Cool! I've seen things like this before and think it's very promising. I wonder if there's value in some utility to make sure changes are under some information threshold—to make sure it can only adjust so many characters, etc.

Using cross-entropy loss to measure the space of narrative entropy is neat. Another good article to throw on to the pile is arxiv.org/pdf/2103.128... from the EleutherAI folk

Hey! Glad to see @dbamman.bsky.social using semantic tags for annotation. I've been using them as well. They're especially nice when you constrain output with a CFG: github.com/ggerganov/ll...