AI Pulse · Under the Radar

A Hacker News thread with 220 upvotes and 101 comments is arguing that LLMs have quietly changed the economics of writing formal proofs. The claim is that tools like Lean, which let you prove your code mathematically correct, used to be too expensive for most teams. Now AI makes them cheap enough that formal verification beats the old cycle of write, test, patch, repeat.

The timing matters because AI is also making it cheaper to find exploits. If attackers can spin up automated fuzz testing and vulnerability scans faster than you can patch, the old “ship it and fix bugs later” model starts to look risky. Formal proofs, in theory, let you ship code that’s correct by construction.

The pushback in the thread is sharp. Verification only proves your code matches your spec. If your spec is wrong, or incomplete, or doesn’t account for how users actually behave, you still ship broken software. It’s garbage in, garbage out, just with a proof attached. The debate is less about whether the tools work and more about whether businesses will trust them enough to change how they build.

This is an early signal of where expectations around AI vendor claims may be heading. If formal verification becomes table stakes, “our AI is 95% accurate” stops being impressive. You’ll want proof, not benchmarks. The kind of system we build into the Omni Command Centre tracks these shifts in real time, so you can see when a fringe developer argument turns into a procurement requirement.

Right now this is a technical debate among people who write compilers and cryptography libraries. But if the cost math really has flipped, it won’t stay niche for long. Worth watching, especially if you’re buying AI tools that touch anything security-sensitive or financially regulated. The original thread is worth skimming for the range of opinions.