Anybody care to comment whether the quality of the existing code influences how good the AI's assistance is? In other words, would they suggest sloppy code where the existing code is sloppy and better (?) code when the existing code is good?
What do you think? (I don't mean that in a snarky way.) Based on how LLMs work, I can't see how that would not be the case.
But in my experience there are nuances to this. It's less about "good" vs "bad"/"sloppy" code and more about discernable. If it's discernably sloppy (i.e. the type of sloppy a beginning programmer might do which is familiar to all of us) I would say that's better than opaque "good" code (good really only meaning functional).
These things predict tokens. So when you use them, help them increase their chances of predicting the thing you want. Good comments on code, good function names, explain what you don't know, etc. etc. The same things you would ideally do if working with another person on a codebase.
