Less than a year ago, a Harvard math professor started an experiment to test the capabilities of AI. They repeatedly challenged the systems to solve research math problems, which people had recently solved but were not published yet. In the latest round of tests AI systems correctly solved all 10 problems. But the models still lacked the ability to clearly show their results and work. Lauren K. Williams (Dwight Parker Robinson Professor of Mathematics and co-organizer of the testing) says “In the last months, AI has changed the ecosystem for research mathematics. Essentially all the media around AI and mathematics focuses on success stories. That makes it hard to get a more nuanced, objective view of the current state of public models. Can they just do everything? If not, where do they fall short?” This shows that AI’s capabilities are still unsure of, but are showing promising results.
Terence Tao, one of the top mathematicians in the world right now, says that for thousands of years, proofs were extremely hard, and that people who found proofs required expertise and also had a bit of luck. However, now, we have these machines (AI) that can get these proofs for us, making it much easier. He says “We are drowning in proofs”. Professor Tao considers the machines as a kind of disruption, as he says that AI is like a helicopter that drops you straight into the waterfall; it skips the process, going right to the answer.
However, AI can still make mistakes. In September, OpenAI announced that its models had solved the Navier–Stokes problem, one of the great unsolved problems. However, this set off a controversy about whether OpenAI had just exploited the unpublished work of people working on the problem. It also left questions about whether the problem really had been solved.
AI models use pattern recognition and trained intuition to create mathematical steps and strategies, a lot like a normal human trying to solve a problem. Then, they test thousands of pathways, testing lemmas (a “helping theorem” to help prove the more complex theorem) until they find some reasoning that holds together. To check the validity, the models interact with verification softwares such as the software Lean.
Examples:
AI made a simplified proof of Fermat’s Last theorem, which says that a^n+b^n cannot equal c^n, where a, b, and c are positive integers, and n>2
OpenAI’s GPT 6 model proved the upper bound for small prime gaps is at most 186, surpassing the old human-proved world record of 212
This means there are infinitely many sets of two consecutive primes that have a positive difference of less than or to 186.
This relates to the twin prime conjecture, saying there are infinitely many sets of two consecutive primes that have a positive difference equal to 2 (AI is still working on proving this)
A total of about 1000 total conjectures proved by AI