AI writes the code — humans do the reviewing, and it's burning them out

A Pydantic engineer argues that LLM-assisted programming has quietly inverted where the effort and the satisfaction live in software work. The dopamine that used to come from solving problems, understanding tricky logic, and watching code compile has been automated away, while the tedious part — reviewing, directing, and course-correcting plausible-but-incoherent machine output — has expanded. Models are smart enough to produce convincing code but not always coherent enough to hold a consistent intent across a complex change, leaving the person to catch failures like ported hooks landing in the wrong file or invented components that don’t exist. The result is a new ‘fatigue of supervision’: work that is simultaneously more productive and less rewarding.

The piece frames this as a broken feedback loop rather than a personal failing, borrowing the idea of a ‘human reward function’ from machine learning. It also names two aggravating dynamics: an intensity trap, where the ease of firing off ‘one more prompt’ pushes people into 2am sessions and parallel agent runs they can’t actually finish because the bottleneck is a single human brain, and a growing isolation, as the collaborative moments of open-source and team programming get replaced by solitary back-and-forth with a model. A maintainer quoted in the article describes waking to thirty AI-generated PRs a day with no human on the other side learning anything.

The author’s broader point is that this discomfort deserves to be treated as an engineering problem in its own right, not brushed off as hype-cycle whining. He draws a parallel to designers’ existential resistance to responsive web design around 2009 — a loss of control that felt threatening at the time but reshaped the craft. The takeaway is less a prescription than a permission slip: the productivity is real, the destabilization is also real, and pretending only the first is happening is a fast path to burnout.