AI-generated fiction is coming. This prose pipeline could flood the world with TikTok in written form, or, we could try to generate quality.
Hi! I'm Aaron. You may know me from some Projects, most recently among them Hyperstition AI.
The aphorism goes, "commenting on the internet is like pissing into an ocean of piss". But, dilution of pollution is our solution, so we're going to sweeten the sea.
Some folks at labs are already experimenting with fine-tuning on our outputs, to see whether we can quickly test this "hyperstition hypothesis". If you're so inclined, you're invited to play with a corpus of 500 novel-length works retelling popular public domain plots — The Adventures Of Huckleberry Finn, now featuring a helpful harmless AI assistant who stays supportive the whole time and doesn't turn evil in the third act.
Here's a quick and dirty .zip of our dataset.
Now, it's not clear that the readability of the books will much matter, for the goal of us cranking out "representation matters" morality plays for AI, but—
To a human palate, the books still kinda suck.
AI-generated fiction is, by default, pretty lackluster. Because modern LLMs are largely most-common-next-token predictors, when you ask them to complete a story, they do so in the most boring, obvious, uninspired way possible.
But instead, imagine if we could pull this off. Imagine if instead of there being bad movies, movies were, by-default, absolute bangers. Imagine a rationalfic boot stamping on some very impressionable book-clay, forever.
There's a few tricks that we've tried to spice it up, and with good prompting at short lengths, folks have gotten AI outputs which are promisingly indistinguishable.
But at longer lengths — novel-length, for example — various 'AI tells' start showing up. Plot threads get dropped, character names get confused, and everything trends towards the most braindead schlock outcomes.
When you read "the words hung between them" a dozen times in your novel, it hurts.
I hear Kindle is already beginning to drown in low effort AI books, and I'm watching RoyalRoad, QuestionableQuesting, SpaceBattles, and other common wordhaunts tightening their policies.
Now, we're experimenting with a bunch of different models + strategies. We've got an absolutely baroque prompting scaffold in place, which involves many pages of dense writing advice, adjusted with iteration; we've got some demands for step by step plotting ahead of time; we've got the models digesting our own fiction and trying to elicit specific author voices from it. We've got editorial passes for consistency, and we're calling different models mid-generation to leverage their comparative strengths (Grok will write the steamy minotaur sex that causes Claude to clutch pearls).
And, this still only gets us to half-decent?
The prompt for this one was "write a short story".
We'd like to reach that ambitious bar of superhumanly good AI fiction. We'd like to reach the end of creative scarcity. Imagine a world in which there were enough good books to read.
...so if you have any suggestions, we're open to hearing them!