AI and the Mathematics Revolution: From Erdos Problems to Fields Medal Proofs

An exploration of how artificial intelligence is transforming mathematics, tracing the shift from AI solving school tests to tackling real research problems, the rise of formal proof verification, and the implications for mathematicians’ craft and future discoveries.

AI and the future of mathematical discovery

The World, The Universe and Us examines a seismic shift in mathematics driven by artificial intelligence. What mathematicians once viewed as a slow, cumulative enterprise is now accelerating under AI, with large language models beginning to tackle problems that have long been considered the domain of humans. The episode starts by outlining the sense of an existential moment among mathematicians, who fear AI could outperform human theorem proving in the near future.

From high school tests to research frontiers

Historically, AI surprised specialists by performing well on elite high school competitions such as the International Mathematical Olympiad. This milestone indicated AI’s potential, but the real leap came when AI started addressing Erdos problems—famous questions proposed by Paul Erdos that mathematicians spend years on during their PhDs. The newest iterations of large language models began to crack these problems, moving AI from classroom challenges toward genuine research relevance.

The 10 problem test and dramatic improvements

A US group of mathematicians created a rigorous 10-problem test to gauge AI capability in mathematics. Early expectations were pessimistic, but AI systems from OpenAI and Google DeepMind solved five to six problems. This shift suggested that AI is moving from assistive tools to potentially handling core mathematical tasks that mathematicians once believed required years of training.

AI scaffolds, problem solving, and verification

The episode explains scaffolds around large language models, designed to help AI explore multiple solution paths and verify them. While these systems are not directly accessible as ChatGPT, researchers have observed that the solutions resemble what a capable PhD student might propose. The discussion also covers formalization, where proofs are translated into computer-checkable language, enabling automated verification through Lean and mathlib. A Silicon Valley startup, Gauss, reportedly formalized a Fields Medal winning proof, demonstrating the power of AI-assisted formalization at scale.

Consequences for the practice of mathematics

The hosts and guests discuss how AI could transform not only what gets solved but how mathematicians work. Terence Tao suggests that AI could enable testing thousands of methods across many problems, akin to randomized trials in medicine, potentially accelerating mathematical progress. However, there is concern about the erosion of deep understanding that comes from struggling with proofs. The AlphaGo example from a decade ago is revisited to illustrate how AI’s creative and intuitive moves—Move 37—changed expectations about what machines can do, foreshadowing similar shifts in mathematics.

The horizon and cautions

The episode closes by considering how to balance practical gains with preserving mathematical intuition and rigor. If AI can perform proof verification and explore new approaches, it could push mathematics to a new era of methodical, large-scale experimentation. Yet the mathematicians warn against outsourcing the fundamental learning process, emphasizing the importance of maintaining human engagement with challenging problems. The go-to-go example remains the Go-game story, reminding listeners that the most profound tools often come with transformative, sometimes unsettling, implications for human creativity.