Using an LLM to lookup syntax, common APIs, and conventions, seems to me like using a calculator to do basic arithmetic. It’s useful to memorize these things because it’s faster.
Moreover, if I know a key term or phrase (which is most cases) I can lookup those things in Google or IDE search, which is also faster than an LLM.
EDIT: to be clear, I’m still writing code. I can do many small tasks and fixes by hand faster than I can describe them to an LLM and check or fix its output. I also figure out how to structure a project partly by writing code. Many small fixes and structure by experimentation probably aren’t ideal software development, and maybe soon I’ll figure out LLMs (or they’ll improve) such that I end up writing better code faster with them. But right now I believe LLMs struggle with good APIs and especially modularity; because the only largely-LLM projects I’ve seen are small, and get abandoned and/or fall apart when the developer tries to extend them.
Moreover, if I know a key term or phrase (which is most cases) I can lookup those things in Google or IDE search, which is also faster than an LLM.
EDIT: to be clear, I’m still writing code. I can do many small tasks and fixes by hand faster than I can describe them to an LLM and check or fix its output. I also figure out how to structure a project partly by writing code. Many small fixes and structure by experimentation probably aren’t ideal software development, and maybe soon I’ll figure out LLMs (or they’ll improve) such that I end up writing better code faster with them. But right now I believe LLMs struggle with good APIs and especially modularity; because the only largely-LLM projects I’ve seen are small, and get abandoned and/or fall apart when the developer tries to extend them.