John Loeber Reports Unreleased OpenAI Model Disproves Hadwiger Conjecture in Automated Theorem Proving

An unreleased OpenAI frontier model disproved the Hadwiger Conjecture using a probabilistic argument for the existence of a counterexample graph. This result was part of a 'math repo' claiming hundreds of major results, generated after running the model against 4,000 major problems.
The model consumed approximately 3.5 hours of ChatGPT-Pro-Thinking equivalent compute per problem, estimated by John Loeber at $100 per problem.
Navier-Stokes existence and smoothness was resolved a few days after GPT-6 Astra's release, leading Loeber to stop his own efforts on the Hadwiger Conjecture as his work became trivial by comparison.
Loeber argues that sophisticated custom engineering, including a private Lean4 fork and custom harnesses, was rendered obsolete by base model improvements, rendering his ~1 million lines of Lean theorems pointless.
The transition to the 'AI-math era' was sudden rather than gradual, with the human-AI collaboration window lasting only about one summer before the field became entirely AI-dominated.
Approximately 50% of AI-generated results appear to be formalized in Lean after being generated in plaintext, a process Loeber finds 'insane' given the cognitive difficulty of maintaining consistency without formal verification.
The mathematical community is described as being in shock or denial, with the individual probability of making a major contribution dropping from 'slim' to 'zero' while access to discovered truths becomes limitless.
Loeber did not use quantitative evals like Frontiermath to measure progress, relying instead on impression, and notes that undecidable problems are excluded from the prospect of limitless AI answers.

John Loeber Reports Unreleased OpenAI Model Disproves Hadwiger Conjecture in Automated Theorem Proving

Key point

Details

Related