The best use I've gotten out of LLMs is as an autocomplete. I use Cursor at work, and it's pretty good at consistently calculating the next 10-20 characters I want to type out. Anything longer, save for some situations where the changes I'm making are super repetitive, the quality dives off a cliff.
I've yet to coax out good/working code of significant complexity from these models without putting an amount of effort into prompting that would be greater than just working it through myself without any LLM assistance.
The use I do get out of Cursor can save a lot of time for me, so I do think it's a productivity boost as is.
I've yet to coax out good/working code of significant complexity from these models without putting an amount of effort into prompting that would be greater than just working it through myself without any LLM assistance.
The use I do get out of Cursor can save a lot of time for me, so I do think it's a productivity boost as is.