AI slop math? 17M-line Lean RH proof compiles, but nobody understands it
gerardsans · x · 2026-11-21
Anthropic mathematician Levent Alpöge released a claimed RH/GRH proof built with Claude: a four-page note, 17 million lines of Lean, no readable paper — openly admitting Claude outpaced him on the technical core. Gerard Sans argues independent checks only verify the code compiles, not that the formalization states the original problem or that the presenter understands it: "The claim is public. The understanding is not." The quoted reply adds that AI shows little conceptual novelty in fields he knows, suggesting some "hard" math was hard only because it was manually tedious.
More from AGI Musings
- Underrated AI use cases: custom accounting, negotiation sims, state capacity — curious_vii · 2026-11-21
- The speed limit for AI science is the cost of verification, field by field — professor_ajay · 2026-11-21
- Comparative advantage won't save you: horses had it too, and still lost to AI — paraschopra · 2026-11-21
- Soviet planning was just a compute/latency problem, plus humans RLHF'd to flatter Gosplan — l4rz · 2026-11-21
- Blogger's AI timeline: math in 2026, physics 2027, biology 2028, aging 2030 — rand_longevity · 2026-11-21
- Why does a Lean-verified proof get more hate than a contested journal proof? — IAmTimNguyen · 2026-11-21