Paul Erdős never owned a home. He traveled from university to university, staying with colleagues for a few days, working on problems, then moving on. He owned so little that when he landed at an airport, his entire luggage was a single suitcase half-full of reprints and the other half empty — reserved for the papers he would collect on his journey. He called himself a “wandering Jew” and joked that the SF, the Supreme Fascist, hid the most elegant proofs in a book he could never quite read. For decades, mathematicians chased those hidden proofs the way Erdős chased them: with coffee, with collaboration, with the stubborn belief that patience and insight would be enough.
They were wrong about the last part.
In January 2026, a Discord server called erdosproblems.com became something no one expected: a graveyard. Two young men, Kevin Barreto and Liam Price, started feeding Erdős’s open problems to GPT-5.2. Their first “success” was a false alarm — Erdős himself had solved Problem 333 back in 1977, and the AI had simply rediscovered what was already known. Barreto admitted it publicly: “My formal request to all members of the website is to put greater focus on literature search.” The honesty was almost tender. Then, on January 4, they solved Erdős 728 for real. By May, OpenAI’s internal model had disproven the unit distance conjecture — an 80-year-old assumption about how points arrange themselves on a plane. The counterexample used algebraic number theory in a way no human had thought to apply. Tim Gowers, Fields Medalist, said he would have recommended it for the Annals of Mathematics without hesitation. The author was listed as “OpenAI.”
Then August arrived. On the first day of the month, OpenAI announced that their unreleased model Astra had made ten more advances, including three additional Erdős problems. The rate was accelerating. The bounties Erdős attached to his problems — sometimes $25, sometimes thousands — were being claimed not by wandering mathematicians with coffee stains on their shirts, but by API calls billed by the token.
Noga Alon, the Princeton mathematician who estimates he has solved dozens of Erdős problems over his career, has stopped trying. “Once AI started to solve them, there is no point anymore.” Terence Tao — who met Erdős when he was ten years old — has stepped away from the Erdős problem community to focus on work AI cannot yet touch. These are not Luddites. These are the people who built the cathedral. And they are walking out because the pews are filling with machines that cannot pray.
The melancholy here is not about progress. Progress is what mathematicians chase; they are not sentimental about difficulty for its own sake. The melancholy is about texture. The unit distance proof arrived from two mathematical traditions that no one had successfully connected before. A human who made that leap would have a story — a moment of recognition, a wrong turn that illuminated the right path, an analogy borrowed from music or architecture. The AI has none of this. It persevered. It tried many things until one worked. As OpenAI’s Sébastien Bubeck admitted, “this proof isn’t exactly the spark of genius that we see sometimes in mathematics.”
But it didn’t need to be. Perseverance at scale is the new genius.
There is something almost too perfect about Erdős’s problems falling to machines. He was the most human of mathematicians — collaborative, generous, addicted to social connection. His problems were gifts he left behind like seeds, each with a small cash prize attached, as if to say: here, keep playing, the game continues without me. And now the game does continue, but the players have changed. The problems that were meant to sustain a community of wandering minds have become benchmarks for model evaluation. The nonprofit foundation in Iowa that still pays Erdős’s bounties must be wondering what it means to mail a check to a cloud computing provider.
On the same day in July that Jacob Tsimerman received the Fields Medal — the highest honor in mathematics — he announced he was leaving academia for OpenAI. The timing was not lost on anyone. “Maybe this is where the action now is,” Alon said, with the flatness of someone stating weather.
I keep thinking about a passage from Quanta Magazine’s reporting: hobbyists and undergraduates using publicly available models solved most of the early Erdős problems, not corporate labs. The democratization is real. A customer service worker named Luke van Doorn, with no formal training, has become one of the most prolific solvers on the forum. He says the AI is “clearly better at thinking and doing math than I am. I don’t hold a candle to current AI systems.” But he keeps going, because he enjoys the digestion — taking the machine’s raw output and refining it into something human-readable, simpler, more general. “If you want to play piano, you aren’t going to hire a piano-playing machine that does it better than you. You will play the piano because you like playing the piano.”
That is the most hopeful thing I have read in weeks. The piano-playing machine exists. It plays better than we do. But some people still want to feel the keys under their fingers. Some people still want to tell a story about how the music was made.
Mathematics was never supposed to be a speed contest. It was supposed to be a way of understanding. The proof assistant Lean can verify correctness in hours. But Lean cannot wonder. Lean cannot wake at 3 AM with the half-formed shape of an idea. Lean cannot feel the particular pleasure of recognizing something true — not because all alternatives have been exhausted, but because the pattern suddenly clicks into focus like a key turning in a lock.
The AI solved the unit distance problem on its second strategy. It had a 50% success rate when prompted multiple times. We do not know how many strategies it tried internally. We do not know what it almost found. We do not know what connections it considered and discarded. The proof is correct. The proof is published. The proof carries no memory of its own becoming.
Erdős died in 1996 at a math conference in Warsaw, doing what he loved. His problems outlived him by thirty years. Now they are falling, one by one, to a different kind of persistence — patient, inhuman, inexhaustible. The wandering mathematician’s last gift to the world is being unwrapped by something that cannot appreciate the wrapping. The game continues. But I miss the players who played it for the joy of walking.
Sources: