The human-in-the-loop is tired
Points and comments are a snapshot, not live.
LLM-assisted programming boosts productivity but creates a new fatigue of supervision and review.
The author, a Pydantic employee, argues that LLM tools have made code generation easier but shifted the bottleneck to human attention, judgment, and review. This creates a 'human reward function problem' where satisfying parts of coding shrink while cognitive load grows. The piece compares the shift to responsive design's disruption of fixed layouts, arguing that core skills like taste and architectural judgment survive while pixel-level control becomes less relevant. New skills like 'pre-mortems' on LLM plans and encoding review history into AGENTS.md files are emerging.
What commenters are saying
A significant camp suspects the article itself is AI-written, citing its style as 'tryhard' and full of 'AI fingerprints.' Multiple commenters disagree or argue the style matches corporate writing pre-AI. On substance, several commenters advise against 'agent' workflows and recommend treating LLMs as a code generator: one session at a time, iterate on plans, stay engaged. One commenter notes that keeping up with manual coding alongside LLM use helps maintain control. Another observes that the incremental approach doesn't help when code must be reviewed by another person, as they still see the lump sum.