Ask HN: State of the art on using RNNs or Transformers to assist with fiction?

This is strictly for fun.

I am sure many of you have seen Karpathy train an RNN to predict the next character (sometimes word) of the Shakespeare corpus.

If you feed it a few input lines, it gives lots of very Shakespearen looking and sounding text.

Similar tricks can be done to train an RNN to reproduce things that look like Simpson scripts and so on.

The other interesting thing is that with GPT3, we see that there is some kind of higher order memory too and the paragraph can be on the same topic.

I am technical enough to experiment with these texts, download and run them and so on (except when there is a dependency collision in which case obviously no one can do anything).

But what is the state of the art in computer assisted fiction? I am thinking specifically of:

1. You outline a story, and then write out the first line of each one, and let it expand the paragraph.

2. You actually use it generate character dialog so that each character sounds different. You could have very authentic Homer Simpson talks to William Shakespeare dialog

3. Perhaps even have the computer generate a half dozen outlines based on feeding it 100s of outlines.

Any tools to make it easier to experiment? Right now I can try to run Karpathy's minGPT on Jane Austen, or just run a transformer example from the pytorch examples on that text but if anyone has already done some work, has resources, paper citations, or just thought through it, I am all ears.