Smut aside, long-form writing is actually not that "easy" to full automate with LLMs. Context windows cause things to break down very quickly. And overfitting

Smut aside, long-form writing is actually not that "easy" to full automate with LLMs. Context windows cause things to break down very quickly. And overfitting of brief and to the point RLHF responses, as well as summary datasets makes coaxing many of the models into slowly narrating things, as one would in a novel is challenging. Agent-based approach is one way to tackle this, but in that case maintaining narrative style coherency is also somewhat difficult to get right. Come to think of it, it's actually pretty hard to automate any sort of advanced artwork generation with LLMs. Once you go beyond your average prooompting, you have to put in a lot of work.