I mean, these models are super useful for small defined tasks where you can check their output.

They're also useful for new libraries or things that you're not an expert in (which obviously varies by domain and person, but is generally a lot of stuff).

I'm a data person and have been using them to generate scraping code and web front-ends and have found them very useful.

But I wouldn't trust them to fit and interpret a statistical model (they make really stupid beginner mistakes all the time), but I don't need help with that.

Like, in a bunch of cases (particularly the scraping one) the code was bad, but it did what I needed it to do (after a bunch of tedious re-prompting). So it definitely impacts my productivity on side projects at least. If I was doing more consulting then it would be even more useful, but for longer term engagements it's basically just a better Google.

So yeah, definitely helpful but I wouldn't say I'm dependent on it (but I'd definitely have made less progress on a bunch of side projects without it).

Note: it's good for python but much, much less good at SQL or R, and hallucinates wildly for Emacs Lisp.