ChatGPT killed a 22-year-old math conjecture during a nap
GPT-5.6 Pro found a counterexample to a graph theory conjecture open since ~2003 while its user slept.
A user asked GPT-5.6 Pro to "find a counterexample to an open conjecture of its choosing" and then went to take a nap. He woke up to find the model had disproved WOWII Conjecture 91 — a graph theory problem that had sat open for roughly 22 years.
The kicker: the user, Matty Hempstead, says he doesn't care about graph theory and had never heard of the conjecture until the model told him it had cracked it. The prompting was almost comically minimal — point the model at "do a breakthrough," let it search, verify.
It's part of a run of AI-assisted math results landing in quick succession. Around the same window, GPT-5.6 Pro was also credited with refuting the ~30-year-old Dinitz-Garg-Goemans conjecture via a 7-node counterexample, guided by just 58 words across four prompts.
Why it matters: when a napping non-expert can retire a decades-old conjecture, frontier models are shifting from answering questions to generating genuinely new results.
Sources
Primary: the company, paper or repository
Independent coverage
Written by an AI pipeline from the sources above. Methodology · Report an error
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.