Okay, admittedly this isn’t exactly my speciality, but I’m pretty sure that if they’re keeping an ever-growing database of all input images, they’re doing the AI training wrong. (I mean, sure you can train a language AI with a few hundred hours of audiobooks, and it probably is the most cost-efficient way, but that doesn’t give you a database of few hundred hours of audiobooks, it gives you a language model.)
Alright, I just looked at an AI writer. Perhaps there are better ones around, but this one was hilariously bad. Just a small segment of a much larger and equally funny script I wanted to share as why I’m not concerned about this particular AI generator at least replacing me in the near future. (I wasn’t even trying to trip it up by the way.)
Chris stared at Sally with the gaze of 6271 melodious magpies.
He said, “I love you and I want your pencil.”
Sally’s blond arm gripped the tattered, jittery newspaper and said “Chris, beam me up Scotty.”
I shared this article a couple of months ago on the writers thread. Amazing AI writing examples:
“the moon was truly mother-of-pearl, the white of the sea, rubbed smooth by the groins of drowned brides.”
"Alice closed her eyes and sighed, savoring the moment before reality came back crashing down on them like the weight of an elephant sitting on them both while being eaten by a shark in an airplane full of ninjas puking out their eyes and blood for no apparent reason other than that they were ninjas who liked puke so much they couldn’t help themselves from spewing it out of their orifices at every opportunity.”
Though we’re still at the silly stage, AI-generated art and entertainment are clearly going to be here in full strength soon…and while I think creative humans will still be competitive with AI, it’ll be one more source of income inequality, as the people with the best algorithms make a lot more than human creators do.
I can’t see myself using AI for writing, because I don’t have a struggle to come up with writing ideas or take a volume-based approach like the indie authors in the Verge article. For my cover art, I’d rather support a human creator than a computer team.
Jesus, that AI is so bad it intentionally gives AI a bad name, won’t be surprised if it’s running in Markov chain
Try chatgpt from OpenAI or character.ai, they are much better if a little bland, both are free amazingly and only requires an account
Also for AI art, the good ones use diffusion model and it’s impossible to actually source their generated image, since they don’t mash different elements from existing images, the vox documentary is great on this topic
To be fair, they probably entered some terrible examples of those really bad dollar store romance novels into that particular generator for it to learn from. Poor computer never had a chance
That’s some next level stuff about the puking ninjas though lol.
Oh yeah I agree. Eventually it’ll get better. And when it does, why would companies pay a human for something a machine will do for free I can actually see it being used the way you’re saying with the computer coming up with the script and the human using it to base their writing on. I kind of do hope it’s still a while off replacing humans entirely, there’s just something about people losing the need to be creative in arts and writing that strikes me as sad, but then again I’m from the oldschool generation now
Maybe they fed it with bad fanfiction?
If the cheap romance novels are using metaphors of that vividness, I’ll have to read more of them.
Kanye used to be anti-nazi when young, so yeah…
The savior complex with religion is still there, so it got that right.
the character definition can be very simple since it’s made by a user, so the quality of any one AI is dependent on user effort, especially for less well known people since the AI can’t grab info and setting off Wikipedia or something.
I’m currently sick and I had nothing better to do yesterday than stay in bed and try ChatGPT (which seems to be going viral at the moment, or maybe it’s just me who’s been out of the loop).
Its capabilities are beyond everything I’ve seen from “chat bots”. It can keep track of conversations and follow-up questions, it remembers context. I had it create a changeling character for me, which it did, a female character named Kaya. I then asked it to come up with a physical description for her, which it also did. This is how it described her hair: “Kaya has long, flowing hair that changes color depending on the light, shifting from deep blue to green to purple and back again.”
I wanted to turn this into a “gotcha”, because while a human might be able to explain why they chose those colors, I was sure the bot wouldn’t be. But this was the answer:
That already seemed impressive to me, but that was nothing. I had it create another changeling for me, and we ended up having this “conversation” (long) about certain aspects of his skills:
TL;DR: Hours later, and we arrived here:
Yes, it knows ChoiceScript. In the hands of someone who actually knows what they’re doing (unlike me), this could be a game changer. In many ways…
Hoo. Boy.
Ayo, that’s a fucking actual AI LMAOOOOOO
It’s as mindbogglingly fascinating as it is scary. I mean, it still spits out a lot of bogus (while sounding awfully confident about it), but I thought it was only image AI that was so far advanced. If you know how to code and are capable of recognizing pseudo-code and nonsense, this can be a great tool. I don’t know any coding languages, but from what I’ve heard from people who can, it’s perfectly capable of coming up with valid and working solutions, especially if you know what you are looking for and able to be precise about it.
I can’t help but wonder what this means for the future of interactive fiction.
Hosted Games has a certain level of quality control via open beta testing, so it should be save, but what about other platforms? What if the market gets flooded with half-heartedly churned out half-assed AI stories, the equivalent to cheap asset-flip steam “games”? (Yeah, this goes for all fiction / novels then, but we’re here in the choice of games forum) Real stories will drown in a sea of noise.
I had the AI create entire parties and a cast of RO’s for me. I had it give me (useful!) tips on how to write engaging romances specifically in the context of IFs. But what if I’m not looking for advice but instead let it do almost the entire work?
And what if it becomes possible to steal existing works and feed them into the AI and have it continue or alter existing stories? Will I - hypothetically, I would NEVER do that!!! - in the future be able to feed it the source code of existing IF and have it write a sequel for me? One that caters to my wishes more than the original ever could? Where I don’t have to wait years for the next part? Will I tell myself that I’m only doing it to bridge the time until the next part? Will I still like the original once I get back to it and realize unlike the AI version, the original doesn’t go in the direction I’d want it to?
I’m worried the question won’t be “Will those stories be good?”, because hell no, they won’t. But I’m afraid they could for many people be good enough.
yup, I’m inclined to agree with this. I think most consumers are looking to scratch their itch, not necessarily on a mission to find some masterpiece.
That’s actually pretty impressive, I’m surprised. I still have my doubts as to whether they’re at the stage where it could track choices well enough to put together a coherent story arc in choicescript. Perhaps with a human prodding it regularly for particular scenes, choices and telling it when to finish up the game it could get there though?
(Edit: I was going to take a look at the site but no way am I giving it my phone number for “verification” grumbles.)
As of yet, it’s definitely not possible, if only because the sessions expire at one point and there’s no save function. It also auto-limits its answers length-wise, and even if you explicitly ask it to create a text with a specific word count, it usually cuts off sooner, but that seems to be a limit set by the developers. You can ask it to continue, though, and it will.
But maybe all this is even still the wrong way to think about the long term implications. Why take the detour via (choice)scripting at all, if you will be able to have it directly create interactive fiction for you without the scripting part. The scripting part is only useful if you want to share a result with others, but if those potential readers prefer their own games too, there won’t be a point anymore in sharing.
It’s already good (I wouldn’t say very good, for reasons I’ll get to) at making procedurally generated text adventures. Yesterday I had it simulate one for me, and this moment, where I tried to bullshit it, stood out to me. Not only did it immediately adjust to my shenanigans, it did so in a way I found genuinely entertaining and fun to read:
What really blew my mind was when I asked it to show a short term and long term goal in the menu and set “leave dungeon” as my current short term and “find kidnapped lover” as long term goal. Not only did it do so, it actually understood the concept behind it: A bit later that dragon joined my party and gave me and my halfling companion the new information that my lover was held captive by an evil sorcerer. We left the dungeon soon after and when I looked into the menu once we were out, it had replaced the short term goal of leaving the dungeon with the former long term goal and set “defeat the evil sorcerer” as new long term goal ON ITS OWN.
Here’s why it was still only “good”: There was clearly no overarching plot and everything remotely creative that happened, came from me provoking it (see example above); “go north” never led outside the dungeon, despite it repeatedly telling me it would do so, I had to directly type “leave dungeon”; Drake (that’s what it named the dragon) decided once outside to leave the party, and said how it had been a long journey together (when he had just joined us); and a bit later it forgot half the stuff that had originally been in the menu.
Thank you for sharing this, it was very informative. I had no idea that writing bots are this sophisticated. Definitely something to keep an eye on.
At the very least, if you’re a game dev who’s more interested in making game systems than stories, you could have them generate simple plots and characters for your game. (Then you could hire a writer to give them more unique dialogue, lore, what have you, but that assumes you care about originality in the first place. But if you’re cutting corners…)
That reminds me–there was a recent controversy with a bot trawling AO3 for fic, which prompted some authors to lock or take down their works. With such a massive database, plus fic tags making it easy to find examples, it’s possible these bots could generate passable fic, even for specific prompts and tropes…
Hmm.
An example of how close the “generated” art is to the stolen.
Makes me wonder what fanfics the text is copied from.
That’s a tricky one. Honestly, I don’t really wanna touch that with a ten foot pole. One could argue, as derivative works, fanfics aren’t truly original content to begin with, and unlike (original) art, could never be commercialized anyway, so no jobs get lost. They’ve always been in a grey area.
I’m saying this as someone who used to write fanfics, is deeply attached to his works and understands why one wouldn’t have their style imitated by an AI. At the same time, knowing the content and quality of so many of those works on there, it is admittedly a bit hilarious to think about. I decided against protecting or removing my own stories from there, but I get those who feel different.
(Having said that, I already made GPT write a “typical slashfic” about the classic Faust and Mephisto. It immediately described Mephisto’s smoldering eyes, introduced Gretchen as the obligatory evil female character incl. “Does she know about us?” concerns, and later introduced MPREG. When I asked if Faust was trans due to his ability to become pregnant as a man, it almost got snarky and explained it was fiction and not real, so it didn’t matter. I’m absolutely not kidding.)
Not really, that website just shows you if your image is included in any training database through image search, not the output of any one model
Good model like stable diffusion will never just trace any image, because of the way it functions and that it has tons of training samples to prevent overfilling