True but irrelevant. If we want to get to decent art, we'll need something well beyond an LLM.
Humans generate stories using introspection. Every writer is their own first reader. They rapidly and iteratively try out story elements to create experiences and feelings. They then find the right words, again seeking out particular feelings. After iterating on the words for a while, they then have others read the words: partners, friends, eventually an editor. Those people all experience feelings, which get observed and sometimes articulated to the writer, driving further revisions.
LLMs can create story-shaped things. They can write glib, imitative dreck. But to write a new, compelling book, they need to be able to iterate against high-quality simulations of reader emotions. That may be something that we can create in the lab, but if we do, it won't be called an LLM.
Right, I think humanity will some day create something resembling an AGI, but the current LLM statistical approach is extremely unlikely to be the path to that. You need many more layers of sophistication.
Like, AI art generators don't just have issues with human hands or eyes, they also create buildings which are architecturally incoherent. I don't know how you solve that with mere data or refinements.
True but irrelevant. If we want to get to decent art, we'll need something well beyond an LLM.
Humans generate stories using introspection. Every writer is their own first reader. They rapidly and iteratively try out story elements to create experiences and feelings. They then find the right words, again seeking out particular feelings. After iterating on the words for a while, they then have others read the words: partners, friends, eventually an editor. Those people all experience feelings, which get observed and sometimes articulated to the writer, driving further revisions.
LLMs can create story-shaped things. They can write glib, imitative dreck. But to write a new, compelling book, they need to be able to iterate against high-quality simulations of reader emotions. That may be something that we can create in the lab, but if we do, it won't be called an LLM.