The lede
The Mathocalypse
Scott Aaronson
On Tuesday OpenAI released 372 mathematical results produced by an unreleased internal model, among them a proof of Subhash Khot's Unique Games Conjecture with a Lean certificate attached, and Aaronson — who is not easily stampeded — calls it one of the biggest days in mathematical history. The mechanics are the part to read closely: roughly four thousand problems attempted at about three hours of ChatGPT Pro-level compute each, a hit rate on the broad list near 5%, and proofs written badly enough that the mathematicians checking them are using AI to help read them. The list underneath is what will unsettle anyone who has taken a complexity course — L = BPL, integer multiplication below O(n log n), matrix multiplication at O(n^9/4), parity not in QAC0, a 2D area law for gapped Hamiltonians. Aaronson's conclusion is not triumphalism but a change of category: mathematics may now work more like an empirical science with the model as a telescope, and the community's job is to keep digesting output it did not generate. He also notes the one conspicuous absence from the release, which is cryptography, and draws the obvious inference about where labs are probing quietly.