I think LLMs are useful for coding, but I think the idea that the workflow should be "computer write -> human review" is backwards. they are excellent at finding your fuckups, but they are also excellent at fucking up. to me it seems more productive to use an LLM as a kind of pre-commit gate. "Take a look over what I'm proposing here and see if it's retarded."
that tends to work and feels somewhat productive. when I have a problem and it spits out what looks like a reasonable solution in a few lines, I might copy that into place, but I don't just let it write my files. Too many times it seemed to understand the problem and didn't, or it decided it just needed to paint a picture of completeness rather than actually completing a thing. That is harder to detect, especially if you're getting mountains of code from the thing and you have time lines to consider.
that tends to work and feels somewhat productive. when I have a problem and it spits out what looks like a reasonable solution in a few lines, I might copy that into place, but I don't just let it write my files. Too many times it seemed to understand the problem and didn't, or it decided it just needed to paint a picture of completeness rather than actually completing a thing. That is harder to detect, especially if you're getting mountains of code from the thing and you have time lines to consider.
