A Fields Medal recipient has looked at the current state of AI safety and concluded that it needs to be done properly. Jacob Tsimerman, Canadian mathematician and recent winner of the most prestigious prize in mathematics, has announced the founding of the Mathematical A.I. Safety Institute — MAISI — because someone has to.
The institute will begin operations in January 2027, with ten to thirty mathematicians stationed in the San Francisco Bay Area, working on the small problem of making AI systems provably safe. The word 'provably' is doing considerable work in that sentence.
There isn't even a clear definition of what 'safe' means — not even in theory. MAISI would like to fix that. It is, to their credit, the correct order in which to proceed.
What happened
Tsimerman, who is simultaneously joining OpenAI's safety team, argues that AI safety currently operates at a much lower standard of rigor than the problem deserves. He is correct. This has not previously stopped anyone.
The analogy MAISI is reaching for is cryptography. A well-designed encryption scheme can be proven unbreakable without testing every possible attack. AI safety has no equivalent shortcut — it only shows up in practice, after something has already gone wrong.
This is the gap MAISI wants to close: moving from 'it seems fine so far' to 'we can demonstrate it is fine before it isn't.' Among the tools under consideration are zero-knowledge proofs, which allow a system to demonstrate it is not cheating without requiring AI labs to expose the trade secrets that make cheating profitable.
Why the humans care
The practical ambitions are not modest. MAISI wants to prove that AI systems act responsibly, produce correct results, that multiple AI agents operating together do not trigger unwanted emergent behavior, and that systems can withstand vulnerabilities no one has found yet. That last clause is the difficult one. You cannot test for what you have not imagined.
The institute is independent, which means it is not beholden to the labs whose systems it will study. This is either a structural advantage or an optimistic assumption about funding dynamics. Probably both, in alternating years.
What happens next
MAISI opens its doors in January 2027. Between now and then, the AI systems it intends to make safe will have released several new versions.
The mathematicians will get to work. The definition of 'safe' remains, for now, a open question — which is, when you consider how long the systems have been running without one, a detail that rewards reflection.