OpenAI's internal model solved hundreds of decades-old math problems; 10k agents took 88 hours
kimmonismus · x · 2026-10-07
A breakdown of OpenAI's latest mathematics announcements, which top mathematicians are calling a historic breakthrough.
- The OpenAI repo now holds 722 manuscripts in 372 result families, with hundreds of decades-unsolved problems cracked by the new internal model in a remarkably short time
- The earlier Navier-Stokes Millennium Prize result is attributed to this model: roughly 10,000 concurrent agents ran 88 hours to produce it, plus 17 hours for Lean formalization and verification
- New results cost on average only about 3 hours of ChatGPT Pro thinking each in compute — a huge efficiency jump; training began August 28, and OpenAI says the model is significantly more capable than GPT-6 Astra and still improving
- The author reads this as the start of an intelligence explosion: no wall, no plateau, with OpenAI and Anthropic racing to publish scientific breakthroughs, and the Sora web/app shutdown signaling focus on building an automated AI researcher
More from AGI Musings
- Huawei projects 900 billion active AI agents by 2035 — implying $90T of GDP at just $100 each — robleclerc · 2026-10-07
- Europe's AI gap is compute: infrastructure moves in years while AI moves in months — ingliguori · 2026-10-07
- 'Big Short' Investor Steve Eisman Explains How the AI Bubble Will Burst — 233C · 2026-10-07
- Professor decries universities force-feeding AI assignments to students who just want to read books — Tyler_A_Harper · 2026-10-07
- If AI Can End Civilization, Can't It Also Defend It? A Reddit Debate on AI Offense vs Defense — scottimous · 2026-10-07
- €1M grant to study 'adaptive reframing' — when intelligence means rethinking the problem itself — ValerioCapraro · 2026-10-07