This is one of the few times I've read a post about using GPT for dev work that has actually made a somewhat convincing argument.

In my day job, there's been various attempts to use GPT to speed up some work, which has generally resulted in me having to fix "confidently incorrect" code. The outputs usually about half right when its usable.

Using an LLM to basically auto complete trivial bits of code that are "just typing" but usually require a referral to the docs is probably the optimal middle ground for productivity - so time can be spent creatively solving the actual problems.

I'll have to experiment with it further - have it emit all the boilerplate for something, do the fun parts myself, then check for correctness might well be the workflow best suited to GPT.