Test, Don't (Just) Verify (alperenkeles.com)

AI is elevating the field of formal verification, making it more accessible and mainstream. Companies focused on AI-assisted mechanical proving are securing funding at billion-dollar valuations, and interest in proof assistants, particularly Lean, is surging. This momentum is driven by AI's ability to tackle challenging problems in competitions and mathematics, attracting the attention of leading researchers like Terry Tao and Martin Kleppmann. The technology addresses two key challenges in formal verification: the lack of formal specifications for most software and the complexity of proof engineering. By leveraging AI-driven programming, developers are incentivized to write more executable specifications, which enhances the verification process. The implications of AI on formal verification extend to automation in software engineering. As AI proves adept at generating accurate proofs, it could lead to the development of reinforcement learning models that improve automated proof generation over time. While formal verification can significantly reduce bugs in software—evidenced by examples like the CompCert C Compiler—the challenge remains to ensure the soundness and completeness of autoformalization processes. Researchers caution that the trusted computing base, a critical aspect of this verification paradigm, needs close scrutiny, as improper formalization can undermine reliability. With AI enhancing both verification and proof writing, the future holds promise for more robust software engineering methodologies, but discerning the balance between verification and testing will be essential as the technology evolves.