The week he won math's biggest prize, he read his own field's obituary

July 23, 2026, Pennsylvania Convention Center, Philadelphia. The International Mathematical Union opened the International Congress of Mathematicians — the once-every-four-years gathering that only matters because of what happens in the first hour — and read out four names. Yu Deng of the University of Chicago, for deriving the Boltzmann equation from hard-sphere dynamics. John Pardon of Stony Brook, for symplectic geometry and virtual fundamental cycles. Hong Wang of NYU and IHES, for cracking the three-dimensional Kakeya problem, making her only the third woman to win in the medal's ninety-year history. And Jacob Tsimerman of the University of Toronto, for the André–Oort conjecture. It was the first ICM held in the United States since Berkeley in 1986. The prize itself is famously modest: CA$15,000 and a gold medal with Archimedes on the front.

Here's the thing, though. The medal wasn't the story that came out of Philadelphia that week. Asked what he was working on next, Tsimerman — 38 years old, the youngest full professor in Toronto's math department, the guy who just took down a conjecture that had stood since 1989 — said he was moving to San Francisco to do AI safety research at OpenAI. Hours after being handed the highest honor in his profession, he announced he was stepping away from it. Not because he'd lost interest. Because he thinks the profession itself may not survive the decade in recognizable form.

His argument is uncomfortably simple, and he has not softened it in any of the interviews since. Mathematics is far closer to automation than his colleagues want to admit, and AI will soon do everything mathematicians do, better and faster. He has said publicly that he expects AI to be better at math than mathematicians within about two years. He has already stopped taking new graduate students, and his stated reason is the sentence that has been circulating in math departments ever since: he doesn't want to train students for a career in mathematics that may no longer exist. A Fields medalist voluntarily shutting off his own pipeline of successors is, quietly, the heaviest fact in this entire story.

So let's work through it. What exactly did Tsimerman prove, and why does that background make his pivot more interesting than the usual "professor joins lab" item? Why AI safety specifically, and why inside a frontier lab rather than a university or a nonprofit? What is machine learning professor Luca Ambrogioni actually attacking when he says this is "like hiring Lionel Messi as project manager"? And what does one person's leave of absence signal about the balance of power between universities and the labs that keep hiring out of them?

Who's actually in this story

Start with Tsimerman himself, because the résumé matters here. Born in Kazan, Russia in 1988, he moved to Israel at three and to Toronto at nine. He won two gold medals at the International Mathematical Olympiad, including a perfect score in 2004. He started at the University of Toronto at 16 and finished his bachelor's degree in two years, graduating in 2006. Then Princeton, where he did his PhD under the number theorist Peter Sarnak, finishing in 2011, followed by a Junior Fellowship at Harvard. He returned to Toronto and became the youngest full professor in the department. Along the way: the 2022 New Horizons Prize, the 2023 Ostrowski Prize, the Royal Society of Canada's John L. Synge Award in 2024, and a Royal Society fellowship in 2025. He spent the 2025–26 academic year at the Institute for Advanced Study in Princeton. Per the University of Toronto, he is the second Canadian ever to win a Fields Medal and the first Canada-based recipient.

The official IMU citation reads: "For his contribution in the recasting of o-minimality as a fundamental method of arithmetic and complex algebraic geometry, and his role in the proof of many central conjectures including Griffiths' conjecture on the algebraicity of images of the period maps, and the André-Oort conjecture for Siegel modular varieties." Translating: he took a tool out of mathematical logic — o-minimality, which constrains how badly a geometric set can wiggle — and turned it into a working method for number theory and complex geometry. The André–Oort conjecture, posed by Yves André in 1989 and Frans Oort in 1995, says that when a lot of arithmetically "special" points pile up inside a Shimura variety, that pile-up can't be a coincidence; there has to be hidden geometric structure underneath. Tsimerman combined Jonathan Pila's approach with his own work bounding Galois orbits to settle important special cases in his dissertation, and later closed out the full picture with collaborators including Ananth Shankar. In 2018, with Benjamin Bakker and Yohan Brunebarbe, he built the o-minimal GAGA framework and proved Griffiths' conjecture — work that sits directly adjacent to the Hodge conjecture, one of the Millennium Prize Problems.

OpenAI needs less introduction, but the specific context of this hire does. On May 20, 2026, OpenAI announced that one of its general-purpose reasoning models had disproved the Erdős unit distance conjecture — a problem Paul Erdős posed in 1946 asking how many pairs of n points in the plane can sit exactly distance 1 apart. For eighty years the working assumption was that square-grid arrangements were essentially optimal. The model found an infinite family of configurations built out of Golod–Shafarevich theory and infinite class field towers, achieving a polynomial improvement of order n^(1+δ), with δ pinned at roughly 0.014 after a refinement from Princeton's Will Sawin. And here's the detail that makes this hire land differently: the human-verified, human-digested version of that proof was posted to arXiv (2605.20695) by Noga Alon, Thomas Bloom, Timothy Gowers, Daniel Litt, Will Sawin, Arul Shankar, Victor Wang, Melanie Matchett Wood — and Jacob Tsimerman. He wasn't commenting on AI-generated mathematics from the sidelines. He was one of the nine people who sat down and checked it.

Two more voices frame the debate. Andrew Critch co-authored "A Taxonomy of Omnicidal Futures Involving Artificial Intelligence" with Tsimerman, posted to arXiv on July 12, 2025 (2507.09369). The paper catalogs scenarios in which all or nearly all humans are killed as a result of AI, organized by who — if anyone — bears primary responsibility. The authors are explicit that they're not claiming any of it is inevitable; the stated point is that naming failure modes publicly is how you get institutions to prevent them. On the other side, Luca Ambrogioni, a machine learning professor, posted the line that has followed this story everywhere: "It's like hiring Lionel Messi as project manager." Meanwhile OpenAI's chief research officer Mark Chen posted that he was "incredibly excited to welcome Jacob Tsimerman to OpenAI," adding that "his mathematical talent is clearly extraordinary, but so is the seriousness and depth with which he engages on AI safety." Other OpenAI researchers were reported to have welcomed the move publicly as well, though Chen's post is the one that's directly verifiable.

What actually happened, in order

First correction to the headlines: this is a leave, not a resignation. Some early coverage said he was leaving the University of Toronto. Tsimerman himself put it plainly — he's going on leave, but he's not leaving; he's still mentoring a number of students and plans to stay involved with the Department of Mathematics. He keeps the faculty post, relocates to San Francisco, and has said he has no firm plans beyond the next year. That structure is now the standard template for academia-to-lab moves: unpaid leave rather than departure, university affiliation retained, actual work done at the lab. The university keeps the prestige, the lab gets the person with no long-term commitment, and the researcher keeps a bridge back. It is also, historically, the structure that most often turns into a permanent move without anyone announcing it.

The reasoning behind choosing AI safety, and choosing to do it inside a frontier lab, is worth taking seriously rather than dismissing. Tsimerman has described his own views as quite extreme by most people's standards, saying he thinks AI will become a serious threat to society and to humanity, and he's expressed support for infrastructure that could slow or pause development. The obvious objection writes itself: if you believe that, why go work at one of the companies building the thing? His answer is that the actual safety problems live inside the training runs. Writing papers from outside gives you the published artifacts; being inside gives you model behavior, failure modes, and the specific points where an intervention could land. Layered on top is a distinctly mathematical argument — given how high the stakes are, safety research needs a very high standard of certainty, and the profession trained to produce certainty is the one that does proofs for a living.

His view of what mathematics actually is helps explain the rest. In his Quanta interview he said that in his experience, and he thinks many people's, math is very much a goal-oriented endeavor — the truth-and-beauty framing is what you see zoomed out, but zoomed in, you just want to win. And if that disappears — if the experience of being the first person to crack a hard thing goes away — he doesn't think that's a trivial issue, and he expects a lot of mathematicians simply won't want to do it anymore. That's the crux. He isn't predicting that mathematics stops producing theorems. He's predicting that it stops producing the specific reward that makes humans want to be mathematicians. He has also said AI could accelerate the field by a factor of a hundred. Both things at once: the field speeds up, and the job hollows out.

The timing wasn't coincidental either. Three days before he stood on that stage, on July 20, the mathematician Levent Alpöge posted a short note on X saying he'd found a counterexample to the Jacobian conjecture — open since 1939 — using an Anthropic model. The counterexample is a polynomial map in three variables whose Jacobian determinant is identically −2 everywhere, yet which sends three distinct inputs to the same output, killing the conjecture in every dimension above two. (The two-dimensional case remains open.) Mathematicians around the world verified it within hours; The Conversation and others covered the aftermath. Worth flagging: Anthropic issued no formal announcement, and the model version, prompts, and full transcript haven't been published, so the "how much was the AI" question is genuinely unresolved. Still, sequence it out. May: an OpenAI model breaks a 1946 Erdős conjecture. July 20: an Anthropic model helps break a 1939 conjecture. July 23: a Fields medalist says he's going to work at a lab. That's not a whim. That's a response.

Item Detail Basis
Medalists and date Four 2026 Fields Medals — Yu Deng, John Pardon, Jacob Tsimerman, Hong Wang — announced at the ICM opening ceremony on July 23, 2026 IMU Fields Medals 2026 page; Simons Foundation
Venue ICM 2026, Pennsylvania Convention Center, Philadelphia, July 23–30 IMU ICM 2026 page
Citation Recasting o-minimality as a method in arithmetic and complex algebraic geometry; Griffiths' conjecture; André–Oort for Siegel modular varieties Official IMU citation PDF
Age and nationality 38; second Canadian Fields medalist ever, first based in Canada BetaKit; University of Toronto
Training U of T bachelor's in two years (2006); Princeton PhD 2011 under Peter Sarnak; Harvard Junior Fellow Quanta; University of Toronto
Nature of the move Leave of absence, faculty post retained, relocating to San Francisco, no fixed plans past a year His own words: "I'm going on leave, but I'm not leaving"
Destination and role OpenAI, AI safety research BetaKit (2026-07-31); Mark Chen's welcome post
His forecast AI better at math than mathematicians within roughly two years; potential 100x acceleration of the field Ceremony remarks and interview coverage
Graduate students Stopped taking new ones — won't train students for a career that may not exist Quanta and follow-up reporting
Relevant papers Co-author, "A Taxonomy of Omnicidal Futures" (Jul 2025); co-author of the human verification of the unit distance disproof (May 2026) arXiv 2507.09369; arXiv 2605.20695

Who wins, who quietly loses

OpenAI is the clearest winner, and not for the reason people assume. Whatever the compensation, this isn't a money story — it's a legitimacy story. The standing critique of frontier-lab safety teams is that they're a communications function wearing a lab coat, staffed by people whose incentives are wired to the product roadmap. That critique gets harder to make when the person joining is a co-author on a paper taxonomizing human extinction pathways, who has said on the record that he supports mechanisms to slow AI development. There's a practical benefit too: OpenAI has been leaning hard on mathematical results as proof of capability, and bringing in one of the people who independently verified its most-cited math result internalizes part of that verification loop. The risk cuts the same way, though. Hire someone with genuinely extreme risk views and you also own whatever he says on the way out, if there is a way out.

The University of Toronto and Canadian academia are in an awkward spot. On paper this is a triumph — the country's first home-based Fields medalist, with congratulations from the president and the department chair, who described Tsimerman as someone routinely found at a blackboard in the lounge puzzling over something with a handful of students. But when that same medalist files for leave the same week and boards a plane to San Francisco, the older Canadian narrative reasserts itself: the country produces world-class researchers and then watches them get absorbed south of the border. Geoffrey Hinton doing Google work from Toronto is the same lineage. The department's consolation is Tsimerman's stated intent to keep mentoring and stay connected — a promise whose value can only be assessed a year from now.

Graduate students and early-career mathematicians occupy the most genuinely difficult position. Think about being three years into a number theory PhD and reading that a Fields medalist has stopped admitting students because the career might not exist. It's one person's decision, but as a signal it's enormous, and it will move people. The flip side is that anyone who already pivoted toward AI safety, alignment, or formalization just got their choice validated by the most prestigious possible endorsement. There's now visible on-ramp infrastructure aimed specifically at this audience — for instance Xiaoyu He's "Existential Risk from AI: An Exposition for Mathematicians," published in August 2026, which argues the case in the register mathematicians actually respond to and explicitly frames the future of mathematics inside the broader question of human survival.

The losers are real. First, pure mathematics itself. The person who cracked André–Oort and Griffiths is not spending the next year or more on number theory, and that's not fungible — there is no substitute Tsimerman. This is precisely where Ambrogioni's Messi jab lands. It's not an insult to AI safety work; it's an allocation argument. When you take the scarcest possible talent and put it somewhere other than where its marginal value is highest, you may be making a locally rational hire and a globally wasteful one. Second, academia's negotiating position. Reporting has counted at least 22 professors at top US universities leaving or taking leave for OpenAI, Anthropic, Meta or Google DeepMind in 2026 — treat that specific figure as a single-source tally rather than a hard number, but the direction is not in dispute. If the symbolic peak of that flow is a Fields medalist, it's fair to ask what cards universities have left to play.

We've watched this movie before — the good ending and the bad one

Start with what worked. In 2013 Geoffrey Hinton joined Google Brain when his startup DNNresearch was acquired, and he did not fully sever his University of Toronto ties. That dual-affiliation arrangement became the template for the next decade, and it produced both commercial deep learning and continued academic output. Yann LeCun did a version of the same thing that year, building Facebook AI Research while keeping his NYU professorship. The most emphatic proof point is John Jumper: a researcher inside a corporate lab won the 2024 Nobel Prize in Chemistry for AlphaFold. That single fact demolished the argument that industry labs can't do first-rate basic science. If Tsimerman's move goes well, this is the shape it takes — access to compute and data that no university could assemble, producing results that would be considered real by academic standards.

Now the bad ending, which is more instructive. In 2015 Uber recruited dozens of researchers out of Carnegie Mellon's National Robotics Engineering Center in one coordinated sweep. The lab was effectively hollowed out, and Uber ATG — the unit those people built — was sold off to Aurora in 2020 and ceased to exist as an independent effort. Talent moved, and neither the original research ecosystem nor the new organization survived intact. That's the failure mode universities underrate: everyone assumes a leave is reversible right up until the institutional memory is gone.

The second cautionary case is much closer to home. OpenAI created its Superalignment team in 2023 with a public commitment of 20% of compute. By May 2024, Ilya Sutskever and Jan Leike had both departed and the team was dissolved. Leike said on the way out that safety culture had taken a backseat to shiny products, and he went to Anthropic. The logic "you have to be inside the lab to fix safety" was in the air then too, held by people at least as smart and at least as committed, and it did not hold. Timnit Gebru's exit from Google in 2020 rhymes structurally. So Tsimerman's strategy isn't untested — it's been tested repeatedly, with mixed and sometimes ugly results.

Which means the thing to watch isn't what he publishes. It's what happens the first time he disagrees with the company in a way that matters. Hinton has said that leaving Google in 2023 is what let him speak freely about AI risk. The real experimental result here is whether a Fields medalist can keep an already-loud voice loud from inside a frontier lab, or whether there turns out to be a category of things he can only say after he leaves. His own statement that he has no firm plans past a year might be read as a man who already knows that's the open question.

How the rest of the board responds

Anthropic pre-empted this, whether it meant to or not. Three days before the ICM ceremony, Levent Alpöge — a mathematician working at Anthropic — dropped the Jacobian conjecture counterexample on X with no press release, and the field verified it in hours. One outlet has reported that Alpöge was an undergraduate advisee of Tsimerman's during the latter's Harvard Junior Fellowship; treat that mentor-protégé detail as single-source and unconfirmed. True or not, the strategic contrast is sharp. OpenAI buys authority and safety credibility by hiring the medalist; Anthropic demonstrates capability by having its own mathematician use its own model to break an 87-year-old problem. Add to that the steady stream of reporting about Anthropic absorbing senior researchers from academia and rival labs — the specific ratios floating around come from recruiting-industry tallies, so I'd hold those loosely — and the two companies are running visibly different plays for the same audience.

Google DeepMind's counter runs on a different axis entirely. AlphaProof reached silver-medal standard at the 2024 International Mathematical Olympiad by learning to produce formal proofs through reinforcement learning over millions of auto-formalized problems, and DeepMind has kept pushing that line. The argument underneath it isn't "AI is good at math" — it's "AI-generated math has to be machine-checkable." Look at the two 2026 results through that lens and the weakness shows: the Jacobian counterexample was verified by human eyeballs in an afternoon, and the unit distance disproof needed nine world-class mathematicians to compress and check it. Beautiful, but it does not scale. That's why the Lean and mathlib formalization community matters more than its visibility suggests, and why Terence Tao's ongoing work cataloging AI contributions to Erdős problems — logging what held up, what didn't, and what got formalized — is quietly some of the most important infrastructure in this whole argument. Tsimerman's own demand for "very high certainty" is, whether he frames it that way or not, this camp's language.

Academia's counterplay is genuinely constrained, and it's worth being honest about that. No university is going to outbid a frontier lab on compensation or compute. That leaves roughly three moves. One, institutionalize the leave — make circulation the norm so that departures become rotations rather than losses, which is what Toronto is effectively attempting. Two, specialize in what labs structurally can't do: long-horizon research with no product payoff, verification infrastructure, and teaching. Three, claim verification and interpretation of machine-produced results as a core academic function rather than a chore. That last one is already happening. The unit distance paper is exactly this — a company produced a result, and nine academics turned it into mathematics humans can read and trust. That's not a subordinate role; it's a bottleneck, and bottlenecks have leverage.

There are other players moving too. The 2026 medal class itself reflects a shifting talent map, and there's been reporting on Chinese tech companies expanding scientist programs aimed squarely at top-tier mathematicians. That reporting is thinner and mostly hasn't been confirmed by official announcements from the labs in question, so I won't put weight on specifics. What's not in doubt is the precedent: it has now been publicly demonstrated that a Fields medalist is recruitable. Everyone drafting an offer for the next one just had their ceiling raised.

So what actually changes

For developers and researchers, two things shift. The first is a signal about what AI safety is. When the person who proved André–Oort decides alignment is worth his time, the implicit claim is that there's formal structure in there worth attacking — that this isn't purely a governance-and-vibes field. If your background is proof theory, probability, or computational complexity, a market just opened that didn't obviously want you two years ago. The second is that verification is becoming the valuable skill. As the volume of machine-produced mathematical and technical claims rises, the scarce resource stops being "can you generate the result" and becomes "can you certify it." Proof assistant fluency — Lean, mathlib, and whatever succeeds them — is turning into a genuinely marketable asset, while positions whose entire value proposition is grinding on hard problems alone are the ones under pressure.

For investors and companies, this marks a new layer of the talent war. Until recently, frontier lab recruiting was a fight over machine learning researchers and infrastructure engineers. The front has now extended into adjacent elite disciplines — mathematics, theoretical physics, theoretical computer science — and that implies two things. One, labs have concluded that scaling alone won't get them where they want to go, and they're buying different kinds of thinking. Two, safety and alignment have acquired a real price, driven by regulatory exposure and enterprise trust rather than pure altruism. What we can't yet tell is whether a hire like this actually constrains a product roadmap or just decorates it. Separating the substance from the signaling will take several quarters of watching what ships and what doesn't.

For ordinary users, nothing changes this month. Your chatbot does not get safer or more accurate because a mathematician moved cities. The plausible medium-term effect is narrower and more specific: one of the most persistent practical failures of frontier models is confident wrongness, and people whose professional training is entirely about knowing when you've actually established something tend to keep asking, loudly, how confidence is being quantified and guaranteed. Whether that pressure survives contact with shipping deadlines is a separate question, but the direction is a good one for anyone who has to rely on these systems.

For universities, funders, and policymakers, the homework is heavier. A Fields medalist declining to train successors is a data point that ought to worry anyone responsible for research pipelines. If AI really does surpass working mathematicians in two years, what does a PhD in 2026 mean, how should it be structured, and who pays for it? And this isn't only about mathematics — the notable thing about this case is that the alarm came from inside the profession that everyone assumed would be automated last, raised by one of its very best practitioners. That said, the two-year forecast is one person's prediction, not a consensus, and plenty of mathematicians who were in that same room in Philadelphia think it's badly wrong. Both of those facts should be held at the same time.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Directly, nothing this week. But if you've got a math, stats, or theory background and you're weighing a career move, this is real evidence that AI safety, alignment, and verification have become markets that absorb top-tier talent rather than sideline it. If you're heading into pure research instead, note that a prediction is still just a prediction.

— Will AI really beat mathematicians in two years? That's Tsimerman's forecast, not a field consensus. It isn't baseless — 2026 alone produced the unit distance disproof in May and the Jacobian counterexample in July, both human-verified. But there's a wide gap between "cracked a few famous problems" and "does everything mathematicians do," and calling that gap closed is premature.

— Is OpenAI ahead of Anthropic and DeepMind now? One hire doesn't settle that. OpenAI bought authority and safety credibility, Anthropic showed a staff mathematician breaking an 87-year-old conjecture with its own model, and DeepMind is betting on machine-checkable formal proof. They're competing on three different axes, so ranking them right now doesn't mean much.

Sources

Numbers are as of announcement and may change.