How to train your program verifier

The research duo behind the a3 framework dropped a spicy bomb: an auto‑generated verifier for Python, a3‑python, that scanned the mega‑popular requests library (183 functions), proved 179 “safe,” and found four real bugs. Cue the comment section turning into an arena. Fans celebrated that a bot finally caught those 3AM crash gremlins—like a string split that blows up without a “/” and a pickle‑unpickling “None” mishap—while skeptics rolled their eyes at the grandiose setup: name‑dropping Fields Medal legend Voevodsky and touting “geometry‑of‑verification.” One top reply called it “Z3 meets ChatGPT with a hero cape”, another called it “a glorified linter in fancy math cosplay.”

Python folks were split. Some shouted “Finally, a babysitter for dynamic chaos,” others warned it’s the fox verifying the henhouse—AI helping build the verifier that checks code AI could write. Mathematicians nitpicked the Hilbert/“Stellensatz” flex, joking “If your theorem hunt ends in catching split('/') bugs, touch grass.” SREs posted pager memes and begged for this on production code, while maintainers fretted about bot‑generated issue spam. The spiciest hot take: measuring “distance from safety” is either genius or pure vibes. Meanwhile, the rest of us hit a3‑python, whisper “please find my bugs,” and hope it doesn’t find us.