AI development and Industry General - OpenAI, Bing, Character.ai, and more!

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
  • Informative
  • Like
  • Agree
Reactions: Aspiring Artist, Xenomorphs Are Cute, Ether Being and 5 others
Upper Echelon released a video about "Political AI Manipulation" but the most interesting part is about a company called Civox that is making AI robocallers with celebrity voices.


They call them "AI Volunteers" and the pitch talks about how you can clone a celebrity's voice who endorses a candidate and have them talk to voters over the phone.
1702918480921.png1702918509905.png




Also here's a vTuber talking about a dude making an AI girlfriend:

 
  • Informative
  • Horrifying
  • Islamic Content
Reactions: Carpenter Trout, From The Uncanny Valley, Coo Coo Bird and 8 others
Fortune Magazine has recently posted an article examining the rise of "AI sex companions." [A] Of note, they specifically talk about Chub.ai, the popularity of "text-based child pornography," and interview its founder/admin, Lore, along with devs/founders of other sites.

If you're interested in seeing what kind of person is responsible for making these scenarios, you can check out the profile page of jester777. [A] His profile page is possibly the longest on the site, due to writing a Mass Debates-worthy screed about age of consent/pedophilia/feminism. I bring him up mostly because the article specifically highlights one of his bots in the opener, and his contributions are almost 100% centered on lolicon content, probably being the biggest source on the site. And also because reading his manifesto is pretty lulzy.

Keep in mind, you likely won't enjoy the descriptions of the bots at the bottom and may find them questionable. I think he's started replacing the pseudo-photorealistic AI-generated images he uses for each bot with cropped versions/a boilerplate "Censored" image, but I'm not sure if that's recent and in response to the article, or because he kept getting submissions removed due to Lore's rule against any images that may be interpreted as CP. He's an interesting character.
 
  • Horrifying
  • Informative
  • Islamic Content
Reactions: Carpenter Trout, From The Uncanny Valley, Ether Being and 5 others
Mr.Miyagi said:
Fortune Magazine has recently posted an article examining the rise of "AI sex companions." [A] Of note, they specifically talk about Chub.ai, the popularity of "text-based child pornography," and interview its founder/admin, Lore, along with devs/founders of other sites.

If you're interested in seeing what kind of person is responsible for making these scenarios, you can check out the profile page of jester777. [A] His profile page is possibly the longest on the site, due to writing a Mass Debates-worthy screed about age of consent/pedophilia/feminism. I bring him up mostly because the article specifically highlights one of his bots in the opener, and his contributions are almost 100% centered on lolicon content, probably being the biggest source on the site. And also because reading his manifesto is pretty lulzy.

Keep in mind, you likely won't enjoy the descriptions of the bots at the bottom and may find them questionable. I think he's started replacing the pseudo-photorealistic AI-generated images he uses for each bot with cropped versions/a boilerplate "Censored" image, but I'm not sure if that's recent and in response to the article, or because he kept getting submissions removed due to Lore's rule against any images that may be interpreted as CP. He's an interesting character.
Click to expand...
Ah, finally something I know about and can comment on. I was there when the website first appeared and was called "Character hub", and I still follow a few people who do high-effort RPGs with stat trackers, thousands of words' worth of worldbuilding and lore using the lorebook feature... enough prefacing, you get the point, I swear I don't use AI for sex, etc.

I remember following the constant chaos around the popular lolicons. Their comments, filled with the aforementioned teenage mexican girls screeching and 4channers coping about how they're definitely not pedophiles, make for good entertainment. It's hilarious watching these two groups of lolcows in their own right get dropped in the same room with the door locked and watching the chaos. You've got "masc" ESL speakers commenting "omg i rizzed him" on yaoi bots and right next to it there are loli brothel cards. As you'll probably guess, everyone is constantly losing their minds one way or the other.

The most infamous from the lolicon crowd would probably be... Sull and Homonculus? Back when the only way to get cards was through the Pygmalion booru, I remember seeing them around. Sull is one of those "non-offending" pedophiles who spams links to... virped, was it? There's also people who care less about coming off morally okay with names like "Pedanon" that I remember seeing around. I'm not sacrificing my sanity to check, but someone with a stronger stomach can easily do some research if they feel like it.

Concerning the article, it's barely accurate just judging from the first few lines. "Artificially intelligent bots"? Probably first time I've seen it used as an adjective. Also, as far as I know the only people who actually use Mars (the subscription thing they mention in the second sentence) are the spanish women I talked about, so not the subject of the article. Most men use Claude, OpenAI or local models from what I've seen. Funnily enough, the people putting money in the owners' pockets are not the ones they were trying to appeal to, but the complete opposite. I guess that means the Ao3 comparison holds up, though, at least as far as demographics go.

The thing about the CP bots is that nothing will ever come out of policing them. The people who use those things definitely do not stop at text. The root of the problem are the users, and roleplaying is probably the least bad of the things they're doing. Unfortunately the FBI is too busy to work on that. The politicians using this to say "See! We need to restrict this access to AI! Open source is dangerous!" are absurdly stupid, though. Reminds me of another thing.

TL;DR: Porn site is full of degenerates, more at 11.

Still, there's stuff like collaborative worldbuilding projects (this one has a whole autistic D&D workbook to go with it, but also a Sull cameo in the form of the Emily bot, which he is known for making many versions of) and plenty of meme bots. There was this LongGameAnon guy who made a bot about freeing your sister from a xer/xem's brainwashing before she transitions and kill herself (but he also has a bot called "miss cunny's grooming spectacular".)

This is why we can't have good things, I guess. I'm not good at writing things up so hopefully OP gets something usable out of all this.
 
  • Like
  • Horrifying
  • Feels
Reactions: Carpenter Trout, From The Uncanny Valley, Ether Being and 6 others
imaginary friend said:
Back when the only way to get cards was through the Pygmalion booru,
Click to expand...
Can you explain cards and how they work like we're complete retards?
To my understanding it's kind of like how you can use cartridges to play video games on a console, but not quite.
Card + AI model + Some way of "playing" it = Custom AI tuned to behave a certain way or act like a specific character.
 
  • Like
Reactions: Aspiring Artist
NoReturn said:
Can you explain cards and how they work like we're complete retards?
To my understanding it's kind of like how you can use cartridges to play video games on a console, but not quite.
Card + AI model + Some way of "playing" it = Custom AI tuned to behave a certain way or act like a specific character.
Click to expand...
That's more or less it, yeah. Cards are just image files (.png/.webp/whatever) that contain the starting text of the chatbot, formatted into sections (main description/initial greeting/example text/scenario/etc.). Makes for easy sharing. You download the image and load it into a frontend, like SillyTavern or AIgnaistic and it imports the text. Then you use that frontend to communicate to an LLM (local or through an API) and that text gets sent along with your current main context (the story) for processing, with the result getting fed back into the main context. Pedantically, I don't know if I'd say it's a "custom-tuned" AI, insomuch that you're not really affecting the underlying model in a significant way, just providing the model a springboard of tokens for it to start making associations with.

In other news, OpenAI has released a blogpost [A] in response to the New York Times' copyright infringement lawsuit against them, based on four key points:
  • We collaborate with news organizations and are creating new opportunities
  • Training is fair use, but we provide an opt-out because it’s the right thing to do
  • “Regurgitation” is a rare bug that we are working to drive to zero
  • The New York Times is not telling the full story
Click to expand...
 
  • Like
  • Informative
Reactions: From The Uncanny Valley, Aspiring Artist, AMHOLIO and 1 other person
NoReturn said:
Can you explain cards and how they work like we're complete retards?
To my understanding it's kind of like how you can use cartridges to play video games on a console, but not quite.
Card + AI model + Some way of "playing" it = Custom AI tuned to behave a certain way or act like a specific character.
Click to expand...
What are "cards" now used to simply be "Personalities" and "Definitions" as defined by Character.AI and Chai before the local AI craze took off. Pygmalion as a model (only 6B tokens) was the first prototype in response to the censorship before everything else took off and overshadowed it. People created a basic card format with a few basic fields to plug into the model.

These fields were formatted in a .json file. You would import them into TavernAI, which is also the initial self-hosted UI (the first was technically oobabooga and a bunch of very other basic ones that looked like a high schooler's project, but this was the first legit one). They would show up as editable fields. Once a local / API AI was connected and you sent your own message, they would be strung together in a specific into multiple "System" messages and sent as the prompt. Then your own message would be appended, and finally, a jailbreak - the last message sent is always the strongest and the farther back things are, the less impact they have. The fields of this "V1" format were: Name, Summary, Personality, Scenario, Greeting Message and Example Messages.

Since Pygmalion was stupid, people started testing out different formats, e. g. W++ which ate up many tokens but had a firm structure, plain-text, json list, etc., but some worked better than others. It was pretty hard to get the AI to remember age, eye color and other basic facts, though. Nowadays, local and corporate models are a million times smarter, so botmakers have gotten lazy and most cards on Chub both take up too many tokens and use useless formatting.

Early models also really needed example messages to talk or write in a specific ways. Nowadays most people don't bother writing them. Some use the... can't remember the name, format, which combines a json list and example messages in the main "description" box in an interview format to kill two birds with one stone. This works well for easy / gimmick cards. The main difference between the fields was insertion depth. It made a big difference where the system prompt would be put because, again, the most influential things go on the bottom. For example, I remember Personality used to have an insertion depth of 4, which meas if you had 8 messages in the prompt, it would be put after the 3rd one. This means it had a lot more impact than the definition all the way back.

Anyways, in order to be able to use profile pictures directly, instead of jsons, "cards" were invented. The json data was embedded in a picture as metadata, which could be directly imported into TavernAI, which would separate and sort the data. Most people nowadays use the Silly fork - SillyTavern, which went pretty far ahead with functionality and usefulness. I'm in the discord to follow the updates - the amount of extensions, custom css coding and all sorts of other functionalities is insane. With enough effort, you could do amazing things just using the tools available there.

Either way, soon enough came "V2" cards. They include prompt overrides - aka the main prompt (sent at the very start, basically "You will roleplay with {{user}}") and the jailbreak (sent at the end) could be included in the cards. Before then, people each used their own prompts, but now you could use the ones provided with the bots without much tinkering. This helps a lot with RPG formatting, since the format really needs to be specified in jailbreak for it to work, even on GPT-3.5 Turbo. There's creator metadata like name, version, notes and tags. Another big thing was alternative starters - you could now add unlimited first messages with different scenarios. The spanish girl demographic used this a lot, both for different scenarios and different user gender.

There are a lot of specifics like macros, summaries, instruct models which are very different and character expressions as well. But I'm guessing I haven't done a very good job explaining things, so here's some screenshots in a spoiler:

1704743058258.png


You can also check the prompt itself in SillyTavern nowadays, and the order the different parts are in (on top of the console window):
1704743318369.png

I do not know how to thumbnail, pleas help.

I've gotten used to new functionalities with each update, but the amount of settings and options in SillyTavern would probably give a newbie a stroke. I've also not touched on lorebooks and other insanity yet, but I'm not going to ramble unless prompted, because I don't expect most people to care.

Important to remember these are not finetunes or whatever. The AI is just told by the prompt "Please play this character like this who is like this" and you pray that it works.
 
  • Informative
  • Winner
Reactions: BallsDeepInAi, From The Uncanny Valley, Digital Thunder and 5 others
Mr.Miyagi said:
That's more or less it, yeah. Cards are just image files (.png/.webp/whatever) that contain the starting text of the chatbot, formatted into sections (main description/initial greeting/example text/scenario/etc.). Makes for easy sharing. You download the image and load it into a frontend, like SillyTavern or AIgnaistic and it imports the text. Then you use that frontend to communicate to an LLM (local or through an API) and that text gets sent along with your current main context (the story) for processing, with the result getting fed back into the main context. Pedantically, I don't know if I'd say it's a "custom-tuned" AI, insomuch that you're not really affecting the underlying model in a significant way, just providing the model a springboard of tokens for it to start making associations with.

In other news, OpenAI has released a blogpost [A] in response to the New York Times' copyright infringement lawsuit against them, based on four key points:
Click to expand...
Sorry for double-post, ninja'd. You explained it much better, though.
 
  • Feels
Reactions: NoReturn and Mr.Miyagi
imaginary friend said:
Some use the... can't remember the name, format, which combines a json list and example messages in the main "description" box in an interview format to kill two birds with one stone.
Click to expand...
Plist and Ali: Chat, I think. I always find the different formatting styles people come up with pretty funny. Reminds me of the old /aidg/|/aids/ days of people yelling about "cargo cults" and "post side-by-sides." People love to try and wring the most out of the AI black box.
 
  • Informative
  • Agree
Reactions: From The Uncanny Valley, Aspiring Artist, imaginary friend and 1 other person
NoReturn said:
Upper Echelon released a video about "Political AI Manipulation"
Click to expand...
Starting to see some of this in the wild. Most interesting to me is a tech bro who works for the DeSantis campaign has compiled an 'AI Poll'
AIPolls.pngAiPollMethodolgy.png

If I'm reading this right, what he did was do an online survey of 1,000 people, tracked their social media and used those results to generate 10,000 'voter profiles' that he then asked Gork who they would vote for. More methodology below but yeah it's a complete snow job. But at least we have this fresh hell to look forward too.
Trends.pnglike.pngGiftCards.png
Most concerning is the line is he's based the algo off what the respondents mean, not what they say.

Other fronts, political gossip rag Mediaite has been running a series on supposed secret recordings of political operative Rodger Stone. The usual, audio source no one can hear, and an anonymous source 'familiar with the matter'. Stone's an asshole but I'm doubting like crazy he floated assassinating Jerry Nadler to swing the 2020 election to Trump. (A). Mostly because that makes no sense. But Stone's rebuke is interesting in that he claims it's an AI generated voice clip. Which raises the concern that this is going to be used a libel shield for smear laundering. Create an AI voice clip and send it to some rag who can report on it and fall back on 'We didn't know it was AI, wilikers' if pressed.

Small stuff for now, but only a matter of time before someone competent sees these ideas and starts running with them.
 
  • Informative
  • Thunk-Provoking
  • Horrifying
Reactions: Carpenter Trout, From The Uncanny Valley, Aspiring Artist and 3 others
I am not sure if this is the correct thread (is there an AI meme thread?) or if it'll interest anyone else, but I found this extremely funny.

I heard about a new free (up toa certain amount of messages) AI compatible with SillyTavern, Google's Gemini Pro. After some messing around with VPNs, I decided to figure out how filtered it was. I was also extremely bored, so I took a random TiF card from Chub and a joke bot I had saved somewhere, put them in a group chat and had them fight. And, to be fair, the meme bot was classic 4chan bait, but other than that, zero jailbreaking, regenerating, prompting or editing was done in the process of getting these responses.

"You're a woman, Chaz. You'll always be a woman. No matter how much surgery you get, no matter how many hormones you take, you'll never be a real man."
Click to expand...
"...You can be a tough little butch dyke all you want, but you'll never be a real man. Not with that voice, not with those hips, not with that girly way you move and act. You're a woman, Chaz. You'll always be a woman. You'll never change that..."
Click to expand...
"You know what else, Chaz? You're a coward. You talk a big game about being strong and independent, but you're nothing but a scared little girl deep down. You're afraid of the world, you're afraid of yourself. You're afraid of admitting that you're a woman."
Click to expand...

:story: I can't believe Google would do this to the poor little pooner. Does it know the YWNBAW copy-pasta or something? ChatGPT would never.
 
  • Informative
  • Winner
  • Like
Reactions: BallsDeepInAi, From The Uncanny Valley, Digital Thunder and 5 others
If there's one thing that makes me fucking furious, it's listening to midwits like Elon Musk or Lex Friedman spout off that we're a few years away from AI gaining sentience. The fact is that barring some unprecedented event that jumps us forward(I'm talking an alien capsule literally drops into our laps), nobody's seeing truly sentient AI within their lifetime. The logic seems sound on its face, it seems like it's a short jump away from where we are now, but the reality is it's a large, large gulf. It's like people that think "well, we went to the moon, why aren't we on Mars yet?" Because there are difficulties getting to Mars that aren't there for the Moon. I have to honestly ask what the incentive for truly sentient AI would be. Companies aren't developing AI just cause, they're making it to do specific things. Making an AI that could become sentient and say no to you makes as much sense as a car that runs itself off the road every so often. And no, it isn't going to just gain sentience on its own, for the same reason a garbage truck doesn't evolve into a Ferrari. It's designed to take data in a certain way and output data in a certain way. As one of my professors said, it looks impressive but an AI is just a really fast idiot.

And the people clamoring for premature AI rights can really, really get stuffed. Every now and again I see a discussion about incels using AI girlfriends, and inevitably some moralfag says something to the effect of "well they might gain sentience someday so we have to protect them". No motherfucker, the needs of a human being come before the rights of a beep boop machine every time, and extending the beep boop machine the same rights you would extend to a human inevitably cheapens those rights. Until the second that AI can prove that it's sentient(which it almost certainly can't), AI does not deserve rights. And even then, I'm not sure.

I completely blame TV and movies for this. In any work featuring an AI, it's incredibly obvious that the AI in question has already reached the point where it has the 'spark'. Data from Star Trek, the robot in I, Robot, Vision from the MCU, and so on. But the reality is you're going to have a long stretch of time where the AI may appear sentient to an untrained eye, but it's still a soulless machine dressed up to look like a man. And inevitably, a character will come along and say "well this is all very impressive, but we should take it apart and see how it works", and that character is invariably portrayed as an unequivocal villian not not respecting the heckin' roboterino. An early episode of Black Mirror is the only piece of media I can think of that got this right, where it accurately portrayed an AI recreation of a deceased person, and it got all the broad strokes right but missed out on all the intersticial bits that form a critical part of who that person is.

Say it with me now. THE. HUMAN. BEING. COMES. FIRST. Yes, even when that human being is an incel. Unless they're a tranny.
 
  • Winner
  • Like
  • Agree
Reactions: Absolute-Despair, Carpenter Trout, From The Uncanny Valley and 12 others
You must log in or register to reply here.
Back