Fields Medalist Jacob Tsimerman Launches MAISI to Bring Mathematical Proofs to AI Safety

Jacob Tsimerman, the Canadian mathematician who just received the Fields Medal, has founded the Mathematical A.I. Safety Institute (MAISI), a new research organization that intends to make AI safety claims provable rather than persuasive. The group's premise is straightforward: instead of testing AI models for bad behavior after the fact, MAISI wants formal mathematical guarantees about what a system can and cannot do. That would put machine-learning assurance on the same footing as modern cryptography, where security rests on proofs rather than on attempts to break a given cipher.

The comparison to cryptography is not decorative. Cryptographers do not demonstrate that a scheme is secure by throwing attacks at it until they run out of ideas; they reduce its security to a hard mathematical problem and prove that breaking the scheme would solve that problem. MAISI's bet is that similar reductions can be built for at least some properties of AI systems, turning vague safety promises into statements that can be checked line by line. Tsimerman, based at the University of Toronto, works in arithmetic geometry and analytic number theory, fields where rigorous proof is the only currency.

The timing reflects a widening split over how to keep advanced AI under control. Most current safety work is empirical: red-teaming, evaluations, interpretability probes, and benchmark suites that sample a model's behavior and extrapolate. Those methods scale with effort but never close the gap between 'we tested a lot' and 'we know.' Formal verification has its own hard limits, notably the difficulty of specifying what 'safe' even means for an open-ended system, and scaling proofs to models with hundreds of billions of parameters remains an open problem.

MAISI is not the first effort to import formal methods into machine learning. Groups working on neural network verification, SMT-based analysis of classifiers, and proof assistants such as Lean have spent years chipping at related problems, mostly on small networks. What is different here is the ambition and the founder's standing: a fresh Fields Medalist can attract mathematicians and funding that a standard lab might not, and can frame the goal as a decades-long program rather than a product feature.

The institute has not yet published a technical roadmap, headcount, or funding figure, so the near-term question is what MAISI actually proves first. A tractable early target might be certifying specific properties of narrowly scoped models, or formalizing the assumptions behind existing safety evaluations. The larger wager is whether mathematics can deliver guarantees fast enough to matter for systems already being deployed, or whether the field will keep relying on testing because proofs arrive too late.