10-30 Mathematicians by January, 30-100 by Fall: Fields Medalist Tsimerman Launches MAISI

One month after winning mathematics’ highest honor and announcing he would leave traditional research, Jacob Tsimerman has revealed what he was actually building. On September 8, the University of Toronto professor — set to join OpenAI’s safety team later this month — formally launched the Mathematical AI Safety Institute (MAISI), a new Bay Area research organization devoted to constructing what he calls the “mathematical foundations” of AI safety.

The timing is not incidental. As The Hill notes, the institute’s founding comes just as OpenAI prepares to release Astra, its more capable next-generation model — the very systems whose risks MAISI hopes to define and measure. “AI safety needs more foundational work,” Tsimerman wrote in his announcement post, and his bet is that the missing foundations look less like benchmark dashboards and more like theorems.

What MAISI actually is

According to the New York Times, MAISI will begin its first full semester of research in January 2027 and is aiming to hire 10 to 30 mathematicians for that initial cohort. Tsimerman’s own announcement is more aggressive on scale: 10–30 mathematicians in the Bay Area by January, expanding to 30–100 researchers by September 2027. A more ambitious “special year” program — modeled on the thematic research years that pure-math institutes like Princeton’s IAS have run for decades — is also planned.

The institute is explicitly modeled on Princeton’s Institute for Advanced Study, the storied home of Einstein and von Neumann, where small groups of leading researchers work with minimal teaching obligations. That pedigree matters here: MAISI is not a product team, an academic department, or a standard alignment group inside a frontier lab. It is an attempt to import the IAS model — long horizons, deep specialization, prestige-driven recruitment — directly into AI safety.

The recruiting pitch is aimed squarely at research mathematicians, the population Tsimerman knows best and the one he believes is under-mobilized. Andrew Critch, the UC Berkeley researcher and AI safety veteran, amplified the launch publicly, signaling that the existing alignment community views MAISI as a genuine addition rather than a competitor.

Why a Fields medalist is doing this

Tsimerman’s path to MAISI has been one of the stranger stories of the AI era. In July 2026, the same day he was awarded the Fields Medal in Philadelphia, he announced he was going on leave from the University of Toronto to join OpenAI’s safety team. As The Atlantic reported, the move sent shockwaves through mathematics: a 38-year-old at the peak of the discipline, walking away from the field’s most prestigious career track mid-ascension.

His stated reasoning was unsentimental. Tsimerman takes the view that AI improves upon human mathematics to its logical endpoint — that the trajectory of automated reasoning is powerful enough to reshape what human mathematicians are for. In the last two years before his departure, he had already stopped taking new graduate students, judging that the traditional apprenticeship model of mathematical training no longer made sense as a career preparation. Rather than watch that transformation from Toronto, he chose to work on the side of it that he considers unsolved: making the systems themselves trustworthy.

That framing explains MAISI’s thesis. If AI systems are on track to exceed human performance in domains like mathematics — the discipline with the strictest standards of verification humans have — then the binding constraint on safe deployment is the ability to state precisely what “safe” means and prove things about it. Current safety practice, heavy on red-teaming, preference tuning, and empirical evals, has no equivalent of the rigor that mathematics itself uses to contain fallible human reasoning. MAISI’s wager is that closing that gap is a mathematics problem, and a hard one.

The context that makes it urgent

The launch lands in a week when the boundary between frontier capability and frontier safety is visibly straining. OpenAI’s Astra model has reportedly solved ten open mathematical problems, and separate reporting documented OpenAI pausing reinforcement-learning training runs over critical cybersecurity capability thresholds. When a system starts producing proofs humans cannot readily check, the question of who verifies the verifier stops being philosophical.

It is also a week of high-profile attrition from labs over safety concerns, with an Anthropic researcher’s resignation letter making headlines. Into that turbulence, MAISI offers a constructive channel: rather than leaving to protest, one of the world’s best mathematicians is leaving traditional math to build the institutions he thinks safety lacks.

What to watch

Three things will determine whether MAISI matters beyond its announcement. First, hiring: can it actually attract 30 serious research mathematicians by fall 2027, in a market where frontier labs pay multiples of academic salaries? The IAS model runs on prestige, and Tsimerman’s Fields Medal buys a great deal of it, but converting that into sustained recruitment is unproven. Second, output: the first semester begins January 2027, so the earliest signal of productivity is roughly a year of hiring and seminar-building away. Third, independence: Tsimerman remains bound for OpenAI’s safety team, and MAISI’s funding structure has not been fully detailed in launch coverage. An institute whose founder works inside the lab building Astra will face fair questions about how sharp its critiques can be.

None of those caveats diminish the structural significance. For two years the complaint inside AI safety has been that the field attracts far more machine-learning engineers than people trained to reason about formal systems at the highest level. MAISI is the most credible attempt yet to fix that imbalance — an IAS for the era when the thing being studied at the institute might itself start doing mathematics.