You’d think LLMs, having been trained on a large fraction of what’s ever been written, would be good at writing. Weirdly, they’re okay but not great. In my previous post about this, I shrugged and thought that we may never know what secret ingredient AI is missing.
Nabeel S. Qureshi, in What Makes Art Great, offers an answer. LLMs, at their core, predict the next word in a sentence. Given a writing prompt, they will complete it in the most predictable way possible. Qureshi argues that this predictability is the reason why AI-generated writing isn’t pleasant to read. By comparison, great human writers and especially poets will go out of their way to surprise us with their writing.
Nehaveigur 