Are readers generating fiction with AI models?
arxiv.org
arxiv.org
Why not? Journalists, lawyers and pundits of various stripes are already doing it. Why shouldn't readers?
The big differences:
- fiction tends to be longer, AI struggles with making a satisfying coherent plot structure after a certain length.
- fiction tends to be subtler. You want characters to have nuance, shades of grey, symbolism, etc. Not everything should be shouted in your face. This is the opposite of writing persusaive literature where you are trying to convince someone of something.
There's money in copywriting, there's money in writing code, but long form creative writing? Almost an opposite of money.
Thus, very little incentive for AI labs to spend any effort on improving long form creative writing performance.
Most of tunes on this generation of AI systems are basic RLAF, and aimed at something like "punchy short form writing". The potential is largely unexplored.
The challenge is length and the context window. I had to own the plot since once it compacts, it has already lost the plot :)
I will give it pointers and tell where to go next. It will do it but either 1) stop after 3-4 paragraphs, or 2) write in a totally different direction than I expected. So I will nudge it to course correct, and it will do it. And then you keep iterating.
It was a good experiment for a couple of days. It saved me typing and it sometimes gave me lines of thinking that were interesting. But nothing that could be published.
But now that I think about it, I found a use case. If only GRRM could use its help :)
If they could reliably generate high quality non-fiction, that would be news.
"non-fiction" doesn't mean "correct".
There are things like the character(s) you interact with, the initial setting, and some background information predefined, but the responses/evolution of the RP vary depending on things like the model used and the user's interactions.
It's somewhat like taking a D&D setting/scenario. Each group plays it differently.
As much as it's nice to know another shares my interest (and will partake in it with me), I can't always get that.
When I can't get that, LLMs are not a worthless substitute, especially if what I actually want is just to play out a situation.
I want to act and then get back a description of reaction. LLMs can describe reaction, therefore I find value in them.
Like how in D&D you go to do whatever and the dungeon master tells you what happens in response. LLMs fit there for me.
They don't fill the void of social interaction or anything like that, but I don't strictly need that here.
Take a look at services like AI Dungeon to see the kind of thing I'm talking about.
It doesn't really replace social anything, but it's kind of like text adventure. It's plenty interactive enough.
Like, "why not both?"
Also finding AI-generated content to similar tastes is a great way to potentially finding buddies. I've made a lot of friends from certain ventures
Why am i not surprised its primarily porn
- Plenty of people are into very niche things. Sometimes almost nothing exists (or is discoverable) that exactly hits the spot. Generative AI lets you hit the spot that nothing else has
- AI is less complex than managing a real friendship. Even if people exist who can help you out with this, it can be difficult to arrange, or you lose access to them etc. AI doesn't lose interest in you, get busy with life stuff or decide you don't deserve a second chance.
Also sharing is still a huge benefit if you can manage it. Sometimes you will only find other people share your spot once you start posting the type of content they've been looking for. (Woe is me that FurAffinity unconditionally bans any LLM assistance now)
That’s why you’re not surprised.
It’s also, on the whole, a much more ethical and engaging form of smut. I don’t really have an issue with the choices of consenting adults, but the (video and image) pornography industry is so seldom that.
My take? It was passable. Over the last year I've read quite a bit less because I spend a lot of times looking at screens during the day and periodically get headaches, but I read something like 12-36 books a year depending on the year. Some of the fiction I read is good, others is... well, it's adequate.
From a purely "was it good enough to make money" standpoint, a book written exclusively with codex could probably sell copies. There are a lot of crappy authors out there too.
Only a few studios produced shows. With Youtube etc., many of the consumers could become producers themselves.
If I read this correctly, books and fiction are headed in the same direction.
LLM-written fiction as explored in the study is generally not published at all. It's treated more like an externalized imagination, a loop of general ideas fed into and expanded on or filled in by the machine with statistical averages. It more closely resembles a sandbox game in my view, a type of media distinct from anything before it in form, and even more distinct in function in that media is generally understood to be a vector of communication between people, and this is instead highly individual.
Actually, it might be closer to say this is similar to a child playing pretend alone with their toys, except perhaps a bit less challenging in that creative roadblocks or narrative building is instantly abdicatable to the machine.
Wouldn't the obvious analogue be a video game? Especially one where you can edit the asset files (making your weapons super-strong, for instance?)
An LLM-authored interactive fiction is a sandbox with no author, with no boundaries and no intent. It's a canvas, and all limitations on content are self-imposed. It's closer to making a game than to playing one, but there is no game in the end that you can share. You're left with a transcript that only really appeals to you. It is not the game. The game fundamentally lives in the player's imagination, and the LLM is there to ease the realization of that imagination and, if I'm being uncharitable, do the hard part of synthesizing ideas and impulses into something semi-coherent, limiting the potential for the player to grow and develop creative skills.
You can speak a world into existence, entirely customized to you’re preferences, and interact with it.
I recommend Super Supportive[0].
[0]: https://www.royalroad.com/fiction/63759/super-supportive
I don't really see how AI 'democratizes' fiction and book writing?
When people say "AI will democratise $FOO" they don't mean it the way it is usually understood: people with skill and motivation can now produce $FOO.
What they mean is that people with no skill and no motivation can now pretend that they produced $FOO.
Remember the "I'm the ideas guy" people who would ask you to build a complete system and offer 1% equity because "the ideas are the hard part"? Well, the future is going to be filled with "ideas guys", making it even harder to find signal in the noise.
Music has sinilar issues, but I think of this like the compression/loudness issue in music production. Everything gets amplified so the range of everything is compressed. And then it gets boring and people slowly jump ship for something else.
I think there will be a wave of AI slop that improves and might actually be exciting on some vector, but we will get bored eventually. Humans crave newness, even if that new thing is worse in exactly the ways that defined good previously, like punk in response to complex rock albums.
AI can combine ideas in interesting ways, but it is by design a predictor of what is most likely. This is directly odds with the concept of newness (and arguably, human-ness) which is baked in to what we consider interesting and relevant.
I'm not sure whether Harry Potter pioneered the "school + wizards" mashup, but the school has many aspects you would expect of a school (save for the wizards) and the wizards have many aspects familiar from previous fictional wizards (save for the school part.) The novelty comes from combining the concepts in more or less straightforward ways.
Even a writing AI that is only capable of taking an existing plot and adapting it for a different setting could produce enough reading material to last a lifetime by taking all existing novels in the public domain and putting each in a different random mashup for modern audiences. And I think many people wouldn't tire of this at all.