Anthropic said that Claude generated mathematical proofs of correctness for existing code for 123 open-source libraries. The amount of generated proof code was about 180 million lines.
During the work, the model identified errors that, according to The Decoder's description, years of testing did not detect. Anthropic aims to make AI-based formal verification as commonplace as standard testing.
Practical significance depends on how independently these results are confirmed and how applicable the approach is to other projects. The presented package contains only a synopsis of the publication, not its full text or separate proofs.
editorial commentary
Why it matters
Likely consequence — increased interest in AI tools for formal verification of code if the results are confirmed. The next observable signal will be independent evaluation of the proofs and the reaction of library developers. Substantial uncertainty is tied to the absence of primary materials in the package and details about the detected errors.