GPT-6 Astra cracked 10 'open' Erdős problems and improved a 1970s bound
stuffyokodraws · x · 2026-09-09
On an a16z podcast, OpenAI researchers Mark Sellke and Mehtaab Sawhney discuss AI and mathematics.
Key points:
- Sawhney pasted an Erdős problem still listed as open into GPT-5; five minutes later the model surfaced a paper that had solved it. The exercise eventually uncovered published solutions to 10 more problems thought to be open.
- GPT-6 Astra went further: told to 'go have fun' with a high-dimensional sphere-packing problem, it improved a bound that had stood since the 1970s.
- Even as AI math improves exponentially, mysteries like P vs NP may remain out of reach; recorded before the latest Navier–Stokes plot twist.
More from Models
- GPT-6 Astra vs. Fable 5.1: One Codes Better, One Has Better Taste — PawelHuryn · 2026-09-09
- OpenAI's NS formalization: 10k agents, 88 hours, ~130B output tokens — nrehiew_ · 2026-09-09
- Researcher says alleged Phi model behavior examples failed to replicate — BlancheMinerva · 2026-09-09
- Jerry Liu jokes the OpenAI codename sequence points to 'GPT-7 Galaxy Brain' — lateinteraction · 2026-09-09
- Vending-Bench sees biggest jump ever: GPT-6 Astra dethrones Claude Fable 5.1 ethically — sandersted · 2026-09-09
- Dev open-sources local prompt optimizer on DSPy+Ollama, boosting llama3.2 3B from 39% to 89% with GEPA — MortisAndTen · 2026-09-09