Skip to content- Watching for: Independent mathematician review accepting or refuting any of the three; peer review of the Kerger manuscript; whether the problems were genuinely open; whether the Lean formalizations match the claimed AI-produced proofs; reproduction of the harnesses.
Notes
- 2026-08-09: Merged from three follow-ups tracking the same question. 2026-07-10: OpenAI published a PDF proof of the Cycle Double Cover Conjecture attributed to GPT-5.6 Sol Ultra, stated to use 64 subagents in under an hour (cdn.openai.com cdc_proof.pdf, HN 48863490); HN noted the released prompt instructs the model to assume a proof exists. 2026-07-15: Star Fleet (Colin Snyder, starfleetmath.com, HN 48914646) published proposed solutions to open Erdos problems from up to 20 parallel Codex instances emitting Lean 4 proofs, with sorry-checkers, pinned mathlib versions, and transitive axiom audits. 2026-07-18: Phillip Kerger (UC Berkeley) published a Lean 4/mathlib formalization and manuscript closing a derivative-free convex optimization gap, a near-quadratic deterministic lower bound for convex Lipschitz functions from exact function values (github.com/PhillipKerger/zero-order-bounds-lean-verification, HN 48957779); the manuscript does not attribute the proof to an AI, so the GPT-5.6 credit is a community claim. None peer reviewed. Machine-checked Lean is a stronger claim than the prose PDF. Covered 2026-07-11, 2026-07-15, and 2026-07-18.
Stories