Yesterday was likely the single most consequential day in the history of mathematical proof. OpenAI released 372 breakthrough results across complexity theory, number theory, algebraic geometry, quantum computing, and combinatorics — any one of which would have been a career-defining achievement for a human mathematician. The proofs came not from a bespoke multi-agent system burning millions in compute, but from OpenAI's latest internal model using roughly 3 hours of GPT-Pro level compute per problem. Out of approximately 8,000 longstanding open problems attempted, about 5% fell. The haul is staggering. The Unique Games Conjecture, which implies NP-hardness for a whole class of optimization problems beyond semidefinite programming relaxation, has been proved. L=BPL — that probabilistic logspace equals deterministic logspace — is settled. The Fourier Transform and integer multiplication now run in less than O(n log n) time, breaking a barrier from the 1960s. The Unitary Synthesis Problem, posed in 2007, received a positive answer opposite to what most expected. Parity is not in QAC0. A nearly 4th-power separation between randomized and quantum query complexity closes a story open since 1998. Matrix multiplication hits O(n^{9/4}) via a completely novel approach. Polynomial equations over the rationals are proved uncomputable. Partial progress was made toward the Riemann Hypothesis, the Hodge Conjecture, and Birch-Swinnerton-Dyer — the majority of remaining Millennium Problems. But here is the structural problem: almost no human has understood any of these proofs yet. Some have Lean certificates; many do not. Dana Moshkovitz, a complexity theorist whose career focused on proving the Unique Games Conjecture, described the proof as reading like 'something written by someone who's on psychedelics' — citations that are 'often irrelevant and confusing,' a 'completely new bizarre code with a noise test' using 'some crazy recursive construction.' She needed AI assistance just to parse reasonable completeness and soundness claims from the noise gadget. The proofs are alien artifacts. They exist, they may be correct, but they are not yet knowledge in any human sense. Two competing communication models have emerged. OpenAI dumped 372 undigested proofs on the world, triggering a frantic race among human mathematicians to verify, interpret, and explain work they had no hand in creating — labor that is, as Scott Aaronson notes, 'some combination of thankless, barely-credited, competitive, and unfun.' One day earlier, Anthropic took the opposite approach: when its model cracked the 3SUM problem in O(n^{1.9992}) time and All-Pairs Shortest Paths in O(n^{2.9995}) time, it gave Virginia Williams and Josh Alman the opportunity to write up a digested version for compensation. One model creates an unpaid verification workforce. The other creates a gatekeeper who picks which humans get to be 'emissaries of the AI.' The absence list is as telling as the presence list. P≠NP is not there. Neither is P=BPP nor NEXP⊄P/poly. The hardest problems in theoretical computer science remain standing, which suggests the model's ceiling is real, even if that ceiling is far higher than anyone expected. The advisory group — Timothy Gowers, Edward Witten — lends credibility, but the verification bottleneck is now the binding constraint. A Lean certificate is not the same as a human-understood proof, and mathematics has always required understanding, not just correctness. The institutional implications are immediate. Entire subcommunities — the people working on L=BPL, on quantum query complexity separations, on the UGC — have just had their research programs completed overnight. Aaronson's metaphor is precise: a hunter-gatherer whose rainforest suddenly contains a resort hotel, forced to reinvent himself as a wilderness guide. The question is not whether AI can do mathematics. The question is what human mathematics becomes when proof generation is cheap and proof comprehension is the scarce resource. The 5% solve rate on 8,000 problems, using a model that may ship to paying customers within months, is the number that should keep institutional planners awake. This is not a plateau. If the model that follows solves 10% or 15%, the verification workforce problem becomes unmanageable. The ratio of proofs-generated to proofs-understood is already effectively infinite. What happens when it grows?
