Originally published on omid.dev.
I was deep in a research thread that was not going to become a weekend post.
The topic started small: juniors asking an AI to format code that Prettier already owns, or to invent a debounce helper the repo already has. It got larger fast. Sometimes the person is not a developer at all — they have an idea, they paste a warning into ChatGPT, the "build" goes green, and they never learn that the message was ESLint. The software can become more sophisticated than the operator's mental model. That is a different problem than "juniors are lazy," and it is too big to finish in one sitting.
Then the conversation itself became the bottleneck. Scrollback got expensive. Earlier distinctions went fuzzy. I caught myself asking the model to remember what we had already decided, which is a bad use of a thinking partner.
The useful question was not "which model has a longer context window?"
