From Math Olympiads to Navier-Stokes: How Fast Is AI Progressing? with Greg Burnham

In May of this year, an AI system solved a problem about how many points can fit in a plane all one unit distance apart from each other—a question the prolific mathematician Paul Erdős had posed decades earlier, and which had resisted significant human effort ever since. That solution was a warm-up act. Soon after, an AI solved one of the seven Millennium Prize Problems, the Navier-Stokes equations, a feat that would have been a career-defining achievement for any human mathematician just a few years ago. Greg Burnham, who leads capabilities research at Epoch AI, studies this trajectory for a living, and his job has become harder precisely because AI has gotten so good so fast.

From School to Work: The Phase Shift in AI Evaluation

Burnham argues that AI capabilities have crossed a fundamental threshold. Up through roughly 2024, the standard way to test AI was with human exams—grade-school math word problems (GSM8K), high-school math competitions, medical licensing tests. AI systems aced those. Now, the frontier has moved from "exam" to "on the job." The problems AI is solving are ones that professional mathematicians have tried and failed to solve. This shift creates a measurement problem: how do you benchmark a system when you don't know the right answer yourself?