Paul Erdős was the kind of mathematician who slept on colleagues' couches, drank coffee like water, and posed problems that the smartest people on earth would spend their careers failing to solve. In 1946, he scribbled down a deceptively simple question about points and distances on a flat surface — what's now known as the unit-distance conjecture. For eighty years, it sat there. Fields medallists circled it. Princeton seminars dissected it. Nobody cracked it.
On 20 May 2026, OpenAI announced that one of their unreleased reasoning models — the kind built not to chat but to think step by step — had quietly disproved it. Not assisted a human. Not suggested an approach. Disproved it, on its own, producing a 125-page proof that nine elite mathematicians, including Princeton's Noga Alon, independently checked and signed off on. Sir Timothy Gowers, himself a Fields medallist, called it a milestone in AI mathematics.
Here's why that one sentence should make you put down your coffee: this is the first time an AI has generated genuinely new mathematical knowledge, rather than reshuffling what humans already knew. And it lands roughly six weeks before OpenAI is expected to file for what could be a near-trillion-dollar IPO.
