Why the Legendary Erdős Problems Are Falling to AI

Something odd happened in math circles this year: the hottest new results aren't coming from tenured professors at Ivy League departments. They're coming from a Cambridge undergrad, a customer-service rep in Belgium, and an unreleased OpenAI model called Astra. All of them are chasing down problems posed by Paul Erdős, a chain-smoking, amphetamine-fueled Hungarian who died in 1996 and never owned a permanent address.

Erdős spent his life handing out cash bounties, sometimes $10, sometimes thousands, to whoever solved the puzzles he scribbled in letters and papers. Thomas Bloom, a number theorist at the University of Manchester, got tired of losing track of which ones were still open, so in 2023 he built erdosproblems.com mostly for himself. He used ChatGPT to write the site's code, which felt novel at the time. Nobody expected AI to eventually help solve the math itself.

That changed fast. By May 2026, an internal OpenAI model had found a counterexample to Erdős's 1946 unit-distance conjecture, a result human mathematicians later refined but still credited as genuinely new. Weeks later, related techniques cracked other longstanding problems. Then in August, Astra knocked out ten more advances, three of them Erdős problems outright. Noga Alon, who's personally solved dozens of these problems over decades, says the tools are reshaping how research actually gets done, not just speeding it up.

What's stranger is who's doing the solving. Kevin Barreto and Liam Price, two twenty-somethings who met on a Discord server, learned to prompt GPT-5.2 by essentially lying to it about how hard a problem was. Their first supposed breakthrough on Christmas morning turned out to be a problem Erdős himself had already solved in 1977 — an embarrassing but telling stumble. Weeks later they landed a real one, Erdős 728, using a formal-verification tool called Aristotle to double-check the logic. Price, who admits he can't fully verify the proofs himself, built a manual process of feeding one chatbot's answer into another chatbot to check it, essentially hand-rolling what AI labs now call a scaffold.

Bloom is thrilled and worried in equal measure. He's watched a genuinely open, democratic community form around his site, where a Belgian customer-service worker can end up co-authoring a proof with Terence Tao. But he's also seeing 200-page papers where no human, including the person listed as author, has actually read the proof end to end. The math is getting done faster than anyone can check it, and nobody's quite sure yet what that costs.