What's this about?
Anthropic's Boris Cherny argued that as models improve, the *kind* of bug they produce has shifted: fewer small mechanical errors (off-by-one mistakes), more high-level problems — bad system design, poor UX, missing broader context. His takeaway: adversarial code review matters more, not less, as models get better.
What exactly is "adversarial code review"?
It's reviewing code by actively trying to break it — hunting for the ways it fails, the cases it forgot, the design that won't scale — rather than just checking it looks plausible. With AI writing more code that *reads* correctly, the reviewer's job shifts from catching typos to catching flawed thinking.
Why it matters
It reframes what human engineers do in an AI-coding world: less line-by-line typo hunting, more judgment about whether the whole approach is right — the part models are still weakest at.