I don't think that AI being somewhat good is the main stopper, it's that we know for certain that it will improve at a giant pace, the very same prompt that triggers a mediocre book today will trigger a good book tomorrow (as an aside, I firmly believe that the word "trigger" is way more apt than "create" when talking about AI output), there is billions of dollars invested to make sure of that, and is not like we haven't seen unbelievable leaps already for art generation.
There is 2 things that are gonna help a lot, one is better classification, in that regard advances will help purge itself, just like our digestive system filters what we don't need models will do the same, e.g. scientific and factual models will have zero jokes in it's corpus, and you will know that the response you received came from such a humourless model, the second thing is just plain feedback, teachers will use AI and give it thumbs up or thumbs down and the machine will use such opinion over the opinion of most students, it will also favour the input by students with the most promise, those who win math championships and all that jazz.
They gonna focus on high value targets, that mean giving it for free to colleges and others in exchange for their cooperation, giving it for free to Hollywood plus custom solutions for their needs, in exchange again, for their cooperation, the entire internet was just a great bootstrap plan not their lifelong strategy.
> the very same prompt that triggers a mediocre book today will trigger a good book tomorrow
"Good" in which sense? That people read it and/or pay for it? But people already did, before LLMs: they read and paid for the most terrible, cliched, trite stuff. I mean, there are whole genres that are basically trash, before anyone even dreamed of AI (I'm pretty sure 90% of mainstream Hollywood script writers can be replaced by an LLM; they already feel like they were written by one anyway. This is not praise of LLMs, it's criticism of Hollywood!).
Surely, then, a "good" book is not merely something people will read or pay for. So why would AI become "good" at it, in which sense?
Reading/writing is a human activity. If you cut humans from a big part of the loop, how can the result ever be good?
This isn't the same context as writing code or building apps.
Years ago I read Luis Bunuel’s biography. When he visited Hollywood (in the 50s or maybe 60s) he made a small device out of paper which allowed him to predict the plot of the typical Hollywood movie.
It was made out of a few wheels, the outer one larger than the inner ones, all attached with a pin in the middle. He had written types and characters and events on the edges of the outer wheels.
He would ask you how the movie starts, what were the main characters, adjust the outer wheels containing these items and get the rest of the plot with very high accuracy.
Hollywood is a home of amazing masterful artists but the suits mostly bet on what has proven to work.
Haha, I didn't know that anecdote! It totally tracks with reality.
I didn't mean to say the people who work in Hollywood don't know their craft, plenty of skilled people who I'm sure would produce wonderful work (and sometimes do!) if given the chance.
I meant Hollywood as this machinery of algorithmic clones, as described perfectly by your anecdote about Buñuel.
