Formal Verification

E247｜对话盛颖：xAI，Infra的浪漫，SGLang，开源，平权与“甄嬛传” adds 盛颖’s practitioner path. Under Clarke Barrett at Stanford, she worked on SMT solver optimization and semantics, then later moved toward AI infrastructure because formal verification felt elegant but too expensive and narrow for many real-world uses.

Formal verification is the use of mathematical proof to show that a program, chip, protocol, or system satisfies a precise specification. In 137. 对洪乐潼的4小时访谈：AI for Math、把数学变成Lean、数学天书中的证明、直觉、被创造与被发现的, Hong Letong / 洪乐潼 presents it as Axiom’s most plausible first market because expensive systems need stronger guarantees than finite test cases can provide.

Key Claims

• Formal verification links AI For Math to software and hardware markets: proofs become valuable when correctness failures are costly.
• The episode separates program, Formal Specification, verification condition, and proof; Axiom is described as focusing mainly on the proof layer.
• AI proof systems could make verification cheaper by writing proof artifacts that existing checkers can validate.
• The bottleneck is not only proof search; humans still have to state the correct property and decide what the system should guarantee.
• This extends AI Coding Verification from tests, reviews, and runtime behavior toward mathematical guarantees.
• Sheng Ying’s source adds a career-level caution: formal rigor can shape excellent systems thinking while still losing to deployment pressure when verification cost and application scope are too high.

Connections

• 盛颖 / Sheng Ying, Clarke Barrett, SMT solver, and SGLang - source-247 path from formal methods to AI infrastructure.
• Axiom, Axiom Prover, Lean Theorem Prover, and Interactive Theorem Proving — source’s formal verification stack.
• Formal Specification, AI Verification, and AI Coding Verification — adjacent verification concepts.
• AI For Science and Mathematical Abundance — downstream domains that could benefit from cheaper proof.