> Is it totally useless or is it the greatest thing ever? If neither, where in the middle do you put it?
For code completion, I think it's close to useless in my experience, traditional code completion feels much more useful.
> How often does it provide plausible but incorrect completions? Is it every few minutes or is it a couple times a day?
It varies with the workload but closer to every few minutes than a couple times a day. For example while writing rust the majority (like 95%) of the code completion suggestions are incorrect. When writing a python website it gets better but you still get bad suggestions that look good, several times a day.
The killer feature is generating code, not as completion but after an explicit prompt. Most models are okayish on that task. But still you have to pay attention.
That's all in my experience across like a year, your mileage may vary.
For code completion, I think it's close to useless in my experience, traditional code completion feels much more useful.
> How often does it provide plausible but incorrect completions? Is it every few minutes or is it a couple times a day?
It varies with the workload but closer to every few minutes than a couple times a day. For example while writing rust the majority (like 95%) of the code completion suggestions are incorrect. When writing a python website it gets better but you still get bad suggestions that look good, several times a day.
The killer feature is generating code, not as completion but after an explicit prompt. Most models are okayish on that task. But still you have to pay attention.
That's all in my experience across like a year, your mileage may vary.