|
I don't think it's real complexity. I mostly do backend and ops. Usually, I never generate backend code with llms, I mostly use them to generate test cases and small helper functions, but each time I try to do complex stuff, even with low domain specific knowledge, llms do not help. Rubber ducking is my most recent use (thanks to a comment here) and is fine, making them still useful _enough_. However, this week I had to do some front-end. I tried llms, once again thinking it would hurt me more than help me, but I was wrong. It can generate 100 good lines of good front-end (granted, mostly htmx) in 3 or 4 iterations. Useful, working stuff, not shit I have to spend hours debugging. I was helped by the fact that I already had an architecture and data models (worked on that yesterday), and it just had to generate forms and containers without real algorithm inside (well, and a FSM but I wrote it by myself, it was easier than try to explain it to a LLM. And it's fun to write). [edit] forgot my point: I don't think front-end is really more complex than it was in 2014. It is more tedious, and use more tools, but also more streamlined and straightforward. My story about how llm fare is also about that: it seems really good at generating boilerplate/streamlined code, and front-end is mostly that. Also now I understand devs who say that llms took 50%+ of their workload, when before today I thought they were lying or not realising the lost time from using them. |