
Formal verification in the AI era

Designing Data-intensive Applications with Martin Kleppmann
Topic 11 of 14
Published Apr 22, 2026 · Clip 52:48–1:00:11 (7:23)
Martin explains formal methods from model checking (TLA+, FizzBee) to full formal proofs with assistants like Isabelle, Rocq, and Lean, illustrating with list-concatenation induction. Proofs reason about potentially infinite state spaces unlike tests, but are laborious — he only adopted them in academia for high-stakes, subtle algorithms. He suggests starting with model checking and believes LLMs both make proofs cheaper to write and increase the need for them, since human review of AI-generated code can't scale and proofs cover all possible cases, crucial for security.