Skip to content- Watching for: A formal writeup or peer-reviewed paper for the Jacobian counterexample and whether the field accepts it; independent mathematician review accepting or refuting the CDC, Erdos, and Kerger claims; whether the Lean formalizations match the claimed AI-produced proofs; clarification of the model's role versus the mathematician in each case.
Notes
- 2026-08-16: Merged from two follow-ups tracking one thread. Strongest datapoint is machine-checkable: 2026-07-19 Levent Alpoge posted a counterexample to the Jacobian Conjecture (open since 1939, Smale #16), a polynomial map C^3 to C^3 with constant Jacobian determinant -2 that is not injective, credited to Anthropic's Claude; HN commenters verified it in Sage and SymPy, and 2026-07-21 Terence Tao published an independent geometric reconstruction via symmetric powers plus resultant normalization, presenting it as established mathematics and disclosing he used a chatbot to check calculations (2026-07-22 he shared the transcript, HN 49010345). Weaker prose claims from the same run: 2026-07-10 OpenAI PDF proof of the Cycle Double Cover Conjecture credited to GPT-5.6 Sol Ultra (HN 48863490, released prompt instructs the model to assume a proof exists); 2026-07-15 Star Fleet proposed solutions to open Erdos problems from parallel Codex instances emitting Lean 4 (starfleetmath.com, HN 48914646); 2026-07-18 Phillip Kerger's Lean 4 formalization of a derivative-free convex optimization lower bound, which does not itself attribute the proof to an AI (HN 48957779). None peer reviewed. Machine-checked Lean and the verifiable counterexample are stronger claims than the prose PDFs. Covered 2026-07-11, 07-15, 07-18, 07-20, 07-22, 07-23.
Stories