A big AI math run still needs proof
Jack Clark describes a researcher’s account of using AI tools on formal math work. The reported run used substantial computing resources and produced many lines of Lean code. The cost and scale come from that account; code volume does not prove the math is correct.
What the source reports
Jack Clark writes about a math project in Import AI. He relays a researcher’s account of using AI to help turn math into code that can be checked. The account puts token costs near $135,000. It also reports about 600,000 lines of Lean code. Both figures come from Resolution’s own account. Today’s digest did not check them on its own. Many lines of code do not prove the math is right. This one case does not show what most teams could afford. The AI Agency digest says teams should count the cost of human checks as well as computing. That is the digest’s advice, not a measured gain from this project.
Original source
- Title
- Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means
- Author
- Jack Clark
- Publication
- Import AI
- Date
- Monday, October 26, 2026