So, I haven’t used Copilot or etc, but I use GPT pretty heavily in coding, and it’s crazy to me that “please autocomplete this thing I typed part of” is the model for AI assistance.

In my experience, GPT functions best when you describe a problem, or a particular refactor (“make log() asynchronous, returning immediately but queueing up the provided string to be written to the log in the background”) (“separate functions A, B, and C into their own separated class”), and then paste the segment of code you want it to operate on, and then paste its reply back into your code. Its generations aren’t perfect, but it can save a ton of time and typing when used in this way; I can’t really imagine that it works well if all the guidance you give it about the code it’s operating on and the task you’re trying to accomplish is what’s currently under the cursor.

I guess most of the issue is that not a lot of the time I spend coding is literally typing code into a blank editor. I would say 90-95% of it is augmenting or fixing existing code with some particular goal in mind. If I were trying to design an effective AI assistant, it would have a sidebar with a class diagram of the currently-relevant classes and their methods, and you could toggle a little eyeball for which ones of them you wanted to “show” to the AI to give it context, and then a box where you could request in a couple sentences the modifications you wanted it to make to that specified code. And, if what you wanted to do was generate some new code, you could do that, but it wasn’t designed around that as the central principle.

IDK, like I say, I haven’t worked with Copilot and friends so I don’t really know, but that’s what I would do.