It feels like most "good" use-cases of LLMs in coding are for cases that don't involve "good code".
In the last few days:
- Implementing 1-off scripts
- Proof-of-concept in hostile environments (VSCode extensions)
- Reviewing massive PRs

So yes, they are useful, but I still don't trust the code they generate.
I've not seen them generate what I'd consider "good" code yet.