What mathematicians should know about the Lean Theorem Prover: reliability & AI (terrytao.wordpress.com)

The Lean Theorem Prover, a powerful tool for formalizing mathematics, has recently gained prominence within the AI and machine learning community due to significant advancements in its capabilities, particularly through autoformalization. Developed by Leo de Moura at Microsoft in 2013 and now an open-source project, Lean allows for rigorous verification of mathematical proofs using a vast library, mathlib, which contains over 300,000 theorems and has attracted considerable contributions. This year marked a breakthrough for autoformalization, where AI technology is used to convert traditional mathematical proofs into formal proofs automatically. Projects like the autoformalization of Fermat’s Last Theorem generated millions of lines of formal code in mere days, showcasing the changing landscape of mathematical verification. The recent "Summer of Soundness Bugs" in Lean has raised vital discussions around the reliability and integrity of AI-driven proof assistants. Several soundness bugs were detected, which, while alarming, are viewed as positive outcomes of rigorous testing by researchers using AI for cybersecurity. These incidents underscore the importance of human oversight in the verification process and highlight the ongoing challenges in ensuring the reliability of proof assistants. Lean's foundational design based on type theory, combined with the emergence of autoformalization, positions it at the forefront of a transformative evolution in mathematics, promising to make formal proofs more accessible and widespread.