OpenAI's "Proof Dump": 722 preprints claim 63% of 372 open math problems solved
gleech · x · 2026-10-13
A Paradigm 3 report analyzes OpenAI's single-day release of 722 preprints on 372 open math problems — possibly the biggest day in mathematical history:
- Claims: 63% fully solved, 27% special cases; includes 90 of the "top 500" open problems, at least 3 of "Fields Medal level" interest.
- Verification: 42% come with complete Lean certificates, but no one has confirmed the formalized statements match the originals; 3 proofs withdrawn, 14 needed repairs.
- Cost: "3 hours" of ChatGPT Pro compute per result (roughly $100–$1000), ruling out massive agent swarms; only 21% are counterexamples.
- Opacity: prompts and "reasoning traces" published for only 10 results; the 372/4000 solve rate is below the earlier announced 45%.
Five years ago, AI systems struggled with grade-school word problems.
More from AGI Musings
- Claymation won't die to AI: the process matters as much as the product — dylfreed · 2026-10-13
- Philosopher Carissa Véliz on how automatic AI translation erases language culture — CarissaVeliz · 2026-10-13
- Every's Thesis: 2027 conference to explore human work after automation — every · 2026-10-13
- Grady Booch: coding agents merely pattern-match and can't do real architecture — Grady_Booch · 2026-10-13
- AI Doesn't Grant Talent—It Changes the Rate of Collisions That Unlock Potential — r0ck3t23 · 2026-10-13
- Scaling laws still work, but pretraining compute alone can't catch the frontier — ZeroStateReflex · 2026-10-13