Grant Sanderson on AI and the Future of Mathematics

Grant Sanderson, creator of 3Blue1Brown, sat down with Dwarkesh Patel to examine what current AI progress in mathematics actually means. The conversation focused on concrete benchmarks rather than broad claims about artificial general intelligence. As Sanderson noted, "an AI earning IMO gold is just one more cleared benchmark," while Patel highlighted how such wins often mask deeper gaps in open-ended reasoning. Patel added, "The real test isn't solving the problem - it's knowing which problems are worth posing in the first place."

Sanderson has begun work on a new project documenting AI advances in mathematics. He argued that an AI system earning a gold medal at the International Mathematical Olympiad - an annual competition for high-school students featuring six notoriously difficult problems - represents one more cleared benchmark, not evidence that AGI has arrived. The same logic applies to harder targets. Even if models eventually resolve millennium prize problems - the seven unsolved challenges named by the Clay Mathematics Institute with $1 million prizes attached - many distinctly human tasks would still lie beyond automation.

The discussion covered verification timelines that can stretch a century, whether an AI proof of the Riemann hypothesis would be understandable to humans, and the difficulty of applying reinforcement learning environments to real economic work. Sanderson emphasized that discovering hidden connections inside existing literature remains a separate challenge from raw problem solving. Recent examples include systems that solved four of six IMO 2024 problems yet struggled with novel research conjectures lacking clean reward signals - for instance, an AI might correctly prove a known result like the four-color theorem extension but cannot assign value to an open-ended conjecture such as linking number theory to quantum field theory when no immediate “correct answer” reward exists to reinforce the choice.

The core distinction he drew is between clearing isolated goals and replicating the open-ended judgment mathematicians exercise when choosing what to prove next. Benchmarks move forward, yet the surrounding context of why a result matters and how it fits into larger theory still requires human framing. This gap explains why rapid progress on contest problems has not yet translated into equal progress on open research questions that lack a single correct answer.

Industry observers often treat each new benchmark victory as a direct step toward broad automation. Sanderson's view pushes back on that framing by separating measurable performance from the wider set of activities that keep mathematical research alive. The result is a more measured picture of what current systems can and cannot replace in the near term. Listen to the full interview on Dwarkesh Patel's podcast or explore related visualizations at 3blue1brown.com.