What's this about?
Responding to the observation that output quality varies more now than it did before AI coding, dotey explains the mechanism. In the hand-written era, output was capped — a few hundred lines a day. Now it is thousands or tens of thousands, quality differs enormously, and multiplying by volume widens the spread. Once generation is cheap and execution is no longer the bottleneck, what separates people is technical judgment and conscientiousness. Judgment means knowing what to do, what "good enough" is, and how to verify. Conscientiousness means doing the parts around the generation: understanding the requirement yourself, making the design decisions yourself, reviewing, testing, confirming it works — at which point handing it to a colleague is no different from the hand-written era, because they are reviewing something you already took responsibility for.
What is the failure mode he names?
Doing only the middle step. Skipping the requirements analysis and the design at the front, skipping the review and testing at the back, committing what the model produced. By line count that person's output is astonishing; in practice they are mass-producing debt for everyone else — the review cost, the cost of the traps sprung later, and the maintenance cost all transfer to the team. His closing line is the one to keep: AI cannot take responsibility, so when the code fails, a person still carries it.
Why does this matter to you?
- It is the Chinese-language counterpart to "AI-written code is still your code" on HN the same day, arrived at independently.
- Judgment is partially recoverable — you can use AI to analyse and check repeatedly. Conscientiousness is not; it is a choice.
- The variance framing is more useful than the usual productivity argument, because it predicts that averages will stop describing anything.