[tptacek]

I was just talking about this in a chat today, because 'simonw had at some point talked about getting to the point where he was letting go of reviewing every line of LLM code, and I am nowhere close to that point --- I'll take Claude's word on Tailwind classes as long as the HTML looks right, but actual code, I review line-by-line, token-by-token, and usually rewrite things, even if just for cosmetic reasons.

[vidarh]

I've had Claude write ~50+ commits of compiler code over the last week, and I've mostly not even read it yet. I will review the end result at some point, and I do spot checks, but for the most part I review whether or not the number of specs that passes is going the right way, and occasionally butt in when I glance over and spot it doing something silly.

I fully get why people won't trust it, but I'm leaning more and more on reigning it in with specs and tests and process, and then letting it run for extended amounts of time, and then giving it instructions to fix obvious issues, before finally reviewing the result, and letting it fix whatever comes up.

[stavros]

Yeah, there's a definite continuum for where LLMs are most to least expert. They seem to be fairly OK with JS, less so with Python, and C is just a crapshoot.

It depends on the project as well, for throwaway things I'm fine to just let it do whatever it wants, but for projects that I need to last more than a few days, I review everything.

[tptacek]

A friend of mine has been doing embedded C stuff (making some kind of LCD wall) and has been blown away by how well it's been doing with C --- he went in an LLM skeptic (I've been trying to sell him for months, it finally clicked).

[stavros]

Huh, interesting. I wanted to turn an old rotary phone into a meeting headset, and I tried to get Opus to make me a sound card, but $35 in API costs later, I had no sound card.