here's my AI cybersecurity apocalypse white pill.
I've been fixing bugs in leftover human code non-stop for over a year now. literally cleaning up years of human slop.
at the same time, i've been building fully AI-written complex software which I regularly review with AI the same way that I'm reviewing the human code.
the difference is like day and night. good models don't just find the bugs, they also produce far fewer bugs.
the difference between a big project written exclusively by modern frontier models and one written by a dozen engineers is just incredible.
all those little obscure mistakes that humans overlook are mostly absent. if you take care of your project, the code quality can be great and there's almost no accumulation of weird edge cases that sneak in inevitably over the course of multiple years of working on a code base with a team of humans.
makes me very optimistic about the future.