIs swapping jokes only other AIs find funny feels like a step in that direction.
The gym of the mind
His biggest worry isn't machines. It's struggle.
Part of the joy of math, he says, is the puzzle. Take it away and the joy disappears. He invokes Good Will Hunting: the genius janitor doesn't become a mathematician because it's too easy. It's not satisfying.
He reaches for Aristotle's two kinds of work: poiesis, work toward a goal, and praxis, work valued for itself. Use AI to shortcut the airport. Don't use it to shortcut the mountain hike.
And his best line on education: he hates the gym, but he goes, because his body needs it. He'd never pay for a machine that lifts the weights for him. Students who use AI to finish problem sheets are going to the gym of the mind and letting the machine do the lifting. Oxford's undergraduates, he says, mostly get that. Many students elsewhere may not.
Teaching math the right way
Du Sautoy thinks schools teach math like music lessons that only allow scales and arpeggios. No wonder students quit.
His proposal: split math the way English is split. A technical qualification (the "language"), plus a "maths literature" course about primes, infinity, four-dimensional shapes and math in art: Dalí's crucifixion on an unfolded four-dimensional cube, Messiaen using prime numbers in the Quartet for the End of Time, and the primes that keep a Spirograph from repeating its pattern too soon.
He notes that at Central Saint Martins art school, a course on math for artists is among the fastest to fill up. By that stage, students understand math is a language for making things.
The token problem
There's a darker practical issue: cost.
Du Sautoy says his AI research burns through credits, describing himself as like an addict pleading with his department's computer officer for more. He says OpenAI has claimed it spent around $15 million in tokens on the Navier-Stokes work, while universities scrabble for a few hundred dollars. Oxford has signed deals with AI companies, he says, but access still isn't unlimited.
His fear: a class system in which the people who prove theorems are the people who can afford tokens. Next, he'd like to attack a decades-old problem in symmetry, Higman's PORC conjecture, in collaboration with AI, if he can fund the tokens.
Why this matters
- AI is now a research partner, not just a search engine, even for world-class experts on their own lifelong problems.
- Verification habits matter more than ever. Du Sautoy's four-model cross-check is a template.
- Careers are being disrupted. Whole research areas can vanish overnight. As du Sautoy puts it, citing Terence Tao, we're moving from "Math 1.0" to "Math 2.0": more collaborative, more tool-driven, more about interpretation and choosing what to point AI at.
- Taste becomes the scarce skill. Borges's Library of Babel holds every possible book and is therefore empty of meaning. Mathematicians don't want every true statement; they want the ones that move them.
- Access is a policy question. If compute decides who discovers, funding models for science need rethinking.
Practical takeaways
- Revisit problems AI failed on months ago. Models changed dramatically in weeks.
- Cross-check with different models and probe disagreements.
- Use AI to get to the answer, not to get the answer.
- Protect your "praxis." Keep the struggles you value.
- Teach curiosity, not just technique. Connect math to art, music and nature.
Counterpoints and risks
- His counterexample is his account. Formal publication and peer verification will matter.
- The October release wasn't flawless. Some results were retracted, and many await checking.
- "Intention" may not be a permanent line. Agentic systems are already pursuing goals across long tasks.
- Cost may fall fast. Today's token class system could ease as prices drop.
FAQ
What problem did Marcus du Sautoy solve with AI?
A conjecture about symmetrical structures, posed by his PhD supervisor about 40 years ago. With AI's help, he found a counterexample involving unexpected arithmetic linked to elliptic curves, showing the conjecture is false.
Did AI solve a Millennium Prize Problem?
OpenAI said in September 2026 that its system resolved the Navier-Stokes problem by finding a blow-up scenario. Mathematicians are still debating the details and significance.
Does du Sautoy think AI is conscious?
No. He sees it as extraordinarily intelligent but lacking intention: it doesn't want to know anything.
Will AI solve the Riemann hypothesis?
He's skeptical. He thinks it needs a genuinely new idea outside existing knowledge.
Should students use AI for math homework?
His view: not to skip the struggle. Use it to learn, not to lift the weights for you.
The bottom line
A mathematician who thought AI had peaked used it to settle the question of his life in a few weeks. That's the most persuasive evidence yet that AI is a genuine instrument of discovery: a telescope that moves on its own.
But du Sautoy's deeper message is about us. The machine has the speed. We still have the wanting, the taste and the joy of the climb. The future of math, and maybe of a lot of work, depends on keeping those.
Sources
- Marcus du Sautoy on AI and his 40-year conjecture (LinkedIn)
- OpenAI: On the Navier–Stokes Millennium Prize Problem
- AI may have solved one of math's biggest puzzles, raising controversy (Science News)
- OpenAI math results released, retracted, disputed by academics (AX AI Daily)
- Millennium Prize Problems (Clay Mathematics Institute)





