Human mathematicians are being outcounterexampled
Points and comments are a snapshot, not live.
AI models are finding counterexamples to major open conjectures in mathematics, with Lean verification.
In 2026, AI models including ChatGPT (Sol), Claude (Fable), and Logos' tool are autoformalizing mathematics in Lean. ChatGPT disproved Erdős' Unit Distance conjecture via a 1.2M-line Lean proof using global class field theory. Claude disproved a 60-year-old Grothendieck question on group schemes (1076 lines). Fable found a counterexample to the 100-year-old Jacobian conjecture (three variables, degree 7). A PhD student used these tools to write 250K lines formalizing a modularity lifting theorem in two weeks. The post concludes that graduate students not paying $200/month for these models are 'crazy.'
What commenters are saying
Commenters split: some argue these are just brute-force searches humans hadn't tried, calling the Grothendieck counterexample 'easy' and 'not interesting.' Others counter that LLMs work across domains and that deniers are in denial. One commenter notes the counterexample prompt was generic, not manually optimized. Another argues counterexamples are unsatisfying-they don't explain why a conjecture seemed true-while others defend their value in redirecting effort and deepening understanding.