The highest prize in mathematics gets handed out at most once every four years. You have to be under forty to win. And you have to be, without serious argument, among the most gifted mathematicians alive.

Jacob Tsimerman accepted his on July 23 in Philadelphia. Then he used the stage to announce he was leaving mathematics.

Tsimerman, 38, is a University of Toronto number theorist who spent his career extending o-minimality — a branch of mathematical logic applied to geometry — into hard problems in arithmetic. His specific contribution: a central role in proving the André-Oort conjecture, a long-standing hypothesis about the structure of special points in geometric spaces that had resisted resolution for decades. The International Congress of Mathematicians awarded him the Fields Medal — math's highest honor, the second in Canadian history — for it.

The acceptance stage is normally where you describe the math and thank your advisors. Tsimerman said something different.

He announced he would take leave from U of T — keeping his faculty position — and move to San Francisco to work on AI safety at OpenAI. Not capability research. Safety specifically. His reasoning, per multiple accounts of the announcement: capabilities are advancing without him, and safety has far more unfinished work with far fewer people doing it. No firm plan past the first year.

Source spread

• BetaKit — U of T professor Jacob Tsimerman to join OpenAI — academic. Solid primary reporting on the announcement and Tsimerman's stated reasoning.
• BigGo Finance — Mentor and protégé join rival AI camps — builder. Best coverage of the Alpöge/Jacobian Conjecture connection and the mentor-student dynamic with Anthropic and OpenAI.
• Quanta Magazine — Tsimerman wins 2026 Fields Medal — academic. Authoritative on what the André-Oort conjecture actually required to prove.
• Yellow.com — OpenAI hires Fields Medalist who says AI will outdo mathematicians — hype. Contains Tsimerman's sharpest direct quotes about AI and the future of mathematical careers.

The week that was

Three days before the Fields Medal ceremony, something happened that probably clarified the view from where Tsimerman was standing.

Levent Alpöge — a mathematician who did his undergraduate thesis under Tsimerman's supervision — used Anthropic's Claude Fable 5 to find a 216-character counterexample to the Jacobian Conjecture. The Jacobian Conjecture (a problem in algebraic geometry about when certain mathematical maps can be inverted) had been open for 87 years, posed in 1939, surviving every human attempt since.

And then, the week after: OpenAI published ten decade-old math problems solved by its Astra model, each with machine-checkable Lean 4 proofs (a formal verification language where a computer checks every step). Including the first explicit construction of a non-sofic group — a type of mathematical object whose existence had been an open question since Gromov introduced the concept in 1999.

Tsimerman's mentor-protégé split lands in that context. His former student is using AI to do the hard math. He's going to work on whether the AI is safe.

What's real

• Tsimerman's move is specific and argued. He isn't saying capability work is evil. He's saying capabilities are advancing without him, and the safety side hasn't attracted the same density of rigorous talent. That's a resource-allocation claim, not a moral one — and it's harder to dismiss.
• The Alpöge result is not a coincidence or a footnote. An 87-year-old conjecture, resolved with a frontier model, three days before the winner of math's highest prize said he was leaving to make AI safer. The juxtaposition is earned.
• Keeping his U of T faculty post matters. Tsimerman isn't burning the bridge. He's treating this as a serious experiment with an open return path. That's intellectually honest in a way most "I'm pivoting to AI" announcements aren't.
• He becomes, as of July 23, the first active Fields Medalist working on AI alignment. What that produces is worth watching.

What deserves a side-eye

• One Fields Medalist at OpenAI is a notable hire, not a structural fix. The safety side is still vastly under-resourced relative to capability work. Tsimerman's arrival doesn't close the gap.
• "Capabilities advancing fine without me" contains an implication worth sitting with: if the best living mathematician thinks he's not needed to advance AI math capability further, that says something about how far the frontier has already moved. Not a comforting observation.
• OpenAI is simultaneously the company building Astra (solving decade-old math) and the company hiring Tsimerman for safety (containing what Astra might become). Both are real. The tension isn't contradiction — it's just the situation we're in.

What to do about it

• If this is new information: The people who understand AI capability best are not uniformly comfortable with what they're seeing. Tsimerman joins Nobel laureate John Jumper (now at Anthropic) and others redirecting expertise toward safety rather than capability. Notice the pattern.
• If you build with AI: Watch for Tsimerman's first publications from OpenAI. A Fields Medalist working on alignment will produce something structurally different from standard ML safety papers — more formal, more adversarial in its reasoning, harder to hand-wave. It will surface. When it does, read it.
• If you use AI tools regularly: The everyday version of AI safety is checking the output. Alpöge solved an 87-year problem with Claude — which is remarkable. It doesn't mean every Claude answer is correct. The tool is powerful and imperfect simultaneously. That hasn't changed.

Further reading

• BetaKit — U of T professor Jacob Tsimerman to join OpenAI — the main reporting on the announcement
• BigGo Finance — Mentor and protégé join rival AI camps — Alpöge/Jacobian context and the full mentor-student dynamic
• Quanta Magazine — Tsimerman wins 2026 Fields Medal — what the André-Oort conjecture actually required, in depth
• Yellow.com — OpenAI hires Fields Medalist who says AI will outdo mathematicians — Tsimerman's direct quotes and assessment of mathematics vs. AI
• SiliconANGLE — OpenAI Astra solves ten open math problems — the Astra results that arrived the week after the ICM ceremony