OpenAI model cracks 90 of math's top 500 open problems in one 722-paper dump
Don't Worry About the Vase (Zvi) · rss · 2026-10-09
Zvi's deep dive on OpenAI's math release:
Scale
- 722 manuscripts (372 families) dumped on GitHub, covering 90 of the top 500 open math problems per Proof of Atlas; presumably the same model behind the Navier-Stokes proof, mostly on a single prompt, 3 hours compute per solution from 4,000 attempts.
Highlights
- Quasi-Riemann: a zero-free strip plus no Siegel zeros — Kontorovich: a human would get an instant Fields Medal.
- Matrix multiplication exponent down to 2.25 (previous record 2.37); integer multiplication barrier broken; new bounds on Pi's irrationality; Unique Games, Hodge/Birch, Hilbert's 10th, Hadwiger coloring.
- Sweeping quantum information results (2D area law, Haldane gap, etc.).
Reactions
- Some call Oct 6, 2026 the most significant day in mathematical history; one ranking found 81% of the biggest discoveries of the last three years came from this dump.
- Kevin Bryan: should be front-page news worldwide; priority #1 for labs is doing this for medicine, oncology, batteries.
- Yet mainstream media largely ignored it; Seth Burn likens it to Sputnik's slow initial coverage.
- roon's counterpoint: in a punctuated exponential, every local maximum looks historic.
More from AGI Musings
- Why companies push autonomous AI: investor appeal plus a legal liability shield — JFPuget · 2026-10-09
- Garry Tan amplifies warning: tech job market instability makes adaptability the key skill — garrytan · 2026-10-09
- AI-assisted review finds code sharing rare in clinical prediction model research — EricTopol · 2026-10-09
- EU productivity now 20% below US: Andrew McAfee urges Brussels to build a startup ecosystem — amcafee · 2026-10-09
- Agents are having their metaverse moment — but this one sticks, says Sokolin — LexSokolin · 2026-10-09
- Matt Walsh flip-flops in two days: AI takes everyone's jobs — or everyone works more? — SydSteyerhart · 2026-10-09