Fields Medalist Establishes Independent Institute Dedicated to Mathematical AI Safety

In July, Jacob Tsimerman, a University of Toronto professor, received arguably math’s highest honor: the Fields Medal, which celebrates abstract mathematical truths regardless of practical application.

Upon accepting the award, Dr. Tsimerman made a surprising career shift, disappointing many in the field: he is taking a leave from academia, largely abandoning pure mathematics to focus on AI safety research, and will join OpenAI’s safety division later this month.

On Tuesday, Dr. Tsimerman also announced the creation of the Mathematical A.I. Safety Institute (MAISI), an independent research body unaffiliated with OpenAI. Based in the Bay Area, MAISI will commence its first full research semester in January 2027, initially hiring 10 to 30 mathematicians, with plans to expand significantly in subsequent years.

Through MAISI, Dr. Tsimerman, who will serve as scientific director, conveys a message to academia: “A.I. safety is a meaningful area of research,” he stated in a Toronto interview. “It is a legitimate intellectual pursuit that is both interesting and useful, with real progress to be made.”

Historically, many mathematicians dismissed AI as the domain of software engineers. However, as Dr. Tsimerman argues, “Engineers can only implement existing ideas; we need new ones.” Since AI systems are built on mathematics—linear algebra, calculus, probability, and statistics—he and his peers view certain AI safety challenges as fundamentally mathematical problems.

“We need more mathematical clarity and rigor in the A.I. industry,” said Andrew Critch, MAISI’s executive director and mathematician who studied the algebraic geometry of machine-learning models. “That’s how nuclear energy works: extensive math is done before a power plant ever turns on.”

Mathematically trained researchers see an opportunity for advanced mathematics to address issues like AI system cooperation, verifying responsible behavior and output accuracy, and designing resilient systems against unknown vulnerabilities.

Dr. Tsimerman highlighted zero-knowledge proofs, a cryptographic tool for verifying a system without revealing sensitive details like model parameters. This could underpin protocols for human trust in AI agents.

Computer scientist Shafi Goldwasser, who co-invented zero-knowledge proofs in the 1980s, described the tool as “essentially a way of proving facts without giving your secret sauce away.”

Dr. Goldwasser co-founded a similar institute, the Institute for Responsible Superintelligence (RESI), launched in August in Cambridge, Mass. Its team includes M.I.T. cryptographer Vinod Vaikuntanathan and A.I. safety researcher Adam Tauman Kalai, who recently left OpenAI. They discussed how their methods must scale with A.I. capabilities.

“It’s a giant, growing ecosystem,” said Lionel Levine, a Cornell mathematician, regarding the community focusing on AI risks. Dr. Levine recently created an online repository of open problems to invite mathematicians into this work. “We need a lot of third-party checks and balances,” he added.

Founding MAISI, Dr. Tsimerman also addressed the public, pushing back against the notion that “it’s not reasonable to expect a high level of safety from A.I.” He insists that “it’s reasonable and in fact necessary, and we should be demanding a much, much higher level of safety standard than we’re currently getting.”

Ravi Vakil, a Stanford mathematician and MAISI scientific adviser, characterized the problem as one requiring prioritization from all relevant experts. “We need to approach this with everything at our disposal,” he said. “We need multiple points of attack.”

“Math might be the silver bullet, or it might not be,” Dr. Vakil added. “But maybe we just need a lot of bullets; this is one bullet.”

Source link