LLMs Plus Formal Proofs: Proof Automation Might Actually Be Practical Now
A post titled 'We have proof automation now' argued that LLMs combined with theorem provers like Lean or Coq might finally make formal verification cheap enough for regular software projects. The thread had a useful meta-comment: there is still widespread confusion about what it even means to use a theorem prover in a project, with some crypto companies confusing 'we used an LLM to write proofs' with 'our code is verified.' The cost argument is the real one. Formal methods have always been prohibitively expensive in human time. If LLMs can draft the proof obligations and humans just review them, the economics change.
Mike Acton, who wrote the Data-Oriented Design PDF also on the front page today, has released an LLM skill specifically for data-oriented programming, which is a different but adjacent thread: using LLMs to encode expert programming knowledge that is otherwise hard to transfer.
The through-line: LLMs are being used as a bridge into specialized technical domains (proof writing, performance engineering) that previously required rare expert knowledge. This is a genuine capability expansion, not just productivity improvement.
So what?
If you are building safety-critical systems, financial infrastructure, or cryptographic protocols, proof automation is worth a serious look right now rather than waiting for the technology to mature further. The window where this is a competitive differentiator, before it becomes table stakes, may be relatively short.