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

Boris Cherny says the bugs in AI-written code have changed kind. In his post, there are fewer off-by-one errors, the small slips where a loop or index runs one step too far or stops one step short, and more problems with system design, interface usability and missing broader context. Those are mistakes in how the code fits the rest of the product rather than in a single line, and he sums it up by saying that some kinds of coding have been solved, but not all.

While models keep improving, his answer is adversarial code review: checking the work specifically for ways it breaks rather than to confirm that it runs. He calls it an incredibly powerful way to catch many of these bugs and says it can take a single prompt, giving the example “use a dynamic workflow to adversarial test every edge case in an iOS simulator”, or running Claude's built-in /code-review command, which also comes in variants such as /code-review low and /code-review medium. The post reports his own experience and offers no measurements of how bug types have shifted or how many such bugs the review catches.