My use of it has evolved quite a bit over time. Mostly what's changed is that I now never assume it's produced correct, functional code, but rather think of what it's done as a kind of template. It does a lot of fast typing and comes up with a pattern that is broadly speaking appropriate. But I think where people get confused is in thinking it's a matter of just describing the problem and getting the working code, so that non-programmers are suddenly programmers. That's not the case at all, in my experience. Basically, I find working with it helpful in that it moves me close to functional code quicker than I'd be able to get on my own.

However, I will say that I find GPT-4o's tendency to ignore instruction and just type like a mad asshole, so that every answer contains repeated, identical code to what it just typed 2 responses before, infuriating. It's actually incredibly bad. What's worse is that it will sometimes retype a version of the code that was already ruled out or fixed, so that if you're not careful it will rollback fixes you made several responses before. Shocking, really. This behaviour seems to be degenerative as chats get longer, so I now tend to start new chats (on the same project) at a higher frequency than I used to with GPT-4. I'm assuming this stuff will get worked out in time, but it can be excruciatingly frustrating working right now.