Formal Verification Emerges as AI's Antidote to Code 'Slop'
Latent Space surveys the growing field of AI-assisted formal verification, where mathematical proof checkers like Lean validate model-generated code and proofs rather than relying on tests or LLM judges. The piece profiles &Truth, a new entrant led by cvc5 co-creator Clark Barrett, whose first tool &Atlas maps open-source C code for memory-safety analysis, alongside Harmonic's Aristotle, Axiom Math's AxiomProver and Theorem. Anthropic and Theorem report that LLM-scaled verification is becoming more token-efficient, with Opus 5.5 solving a set of small verification tasks using roughly 40% fewer tokens than Opus 5. The article frames a shift from 'vibe-coding' to 'veri-coding', where developers specify requirements and agents collaborate with verification tools.