Chris Coyier asked whether the smaller Sonnet model beats bigger Anthropic models at mechanical jobs like checking a design system's components. Pitre says from trial and error that the bigger models add their own ideas, which is sometimes great and sometimes not what you asked for.
…that feels more deterministic. And from my perspective, I would prefer to have a script be written to script those things. But of course, you can unleash AI on it and it'll do it itself. And I would probably trust one of the more like doer models, like a Son of 5.5 to do something like that. That's
a little interesting nugget we should hang on to for a minute. You think Sonnet is better in the anthropic universe at a job like that than Fable? Because Fable will be like, let me, I need to think about this. So it's literally better. Sonnet will just be like, have job, must do job.
Sonic, Sonnet, and just my trial and error has been like, Sonnet has been like, okay, I'll do what you asked me to do, and like nothing more. But then if you unleash a Fable or 5.5, it's going to insert its own opinion. It's going to say, I see what you did here, but I'm going to do you one better. Like it's the kid who like is like giving a cheat. It's the extra credit kid who's like, hey, you forgot to assign homework today, teacher. Sometimes that's great, right?
Like, oh, there's my Granger. There's a documentation file tucked away in your thing somewhere. Let me update that to you while I'm at it. That kind of thing. Which can be because it could call out
things, but a lot of times it'll end up actually doing it for you. And sometimes you don't know that it did the thing until you see it in a pull request and you're like, wait a minute, it inserted this file. I see the benefit of it, but it was something that I was unaware of. And I'm very keen on…
seen you do that. It's kind of fun. So this does round it out, right? We talked about making a design system. We talked about expanding and improving a design system, but we now we're jumping into using the dang thing, which is where literally all the value is, you know? And