It's still an extremely useful tool if you accept its limitations, and I think it's being reductive to say it can only solve "dumb" problems or suggest boilerplate code.

I used GPT-4 the other day to refactor/optimize an extremely bespoke and fairly complicated geoprocessing script that we use at work that was written by a former employee who's no longer with the organization. Yes, it got some things wrong that had to be corrected (sometimes all it took was feeding it a stacktrace, other times this wasn't enough and I'd have to figure out the issue myself)

But at the end of the day (literally, this was over the course of an afternoon,) I'd managed to cut the runtime by more than half, using libraries I'd never before touched and wasn't previously familiar with. It probably would have taken a week to implement otherwise.