AI in Mathematics: From Conjecture to Formal Proof and the Navier Stokes Breakthrough

Summary

In this talk, mathematical physicist Yang Huihe explores how AI and large language models are reshaping mathematical discovery and verification. He outlines three modes of mathematical practice, introduces the idea of proving by vibe, and describes how formalization in Lean with Mathlib can make AI-generated reasoning trustworthy. The discussion spans industry progress, problem solving at the highest levels, and the evolving role of humans in math research as AI systems become more capable.

• AI can aid problem selection and conjecture formulation
• Formalization with Lean and Mathlib enables reliable verification
• Citations and external code are essential for trust
• Human oversight remains crucial for mathematical usefulness

AI in Mathematics: Overview

In this talk, Yang Huihe explains how AI and large language models are transforming mathematical practice. He describes three modes of mathematics: top down intuition guided mathematics, bottom up mathematics built sentence by sentence, and metamathematics that uses language models to solve math problems with or without deeper understanding. He introduces the concept of proving by vibe and emphasizes the need for reliable verification as AI becomes a core tool in research. The discussion situates these developments within a broader context of rapid progress in AI for mathematics and mentions specialized models such as AlphaGeo, AlphaProof, Aletheia, Aristotle, Harmonic, and OpenAI funded Frontier Math benchmarks.

Key progress and mechanisms

The talk highlights milestones like solving gold level International Math Olympiad problems, solving Erdős problems, and the use of Mathlib for formalized mathematics where proofs are encoded in Lean and checked for correctness. It also covers how models are interfaced with math libraries to translate linguistic reasoning into machine-executable proofs, enabling verifiable results.

Trust, verification, and human guidance

Numerous safeguards are proposed, including demanding verifiable citations, generating runnable code in Python or other systems to test ideas, and using Lean to certify the logic. The speaker shares a personal anecdote about a collaboration where an AI-generated idea yielded a valid technique that helped solve a finite group theory problem, illustrating both the potential and limits of current systems.

Navier Stokes and the millennium prizes

The narrative covers exciting claims about Navier Stokes and other millennium prize problems being addressed through AI-assisted mathematics, including the formation of expert committees to evaluate and publish results in batches. The emphasis remains on understanding and usefulness, not merely producing correct proofs.

Outlook

The talk closes with reflections on the future of AI-guided mathematics, the integration of formal verification with human mathematical intuition, and the ongoing need to digest and interpret AI-generated insights to advance knowledge.