Theorem co-founder Rajashree Agrawal breaks formal verification into 3 problems and says AI already solved the one everyone thought was hard: "The core challenge is taking these programs and shoving them into a proof assistant, and then being able to phrase these questions. That's the theorem statement generation problem."... "The second part is generating all the proofs. That one, the AIs solve fully, because they can reason about these programs perfectly." "The third part is checking this proof, which is where you need to spend your CPU cycles, not GPUs. The proof assistant is asymptotically too slow, so even though you wrote the proof, the checking time is so long that you can't actually get the answer." Theoremshow more

@rajashree If the program is compiled, then maybe verify the executable, too. Here's a verified GPT2 executable: More at

@rajashree Breaking verification into these layers makes the challenge much easier to understand.

@rajashree Proof generation scales cleanly with GPU search, but proof checking gets choked by single-threaded CPU kernels running type elaboration on massive proof terms.

They did not kill the mathematician. They moved the hard part into a sentence. Statement and checking are still the job. A generator that grades its own proof is the part you do not want to automate first. Witten is not unemployed. He is the remaining bottleneck you just advertised.

@rajashree Tools like deepMet and Dream RSI make this 🚀🛸

@rajashree That third bottleneck is where I've actually lived. Chasing a floating point mismatch across CPU vendors on an inference migration ate weeks, generating the answer was never the hard step, getting two chips to agree on it was. Verification is where the real clock runs.

Formal verification is the part of AI reliability that almost nobody is talking about in policy circles. If AI can now help prove program correctness, we should be baking verification requirements into high-stakes AI procurement and critical infrastructure. #AIPolicy #AIGovernance