Comment by Nextgrid
6 hours ago
> hard problems are best solved by breaking them down into smaller, easier sub-problems
I'm ok doing that with a junior developer because they will learn from it and one day become my peer. LLMs don't learn from individual interactions, so I don't benefit from wasting my time attempting to teach an LLM.
> much like compilers did for Assembly programming back in the day
The difference is that programming in let's say C (vs assembler) or Python vs C saves me time. Arguing with my agent in English about which Python to write often takes more time than just writing the Python myself in my experience.
I still use LLMs to ask high-level questions, sanity-check ideas, write some repetitive code (in this enum, convert all camelCase names to snake_case) or the one-off hacky script which I won't commit and thus the quality bar is lower (does this run and solve my very specific problem right now?). But I'm not convinced by agents yet.
>often takes more time than just writing the Python myself in my experience
I guessed you haven't tried Codex or Claude code in loop mode when it's debugging problems on its own until it's fixed. The Clawd guy actually talks about this in that interview I linked, many people still don't get it.