Earlier quoted context omitted.
> That said, I have fed thousands of lines of code into both of them and asked them to refactor it, and neither one of them made more than one error. All code otherwise compiled and worked first try. I would be very interested to get a more detailed scope of what you did here. Feeding thousands of lines of code into GPT4 and getting a near perfect refactor does very much NOT sound like my experience, but it seems hig…
Yes, I've observed the same phenomenon. The more detailed my prompts are, the more errors GPT tends to make. I use it as a partner to discuss implementation ideas, before I start coding. That works very well, because gpt and I usually find somethings, that I missed at a first glance. But coding with gpt or co-pilot is too disrupted for me.
This is interesting; I do a third thing that's the exact opposite of both of you.
I generally know the architecture I want, so I just slowly spoon-feed it to GPT-4 and have it build each of the components with the right interfaces to fit together. Usually, building each component requires a few rounds of hand-holding it through various refactoring. Its first draft of code tends to be, for lack of a better description, totally lacking in any creativity or elegance --- generally bland, insipid, and mundane stuff.
But it'll gladly take refactoring directions to get it into shape, so it's still very quick to get a whole bunch of components written up. Then I just glue everything together on my own. And the whole experience is still way better than writing a bunch of boring boilerplate code by hand.