What's this about?
The argument in one line: writing code always bundled something we never had to price separately — *understanding* it. You could not write something you didn't understand and get a passing grade. Now that agents do the writing, understanding has become optional, and if you also let agents review and test it you can skip mental models altogether.
Why was understanding free before?
Because writing forced it. To produce working code you had to hold the problem in your head, and that act of holding was the understanding. It was never a separate line item, so nobody costed it — and now that generation is decoupled from comprehension, the cost shows up as a choice you did not previously have to make. The author's caveat matters: for sufficiently low-stakes software, not understanding it is a perfectly reasonable trade.
What's the community saying?
- The most-defended position: "I review all of my AI-generated code and I'm still sufficiently sped up" — the ability to emit large quantities of text is already the productivity leap.
- The responsibility argument, stated flatly: if I ship code I don't understand and it has major problems, I am responsible, not the LLM. One commenter adds that sending unverified output for review is disrespectful to teammates.
- The most interesting personal report: coding skills have noticeably atrophied — to the point of scheduling manual practice — while the ability to read and comprehend code quickly has improved dramatically.
- The awkward question nobody answers well: what do you say when your boss asks why a colleague produces 10–100× your lines of code?
- A genuine counter-position: for someone who was never fluent at getting thoughts into code, agents removed a bottleneck that was never about understanding at all.