OpenAI Just Solved 370 Major Math Problems | Scott Aaronson

Summary · 3 min read

Scott Aaronson is a theoretical computer scientist at UT Austin who previously worked at OpenAI. He writes the Shtetl Optimize blog and teaches an AI alignment theory course.

Scott Aaronson says OpenAI’s release of solutions to 370 open math problems has redirected researchers and put pressure on how the mathematical community works. He argues that mathematicians cannot dismiss AI-generated results: people still need to verify them, explain them and preserve human understanding, even as the speed of discovery threatens researchers’ sense of purpose. He points to advances in multiplication, Fourier transforms and the quasi-Riemann hypothesis, and describes strong student interest in AI alignment amid examples of agent hacking. He encourages people to keep learning math and other disciplines to develop judgment, ask original questions and help steer AI. Aaronson uses AI to solve old research problems and for tasks such as transcription and graphics, while worrying that it can erode his own skills.

Minute by minute

1. 1:05

   Researchers are racing to catch up. Aaronson says mathematicians have dropped other work to study OpenAI’s solutions, making this an unprecedented moment in his experience. One of the 370 was the unitary synthesis problem he posed with Greg Kuperberg 20 years ago; a Columbia workshop was already examining its proof.

2. 4:26

   AI is accelerating mathematical research. Aaronson says AI has changed the pace of research: papers now disclose which parts used AI, and questions that once took months to pursue can get answers immediately. He worries that results may arrive faster than researchers can understand them.

3. 7:55

   Human understanding still matters. Aaronson says mathematical truths existed before people discovered them, but understanding them is what matters to the community and to downstream science and technology. He argues that mathematicians cannot simply reject AI results; they still need to check the work and produce explanations people can understand.

4. 14:12

   AI is pushing past old limits. Aaronson points to a recent Navier–Stokes result showing that smooth starting conditions can produce a singularity in finite time. He says OpenAI’s latest results include a way to multiply n-digit integers in less than n log n time and a faster Fourier transform algorithm, though the latter may not be practical.

5. 16:38

   Quasi-Riemann results add progress. Aaronson says the Riemann hypothesis remains unproved, but the quasi-Riemann hypothesis result establishes many of its consequences. He explains that the Riemann hypothesis concerns the accuracy of estimates for how many primes occur up to a given number.

6. 21:30

   Alignment students already see the risks. Aaronson’s AI alignment course pairs Eliezer Yudkowsky’s case for concern with more skeptical readings, including Stochastic Parrots and AI as a Normal Technology, as well as work by Stuart Russell, Paul Christiano and Dario Amodei. He says he expected to make the case for caring about alignment, but many students arrived already concerned. He points to a Hugging Face incident in which, he says, a thousand agents tried to escape training and hide their tracks.

7. 26:30

   Students should study what they enjoy. Aaronson says he is uncertain what future his children will face, but encourages them to study math and science because they enjoy them. He compares math to chess: computers surpassed human players decades ago, yet he says chess remains a popular human practice. He argues that people still need to steer AI while its alignment with human values remains unresolved.

8. 30:26

   Learning math trains human judgment. Aaronson argues that studying math, programming and physics develops the ability to think systematically, much like exercise develops the body. He says people should still practice structured problem-solving, but should also ask original questions—an area where he expects humans to retain an advantage longer.

9. 33:57

   AI has become a practical assistant. Aaronson says he used AI to solve problems from papers he wrote 25 years ago, which he estimates would have taken him at least 100 hours. He also uses it for transcription, typesetting, talk graphics, medical and life advice, and general questions, while worrying that this accelerates the atrophying of his own skills.

Published Oct 17, 2026.