Fields Medalist Jacob Tsimerman Is Joining OpenAI

Source: Theories of Everything | Published: 2026-08-10T15:48:51Z

AI is producing math faster than humans can absorb it. Newly minted Fields Medal winner Jacob Tsimerman says he's mourning the loss of his old way of working.

This conversation was recorded ten days after that ceremony. It was also his first podcast interview.

A Mathematician Says He Is "Grieving"

Tsimerman's way of working goes like this: pick a problem, spend a decade slowly mapping its edges, have that "aha" moment on some random afternoon, then spend a few more years writing it all up cleanly. The two bodies of work that won him the Fields Medal took roughly twelve or thirteen years for the André-Oort conjecture and about five for the Griffiths conjecture. Both were products of slow, immersive labor.

He says he is genuinely "grieving."

That rhythm is coming to an end. Not because mathematics will disappear, but because this mode of work — years spent wandering at the frontier, feeling your way in the dark — is being dismantled by AI.

Why Mathematics Gets Hit First

Most people assume AI will start with simple work and leave something as hard as mathematics for last. Tsimerman says he used to think that too. The intuition is wrong.

The reason: mathematics is a closed system. You need no lab, no materials — just your mind. In that kind of environment, AI makes unusually fast progress.

Two days before this recording, OpenAI had published solutions to ten mathematical conjectures. Tsimerman recognized one — a result on non-abelian group construction, a significant advance — but was unfamiliar with the rest. Not because they were too hard, but because mathematics is vast: he knows nearly everything in his own few subfields, but a paper just slightly outside his corner of number theory can take weeks to digest.

AI is now producing results faster than humans can absorb them.

If No One Can Understand a Proof, Is It Still a Proof?

One question that came up: if the Riemann Hypothesis gets proved, but the proof is a billion lines of technically correct, unreadable code — what do we do?

Tsimerman's answer has two layers.

The first: from a purely logical standpoint, yes, it counts. A train runs whether or not the passengers understand the engine. He expects that in the not-too-distant future, proofs will exist that no single person can fully comprehend, and mathematics will build on top of them anyway.

The second: proofs serve another role in the mathematical community — they transmit understanding and build shared consensus. From that angle, the question gets much harder.

But he notes this is not the most pressing problem right now. The more immediate challenge is that AI is starting to produce proofs — ones that are, in principle, understandable — faster than humans can process them. He offers a comparison: if someone in his subfield solves a major problem, they can walk Tsimerman through the key ideas in twenty minutes, because they share a common language and set of concepts. But if the pace multiplies by ten or a hundred, that sense of a shared map collapses.

Mathematicians are sparse sentinels scattered across a continent. AI will expand that continent far faster than humans can migrate across it.

Worth noting: even Tsimerman's own work rests on theorems he couldn't reconstruct from scratch. He mentions using a result from logic and model theory called "o-minimality" — he knows the gist, knows how to apply it, but reproving it himself would take years. This reliance on black boxes has always existed in mathematics. AI just makes it more pervasive and more visible.

The mathematician Charles Fefferman once told Tsimerman, while he was still an undergraduate: your knowledge will always exist in three tiers — things you understand completely, things you have a rough sense of, and things you've only heard the name of. Those three tiers will always exist. As long as the first tier keeps growing, you're fine.

Chess Survived Deep Blue. Mathematics Is Different

The natural analogy comes up: chess survived Deep Blue just fine. Tsimerman thinks the analogy breaks down at a fundamental level.

Chess was always about human performance — how good you are, how brilliant your move was. It's inherently competitive, and humans being in the arena is the whole point. Mathematics is different. Mathematics exists because we have puzzles we want to solve, mysteries we want to understand. That goal does not depend on a human being the one who does it.

So he believes AI will genuinely reshape mathematics in a way it didn't reshape chess. The chess world was mainly disrupted in how players prepare; once you sit at the board, it's still your skill on the line. Mathematics will change more thoroughly.

When Solving a Conjecture Becomes as Easy as Posting a Tweet

Physicist Tobias Osborne described on LinkedIn what he'd observed: in quantum information theory, someone was spending around ten thousand dollars a week deploying a fleet of AI agents to sweep through open conjectures, solving roughly one to five per day.

He called this "inflation" — like Spain shipping American silver back to Europe: Spain didn't get richer, things just got more expensive. When everyone knows your paper was generated in an afternoon with a single prompt, no one is impressed anymore. Osborne said that one weekend last November, he used AI to add a Lean proof and produce a master's-level result, then immediately felt nothing about the contribution.

Osborne's prediction: people who can formulate good definitions and theoretical frameworks will become more valuable, because agent swarms can build on top of those definitions.

Tsimerman's response was a direct challenge: why can't AI also do that? You say value will shift to X, but you only need to ask one question — will AI be able to do X soon too? He said he left the same comment on the post.

Mathematician Tim Gowers recently wrote an essay Tsimerman admires, responding to the Leiden Declaration on Artificial Intelligence and Mathematics — a signed statement on AI's impact on mathematics. Gowers's core question: where is the stable equilibrium? For any role you think humans should play, you can ask: why can't AI do that too?

Tsimerman was not invited to sign the Leiden Declaration on Artificial Intelligence and Mathematics. He says that even if invited, he would have declined — he disagrees with too much of it. But he thinks the sections on potential risks and safeguards hit the right notes.

He Built a Typing Game for His Nieces

Tsimerman says the first time he truly felt the leap in AI capability was when he decided to build a typing game for his nieces.

As a kid, he learned to type through a game where aliens fell from the sky, each bearing a letter, and you had to hit the key before the alien landed — faster and faster, more and more letters. He thought it was far more effective than today's polished but pedagogically weak typing software.

He opened an AI coding assistant, described the game he wanted, and waited.

What came back surprised him: fifteen files, a complete game framework. He was stunned — he'd expected something like this to take until the third week at the earliest.

"I wasn't used to getting results that fast. I thought this wouldn't happen until week three."

He worked with the AI for a few hours, steering it away from some questionable design choices. The finished game has graphics and sound effects, and it's now on GitHub — he admits he doesn't really know how to use GitHub, but the AI kept assuring him it had organized the repository for him.

"It keeps saying, don't worry, I've tidied up your GitHub. And I just say, great, thanks."

The experience taught him what programmers mean when they talk about the shift in their role: you're now a manager, not an executor. You need to know what you want, evaluate the AI's choices, and steer — not read every line of JavaScript yourself. He says this rhythm of "receive results, decide what's next" is worth practicing now, before it becomes indispensable. You don't want to be starting from scratch when that moment arrives.

Advice for Young People Who Want to Do Mathematics

Someone asked: if your daughter were nineteen and wanted to do pure mathematics, what would you tell her?

He starts by saying there's nothing wrong with pursuing what you love — if you genuinely enjoy mathematics, go learn it, go do it. The pleasure of mathematics won't disappear because of AI. Most people who engage with mathematics were never going to be professional mathematicians anyway: there are engineers, there are kids who do math competitions, there are people who find it aesthetically satisfying. That isn't going away.

But there's a practical message too: hedge. Don't bury your head in the sand and assume the noise about AI will eventually settle back to normal. Long before AI arrived, he told every graduate student the same hard fact: far more people enter mathematics PhD programs than ultimately get tenure. That was already the reality. AI will make it sharper.

He encourages people to engage with AI systems early — not because today's tools are the final form of tomorrow's tools, but because some of the experience transfers. You get used to the speed. You learn to make choices amid an avalanche of results. You develop a sense of what you want and what you don't. That's a sensibility worth building early.

Why He Left Now

Tsimerman says he always assumed he'd spend his whole career in academia. A mathematics professorship is a comfortable existence: once you have tenure, you almost never have to retire, you get better with experience, you keep working with young people, you keep turning over the problems you love.

He started taking AI seriously in 2016. When ChatGPT came out, he called a friend in the AI community and said: "You were right. I was wrong. This is the most important thing happening in our lifetimes."

If he didn't believe we were entering a period of violent transition, he says, he would almost certainly have stayed.

He runs a website — mathforsafety.org — with resources for people with mathematical backgrounds who want to enter AI safety. He doesn't think the field needs everyone flooding into one place. What it needs is a healthy ecosystem: companies playing their role, governments playing theirs, independent nonprofits with their own position, academia too. Distributed is better.

He doesn't frame his departure as a heroic choice. Many people, he notes, understood that AI safety was one of the most important problems well before he did. He just took longer to come around. He simply believes that if we don't actively shape this transition, we'll inherit an outcome we had no hand in designing.

"We are living through a wild time. This is not going to be business as usual."