Caltech to Host First Hackathon Aimed at Solving Open Math Problems with AI

Caltech is convening what organizers bill as the first hackathon built around research-level mathematics. From October 30 to November 1, roughly 100 teams of top mathematicians will get access to frontier AI models — backed by more than $2 million in compute credits — and 40 hours to attack unsolved conjectures and sketch out new theory. Teams won’t just submit answers; they’ll defend their work in front of established mathematicians who probe whether the humans actually understand the results their models produced.

The event rides a wave of recent AI-assisted breakthroughs in pure math, including the claimed disproof of Erdős’s 80-year-old planar unit-distance conjecture, the construction of the first explicit non-sofic group, and a still-unverified result giving the six-sphere a complex structure. Those wins frame the two questions the organizers say they’re really chasing: how far AI can compress the path from an idea to a peer-reviewed proof, and what a mathematician’s job becomes once machines can crack conjectures faster than people can.

Prizes come in two waves — one awarded on-site for the most promising and clearly explained results, and a second after the wider math community has had time to verify the claims. That staged structure, plus the emphasis on comprehension over raw output and a stated commitment to responsible AI use, signals awareness that AI-generated proofs still need human scrutiny before they count.