...are you sure that Claude Sonnet writes code that you'd write?

For me, it writes code that mostly works, which is great. It insists on adding some of the least readable comments I've seen. It also likes to pad out the code to about double or triple what I would write.

It doesn't seem to have that little voice that says "hey, this approach sucks" so it needs the harness or the prompter to tell it to go back and refactor. Even Fable and Opus struggle with this.

▲

zahlman 16 hours ago | parent [-]

> It doesn't seem to have that little voice that says "hey, this approach sucks" so it needs the harness or the prompter to tell it to go back and refactor.

Well, yes. For humans, the "this approach sucks" voice is still very separable from the thought process that just figures out the syntax and algorithms etc. The important thing is whether the overall system has this feedback and that it's having the desired effect. Better harnesses are progress in the same way that more capable models are.

And yeah, there's a wide range in the standards that programmers hold, in practice. Plus they don't at all agree on what's "better" than which.

▲

strken 14 hours ago | parent [-]

Regarding the "this approach sucks" voice, there's an important difference between humans and AI which I've noticed and which I didn't explain well.

The general human approach to programming at most companies I worked at prior to 2024 was to head off in vaguely the right direction and refine the approach as we went. There might be a design and a plan, but at some point we just winged it and fixed our work after the fact.

This seems to be the worst way of getting an AI to do anything. They do better if they can one-shot an implementation from a clear spec then only need to fix bugs. Better models can handle a less clear spec but they still can't really iterate in a major way. To have an AI model build a bigger thing, you one-shot increasingly detailed specs and decompose them into smaller ones.

	▲	zahlman 5 minutes ago \| parent [-]
		I can only speculate, but my guess is that this limit is pretty fundamental, yeah. However capable the models get, they can't read your mind. The kind of iteration you have in mind is difficult because they don't have a basis to iterate upon; to the extent that they could pretend, that gets into what people have been calling "taste". And even if they somehow got better taste, it wouldn't be your taste. If that were acceptable, why bother prompting anything in the first place? At that point the system might as well choose its own projects entirely. But I don't think humans would want to live in that world; even if not paperclip-maximized, we'd eventually become an underclass. That's assuming basically unlimited capacity for the models to get more powerful. If there are limits (and I hope there are!) then it'd be even more certain that projects continue to require human guidance. But even despite that, I hope I don't sound AI-psychotic if I suggest: the models are capable of at least assisting with that spec-writing and decomposition.