[vineyardmike]

> an LLM is regularly asked to produce a valid block of code top to bottom, or repeat a section of code with changes, when that's not what we do.

Eh, it's mostly what we do. We don't re-type everything every time, but we do type top-to-bottom when we type. As you later mentioned, "next edit" models really strike that balance, and they're like 50% of the value I derive from a tool like Cursor.

I'd love to see more diff-outputs instead of "retyping" everything (with a nice UI for the humans). I suspect that part of the reason we have these "inhuman" actions is because of the chat interface we've been using has lead to certain outputs being more desirable due to the medium.