They always talk about a model scoring 99.7% on a test. Well, how about a joke test? I confidently presume the answer is currently no, and I think it stays no for a long time, even at their most capable. This isn’t a moralistic “go humans” judgment; it’s the technical reality. They operate on a probability distribution of data. I love the idea that... See more

When you prompt an AI model to write a whole story, do a revision, or write part of it, you’re essentially making one big decision. That’s very different from the real human process of writing, where every single thing is a small decision.