LLMs still produce bugs, but those bugs are different than what they used to be.

Boris Cherny argues that reviewing AI-written code increasingly requires attention to system design, usability and missing context. In his account, these problems have become more prominent relative to small implementation errors.

He recommends adversarial review: deliberately trying to expose weaknesses and edge cases instead of accepting the first working implementation. That recommendation follows from his diagnosis, because locally correct code can still fit the wider application poorly. The post offers an experience-based review strategy, without measuring a general change in bug rates.