literature remixing
#concept #remix #seed #generative #llmsLiterature remixing: take a book whose ideas are worth reading but whose writing isn't, and have an LLM do the labor of rewriting it while a being with a soul makes the creative decisions.
Background: Alongside Night Sucks
Spoilers: some writing sucks.
I recently tried to read Alongside Night (1979) by J. Neil Schulman, which received promising praise from the likes of Nobel Prize-winning economist Milton Friedman, and is oft referenced by cypherpunks for its depiction of "counter-economics," wherein a superior economic system replaces the status quo simply by being better. Economics are a theme in many of my own works and can make for thought provoking science fiction.
The very beginning was engaging enough: hyper-inflation and economic collapse in the United States and a young protagonist who needs to escape the country with his family, led by his economist father. Maybe a bit on-the-nose, but maybe the nature of economics in fiction sometimes requires shoehorning. The writing is terrible though. Extended thoughts of the protagonist that end in dead ends. The unbelievability of the aborted rappel from the 50th floor of a New York skyscraper which still somehow leaves the beginning of the book slow and boring; the prose and pacing are not compelling at all.
I lasted all of an hour before I had to delete it completely from my ereader, which is a shame because I like the idea of a scifi book whose main premise is an economic one. I will slog through a book for its merits (a few years ago I finished Sundiver by David Brin for talking-dolphin prior art even though it was also terrible) but the third, different paragraph about the protagonist not having any plan and deciding to walk in a random direction did me in.
A 1980 review backs me up: the book "does not have ... credible characters or a convincing plot," has "imaginative ideas, inadequately fleshed out," and "may be good propaganda, but it is bad melodrama."1
With a recently renewed and underutilized Claude plan, my first thought was whether an LLM could re-write it better enough to be worth reading. My second thought was LLM slop might be an even worse waste of time than regular-old-bad-writing.
Can an LLM write a book?
- Yes they can. Apparently Amazon is flooded with LLM-produced books.2
- Are they any good? Maybe some of them, who knows, but in general they are what is colloquially called "slop."
- Writing straightforward prose is not their weak point, it is the creative part of coming up with a premise worth reading.
- In the case of Alongside Night the premise is strong but the prose and structure are weak.
- Writing straightforward prose is not their weak point, it is the creative part of coming up with a premise worth reading.
My premise: in this case I don't need the LLM to write a book or even re-write a book, I primarily need it to read a book I don't have the willpower to read myself. Later it may draft, but only based on my unsloppy decisions.
Can an LLM read a book?
Yes. Much better than it can write. LLMs are very good at reading books:
- Starting with a corpus of text is the ideal state for a contemporary LLM.
- for example if you need help with something, providing the LLM with primary sources (documentation, a guide, etc) will generally result in much more useful output than asking it to hallucinate from scratch.
- An LLM doesn't get bored.
- It can cross reference, annotate, and even hold the entirety of some novels in a single context window.
How I might use generative LLMs to produce not-slop
At the time of writing this post I have not tested any of this, these are just notes that I may apply one day to the task of "remixing" existing literature.
I am confident even a one-shot pass from an LLM could tighten the prose, fix bad pacing, shorten. It could summarize the story no problem, we can all agree, and it could read the story, so it could probably do something part way between these two things.
Some of these could happen in parallel, and you could go through and fix them all in fewer steps with less feedback.
- HIGH-LEVEL PASS: Have an LLM read a book and note its high-level structure, characters, etc. A document with the story arcs, characters, premises, themes numbered.
- Require chapter numbers, page numbers, cross-references for each item
- for example ARC 1 involves characters A, B, D
- Optionally, let the LLM score what of these are worth keeping (for its novelty, or centrality, or simply good scenes and premises).
- Respond to this high-level outline with a first pass critique. What structural components are worth keeping or if they could be wholly better.
- Be specific: "make character A a cockaroach"
- These improvements are done by a human. Choose whether any of the LLM's "worth keeping" was tasteful.
- Ask the LLM to restructure the outline as much as necessary to match my improvement plan.
- Ask the LLM to pinpoint what breaks based on these changes.
- Using the format of the HIGH-LEVEL document
- "Character A is never heard from again after this scene if you change this."
- Come up with fixes to these breaking changes myself.
- MEDIUM-LEVEL-PASS: Have an LLM now make an outline of secondary components: lesser plot points, characters etc. Optionally ask LLM to score them.
- Respond to these points with better ideas.
- Ask the LLM to restructure again based on these new changes
- LLM should pinpoint what breaks
- Human should fix
- As in any writing, changes in one place may present choices elsewhere.
- Do this as many times as it takes to get a good story.
- RE-DRAFT: Based on the resulting new document with the incorporated changes, ask the LLM to rewrite, chapter by chapter, noting how the plot of any given chapter interacts with the others to keep them consistent.
- Judge the new synopsis myself.
- If I want, ask the LLM to argue the original was better.
- Tweak as necessary.
- Next ask the LLM to identify the weakest prose, list it in a comprehensive document.
- Either go through the list myself or, for example, ask the LLM to rewrite the bad prose with the prowess of Michael Crichton.
- Ask the LLM to read the whole thing again for consistency checking.
- Finally, read the book myself.
Why this might work
The bar for the book selected for remixing, at least in the case of Alongside Night, is extremely low. It was so bad I was not going to read it at all.
Further, simple Markov chaining and rudimentary neural nets (whose raw output was objectively worse than the sloppiest current LLM) lubricated with enough elbow grease can result in print-worthy poetry. A novel may be orders of magnitude larger than a poem, but current LLMs are even more orders of magnitude larger than the tools I used to write those poems. My GPT-2 Tao Te Ching project (hundreds of human-written translations in, a new one out) is an example of "literature remixing" except I never finished the elbow grease steps (lots and lots of unfinished notes exist though).
How this might not work
The real problem is bad plot, and if you let an LLM "fix" a plot it is probably going to give you blockbuster slop, which may be enough, but not very interesting. That is why my Steps 2, 3 and 5 are me, personally, using my own brain, coming up with better ideas than the original text and letting the LLM write the copy around these creative contributions, in a feedback loop. It's also why I am not interested in concepting something similiar for writing new books from scratch; I only write about literature "remixing" because I think starting with a specific piece of writing and working closely against it suits some of the strengths of LLMs.
An LLM reads well but quotes loosely. There may need to be some scaffolding and integrated diff checking to keep as closely to the original text as desired. Or lots of spot checks, references, and cross-linking.
Legality
Translations, adaptations, abridgments, sequels and screenplays are all derivative works requiring copyright holder permission. A "remix" probably falls under one of these types. Public-domain books are fair game. Alongside Night (1979) is still under US copyright.
In music licensing a remix requires permission to use the original recording. Interestingly, covering a song requires no permission. So long as the cover keeps the songs basic melody and chracter, it can be rearranged in another style. Covers just require payment of a mechanical/automatic royalty fee to the rights holder. Maybe the concept I'm describing is more like a cover than a remix, though in a musical cover you can specifically not change the lyrics. Something something don't judge a book by its cover.
Legal precedents (in the US) for using another text as the basis for a subsequent one are narrow:
- Fair use can allow, among other things:
- Criticism
- Is rewriting a book better than the original a form of criticism?
- Parody
- Is rewriting a book better than the original a form of parody?
- Criticism
Literary "abstractions" (tropes like a boy discovering he's a wizard and goes to a magic school) can't be copyrighted, and I imagine even niche economic premises can't be copyrighted so long as they aren't fully original coinages. In the case of Alongside Night the actual writing may be so bad that the only salvageable parts are the abstractions, so a "remix" would be practically unrecognizable from the original. A second defense is anybody who is willing to call you out for ripping off Alongside Night has to admit they read the whole thing, which is kind of embarassing, and also outs them as probably a libertarian (who is for strictly enforced copyright laws), which is probably more embarassing. I don't know, maybe copyright is the one thing some Libertarians want the government to enforce.
The fact that the author of Alongside Night is dead, apparently did not sell a lot of copies, and wrote in support of black markets is also no protection.
Next
I think the Wikipedia3 of this book is enough for me, so I probably won't remix it unless for the benefit of future readers more interested in the story than myself. I will document the process if I do.
Remixes
All remixes, including any not listed here yet: /tags/remix/
Footnotes
Michael Grossberg, "Adolescent Dystopia", Libertarian Review, March 1980. ↩
Ella Creamer, "Amazon restricts authors from self-publishing more than three books a day after AI concerns", The Guardian, 20 September 2023. ↩
replies — join on mastodon