I think this distinction sometimes elides people who are quite steeped in day to day LLM usage, because you can imagine a system that you can cobble together which would converge on correct output after you have spent hundreds of hours on being human scaffolding for it.

→ View original post on X — @patio11