[spankalee]

> the ones who are having a magic robot write a thousand lines of code that almost, but not quite, does something sensible

Gee, I wonder which "side" you're on?

It's not true that all AI generated code looks like it does the right thing but doesn't, or that all that human written code does the right thing.

The code itself matters here. So given code that works, is tested, and implements the features you need, what does it matter if it was completely written by a human, an LLM, or some combination?

Do you also have a problem with LLM-driven code completion? Or with LLM code reviews? LLM assisted tests?

[rsynnott]

Oh, yeah, I make no secret of which side I’m on there.

I mean I don’t have a problem with AI driven code completion as such, but IME it is pretty much always worse than good deterministic code completion, and tends to imagine the functions which might exist rather than the functions which actually do. I’ve periodically tried it, but always ended up turning it off as more trouble than it’s worth, and going back to proper code completion.

LLM code reviews, I have not had the pleasure. Inclined to be down on them; it’s the same problem as an aircraft or ship autopilot. It will encourage reduced vigilance by the human reviewer. LLM assisted tests seem like a fairly terrible idea; again, you’ve got the vigilance issue, and also IME they produce a lot of junk tests which mostly test the mocking framework rather than anything else.

[queenkjuul]

LLM code reviews are completely and utterly worthless.

I do like using them for writing tests, but you really have to be careful. Still, i prefer it to doing all the testing by hand.

But for like, the actual code? I'll have it show me how to do something occasionally, or help me debug, but it really just can't create truly quality, reliable code.