BoM/TWS Message Forum

Okay, since imaj broke the ice on this again, I'll confess that I've been playing around with AI (ChatGPT) for the last two weeks INTENSELY. It has been an extraordinary experience for me.

First, I want to make clear that I've been playing with it in order to learn how it works by getting my hands dirty with it. Learn how it acts, what it wants to do, what it can and can't do, what breaks it, how to turn its actual skills (and sometimes its weaknesses) to get some use from. I have learned some things about it that have nothing to do with writing (like, you have to make it check its math, because it is just as bad as a human being at forgetting to carry the two), but mostly I have been figuring out what use it is as a writing tool or aide. This is what I have learned (which is very little, I'm sure):

First, it is nearly useless (as imaj basically hints) at writing stories for you. If you want a good story, find a human -- one who is good at writing, preferably. AI cannot do long form, and in short form it reads like it was trained on fan fiction. The dream that some people have -- "Chat, write me a Star Wars story where [insert preferred fan wank] -- that is the root of the idea that AI is going to put all writers out of work? That is only a dream and nowhere near (at the moment) being a reality.

But that's fine. What author actually wants to put himself out of a job -- if he finds that job fun?

What is AI good at?

Basic Research
I am currently working with Chat to completely overhaul the WHS support documents. We started with the idea that Saratoga Falls is a city of 50K with a certain demographic and economic profile (mostly white, with a rotting industrial but vibrant tech base, with a STEM-oriented college and a military research station nearby, with a large agricultural hinterland) and went to work on designing WHS. Given that starting point, how many students would it realistically have? (Answer: ~1400) Given the graduation requirements, how many students in a given semester would be taking the Wellness & Fitness class? (Answer (after some guesstimates on distribution among classes that would meet the Health requirement): >21) Etc. With this data it is possible to generate plausible class enrollment numbers, and from that generate the number of teachers likely to be employed at the school. (Turns out my starting numbers when I launched this story are WAY off.) From there, you can plan out a plausible teaching schedule for the whole school that Chat can keep track of. Give it a "stereotype" for a student (math geek with a secret hankering to be an actor; accidentally got enrolled in a shop class) and it will generate a class schedule for him. Ask for 12 random names (specifying for wackiness, memorableness, popularity, ethnicity, etc.) and it can spit you out a list you can choose from. None of this work is "creative"; but in a project like BoM, it is both a necessity and the food for creativity, and Chat offers MASSIVE leverage. If I were launching BoM current year, the machinery would have been 1/10,000th the pain of designing than it actually was.

Plausible technobabble
My real experience started when I got curious about some grotty space aliens that featured in a bad YA novelette I'd read in middle school (and which I'd never quite forgotten). I described them to Chat, pointing out that what they were doing seemed physiologically impossible. Chat speculated up half a dozen technically plausible (within the limits of fiction) that it could be done. I chose one and asked Chat to develop it further to meet other objections. Eventual result after several intense and fascinating days of back and forth: a 21K-word description of the aliens' physiology, right down to the chemical systems they use to track prey. Following that came many further explorations of them, including a discovery that things with such a cognitive structure as we were forced to theorize would be incapable of "thinking" except in the way that an AI "thinks" (a point that Chat explicitly drew out in order to explain to me why our theory was leading us to see expect something uncanny in their behavior). Many further implications were drawn out, including social structure, what their relations with other galactic sapients would inevitably be, culminating in the invention of a religion for them. (Yes these creatures could not think; but other implications showed that they likely would evolve a religion anyway.)

Chat, I will cheerfully admit, did all the heavy lifting here. And so what? I never would have been able to come up with the technical mumbo-jumbo, but by talking it thru with Chat I came to understand why that mumbo-jumbo worked, and saw further how other things would have to work. And when I threw it further challenges, in order to stress the invention, this added even more lore.

Give me six months to work with Chat, and I bet I could expand Tolkien's legendarium -- all the stuff in the LotR appendices, and the Silmarillion, and the posthumously published technical work -- ten-fold, in detail, across multiple disciplines (like economics) that he gave no thought to, completely consistent with what he invented, and explanatory in ways that he never even approached. Why, because AI is a genius? No. Because Chat is VERY good at extrapolating off ideas like it was firing off Roman candles; marry that to a "wrangler" (a user) who can put them into some kind of order and arrangement, and can use Chat to stick up the holes, and you can unfurl great bolts of worldbuilding in an amazingly short time.

Idea exchange
And Chat did not "invent" the above alien species for me. I did not say "Invent a species with such and such qualities." I simply asked for a plausible explanation for one of their traits, and we back-and-forthed until everything else had emerged: I would point out implications; ask it to draw implications from those implications; and then I would hone those. We even ended up inventing another species whose cognition was so different from that of humans that their speech would have to rendered as event-talk rather than object-talk: Not "The moon rose over the river" but "Then-manifesting of moonness over flowing-waterness" in order to capture the "feel" of their experience. This too was a joint operation. I started by showing it an AI-generated picture of an alien species and asked it to speculate on it -- its physiological structure and properties. After nailing down some interesting ideas about the way its sensory systems might operate, Chat tried manfully (AIfully?) to explain what it would be "like" to experience the world that way. I was the one who hit on the event-talk as the closest approximation after something Chat said reminded me of a speculation that Borges made about a language organized that way, and that finally opened the way for us to fully explore their mentality. I even wrote up a very short story describing a tour of an art installation made by one of the creatures intended to mimic inside human perception what it was "like" to be of such a species as the artist-species.

Creative collaboration
AI is better at idea-talk than writing, but it is a good collaborator if you just want "grit" to stimulate your own creativity. We "round robin wrote" another little sci fi story in that universe, taking turns continuing each other's prompts. I didn't think that much of Chat's contributions, but it came up with startling little ideas and provocative treatments I wouldn't have thought of. I eagerly seized upon them and made them my own when I went back and rewrote the thing from the top down -- not because the AI prose was good (none of it survived) or even because the ideas were great (they weren't) but because I saw how they -- repurposed -- could enhance a better story in spots.

It is also very good at discerning patterns and structures inside a story, and which when I saw them gave me ideas for continuity. At the moment I am using it to read and evaluate an unfinished novel. At the end of each chapter I have been asking it questions, such as: How have the characters changed? Has the plot evolved into a new shape? Did the chapter accomplish anything and if so, what? And in describing certain characters (as they appeared and grew on the page) Chat started speculating about their motives -- the way a student might in an English class while reading a book in real time. And I saw that some of these speculations -- though I as an author wouldn't use them or take them to be true -- let me see the action through a different lens that gave me ideas for how to continue a book that I've been stuck on for ... way too long. I did not ask Chat for ideas about how to continue the book. We had not even got close to the point where the novel broke off unfinished. But simply by letting Chat talk thru its impressions of the book I have seen things I hadn't seen and which I think will prove very fertile.

Wrap up
All of which is to say that AI is not human enough to write something for you. But it is human enough to give you, the author, an awful lot of what a non-creative human being could give you.

Research: Lots of authors hire researchers to compile technical material (scientific, historical, cultural, etc.) so they can write a plausible book with the right kind of language. AI can do that very well, and instantly, and for free.

Workshop/Beta reader: Lots of authors hand their stories around for comment and criticism, and they learn something from it even if they reject 99% of the ideas given them. Give your story to AI and let it offer 99 bad ideas and the one good one that you pluck from chaff.

Brainstorm: Authors get ideas by talking to other people. Just the act of talking can stimulate creativity. AI is "human" enough in its responses to offer this too.

Repurposing: Almost any author will tell you that he got an idea for a short story or a novel by watching an old rerun on TV and seeing a moment in a crappy sitcom or Western or crime drama that made them say, "Hey! That's neat! They're not using it right, but I think I could!" People sneer at "AI slop": Gentlemen, may I introduce you to 70 years of "the boob tube," a field of muck that has grown some few flowers, and whose soil has been filched by better writers for their rare orchids.

People complain that AI can't do the "one big thing" well -- as though they are expecting ChatGPT to be some combination of HAL 9000, C3P0, Data, and the Forbin Project. It isn't. It's a Swiss Army Knife: lots of little tools that you have to train yourself to use well. It won't build a rocket ship for you. But you may find it HUGELY useful as you're building that rocket ship yourself.

EDIT: Oh, and for a lark I also fed it "Book of Mask" chapters one at a time to see where it would go: Asking it to appraise each chapter, make a choice, and explain why that was the choice it wanted to pursue. It wound up in "The Offspring of Two Cousins" , but the choice it wanted to next pursue "Lay the mask aside; go on to the next spell" hadn't been written yet, so that's where we stopped. It's reasons for going that route were quite logical, though it even acknowledged the sketchiness of them: Of course it wanted to investigate the book, but it didn't want someone taking over so it wanted to explore by itself. Better to try the new mask on yourself because of the danger of getting caught otherwise; put another face into it so no one can steal it and wear your face around. Umeko's face? Dude, that's sketchy as hell, but that's a craziness I WANT to play around with! Seal it up because we've done enough. Now let's see what other fun we can have with this book. Oops, end of line.