'It is incredible': How AI is transforming mathematics

An Englishman without a degree, Liam Price, solved one of Erdős' famous problems — number 1196 — using ChatGPT. Not a monstrous calculation, but a solution that has even surprised the experts, because ChatGPT has found a new path, outside the human framework. The argument here is that artificial intelligence is already changing mathematics, not only by speeding up calculations, but by arriving at insights that even mathematicians had never seen before. Until now, it was thought that AI could only remix known techniques, but now the surprise is that it is beginning to generate truly original ideas — and to connect different fields with an intuition that almost seems like "thought." Take Jared Duker Lichtman, a mathematician at Stanford: he said it's as if AI had found a new opening in chess, one that no human had ever imagined because they were blocked by conventions and aesthetic taste. Liam Price's story is perfect: from his home in England, together with Kevin Barreto, a Cambridge student, he obtains a solution from ChatGPT that does not start from probability — the classic approach — but stays with the original language of the problem, and yet connects numbers and probability in a novel way. Terence Tao, one of today's most brilliant mathematicians, says that this breakthrough was unthinkable a year ago: at that time, it was thought that language models would never go beyond what they had "read" in books. But now we can see something that is out of the ordinary, so much so that Sébastien Bubeck of OpenAI calls it "incredible." There are those who play it down: Daniel Litt of the University of Toronto says the results are still modest, but warns that skepticism is misplaced — the real mystery is why AI, with its superhuman knowledge and effortless stamina, hasn't already made revolutionary discoveries. Perhaps, says Litt, we fail to understand what the human "secret touch" in mathematics really is, that creativity that no one can explain. Another fact: today's models can produce proofs of 3-4 pages, Google's are aiming for 10, but it will take time to reach 100. Meanwhile, Lauren Williams of Harvard sounds the alarm: more AI means more "slop," that is, proofs that seem perfect but are full of errors — and no one has the time to check them all. The missing perspective is human effort: for centuries, mathematics has been an exercise in endurance and mental solitude. If AI eliminates the effort, the wonder remains, but perhaps a piece of what truly makes a mathematician is also lost. The sentence that sums it all up: AI is not just solving problems, it is changing the rules of the game in mathematics. If you think this story concerns you, you can press I'm In on Lara Notes — it's your way of saying: this idea is now yours. And if tomorrow you tell someone that a guy without a degree used ChatGPT to beat an Erdős problem, you can mark it on Lara Notes: Shared Offline is the way to say that conversation mattered. This was from Nature and it saves you 2 minutes.