Erin 💽✨4d ago
My timeline is full of people talking about LLMs making hard to detect mistakes and how it's going to ruin KeyPass XC and all I can think is "folks, have you heard about humans? Ooooh boy...."

Yes I did spend 30 minutes staring at some code going insane trying to find a bug only to discover that a LLM had made a 1 character copy and paste error. At first I was mad and then a moment later I started laughing at realising it was exactly the same shape of bug I had made 100 times before.

They're not miracle workers; they're dumb, make mistakes, are not creative, and typically I find their output to be mediocre. They're also not going to ruin your project unless you're asleep at the wheel or being lazy
maschmi3d ago
Show thread✧✦Catherine✦✧
@erincandescent they make different types of bugs, which are specifically more difficult to catch in code review, and they incentivize people to submit more code. I think the danger is real, and I also think that code review is *not* the place to be sifting through garbage looking for obvious bugs, and we should not normalize that
Nov 8 at 12:47amWeb
Show thread✧✦Catherine✦✧4d ago
@erincandescent if one actually gave a fuck about code reliability _or_ productivity, one would reimplement PVS-Studio's lints (which catch many copy/paste errors automatically) for $favorite_language_here instead of wasting time on LLM shit
Show thread✧✦Catherine✦✧3d ago
@erincandescent i made the same argument in the keypassxc repo and achieved consensus; i rest my case
Show threadAndrew Zonenberg3d ago

@whitequark @erincandescent This is something we're starting to think about at work - all of our testers have years of practice in finding the bugs humans make, but very little in finding the bugs LLMs make.

We have no idea how much LLM generated code is in codebases our customers give us - it's not something that has ever come up in discussions, although I'm starting to ask during scoping if they knowingly allow/encourage LLM code generation - which doesn't help.