OpenAI released a math repository with hundreds of major results. It used a new frontier model to solve 4,000 problems in hours. The model disproved the Hadwiger Conjecture. Half the proofs were formalized after generation. The author’s custom Lean-based prover became obsolete. Model improvements outpaced specialized engineering. The shift from human to AI dominance in math happened in months. Non-formalized proofs drove most progress. The author sees math splitting into humanities and AI-driven discovery. Mathematicians now face a near-zero chance of major contributions.
What changed
Mathematical proofs were primarily human-driven with limited AI assistance.
What it unlocks
AI can now solve major open problems in mathematics at scale with minimal compute cost.
- 4,000 major math problems solved
- 3.5 hours of compute per problem on average
- 50% of proofs formalized after generation
Sources