[wyldfire]

Rather than a natural language code review, having a machine learning model that can spot code that "looks wrong" might be really valuable, especially if it was able to get a "good enough" SNR. Static checkers today can be limited in their applicability. But if you trained a model on { block of code } => "defect class <X>" it could be really powerful. Perhaps seeing examples of applied fixes might be a good way to convince someone whether or not there was really a bug there.

Maybe it could even collaborate with static checkers as a quick screen -- seeing as how some of static checkers today are a compromise between the execution time of the checker and the comprehensiveness of the check.