I will say there's a few common boilerplate things that is easier to express as quick prose and get 4-5 lines out of it, highly predictable lines that I could have probably done myself, a little more convenient when I'm rusty on a language or library. Also code completion, same thing of only good for like 3-5 lines at a time.

Still have to watch it like crazy and carefully fix the mistakes and still throw out what it suggested and it can be pretty annoying, but I'm willing to believe it saved me a smidge of time on those. But anything beyond that and it's a horrible mess I would never abide that's more trouble to try to fix than just scrap and start over. It's hard to stay vigilant and micromanage the details as I'm used to reviewing human code and the sorts of mistakes the LLMs make are distinct from what humans make.