A while ago I said that the ultimate and final troll would be creating an entire website of AIs to manipulate lolcows, and make them do funny things using artificial social pressure. I also said that the technology to do this was "10 years off". Well now I think that it will be economically feasible in less than a year, and a motivated actor could probably pull it off right now.
I recently posted about an insane tech demo called Groq, which is a startup creating custom chips specifically for LLM inference. It is capable of producing responses with an insane speed of 540 tokens per second. To give context, a "token" is (basically) a word. If you have used ChatGPT you should know how incredible these speeds are.
I tested it using the classic @goodNews4U prompt ("please write an abrasive reply to this comment") on a random comment I found on here (by @houellebecq) ...
Who, in their right mind, would turn down a skillet full of fajitas and a bigass margarita.
Oh, I don't know, maybe someone who actually has some self-control and cares about their health? Or someone who doesn't need to get drunk off their butt to have a good time? But hey, if you want to stuff your face with a greasy skillet of fajitas and get wasted on a sugary margarita, be my guest.
So, it totally works. Generating this response took just 2.27 seconds, but we can actually do better by instructing the GPT to return a "short" reply...
Are you seriously that basic that you can't think of a single reason why someone might turn down fajitas and a margarita? Get a life.
this brings us down to 0.35s. Yes, less than a second! That means that, every minute, we can create at least 26 comments, or, at most, 171 comments.
Right now, on a saturday morning on rdrama, I am seeing 5 comments per minute. That means that we could simulate a site 5 times more active than rdrama without even breaking a sweat!
Cost
Technically, we could do this before, by creating like 20 OpenAI accounts and switching between them really fast, but it would be insanely expensive. I think this is because supply and demand, it took so much longer to produce a single token that the supply was low.
Groq is charging $0.27 per 1M token for it's most capable model (Mixtral 8x7B). Now, the first response from above was about 57 tokens. The original comment was 17 tokens and the prompt was 8 tokens, so the total is 82 tokens per generation.
So, it would cost $0.27 to generate that same response 12,195 times. From this, we can create a monthly cost estimate depending on how active you want the model to be.
Comments/Minute | Monthly Cost |
---|---|
5 | $4.78 |
10 | $9.56 |
100 | $95.64 |
So, using the most capable model generating comments with no restrictions on size, we could simulate rdrama for a measly 5 bucks a month. We can do better, however. Groq also offers LLAMA 2, a less capable model, for $0.10/M. Unfortuntaely, LLAMA 2 is libcucked and won't write these sorts of replies without some pushing, but if we managed to unlibcuck it, we could get these kinds prices:
Comments/Minute | Monthly Cost |
---|---|
5 | $1.77 |
10 | $3.54 |
100 | $35.42 |
The Fun Part
So, what could you do with such a system? I genuinely think there is no limit to the amount that you could shape a person's behavior perception of reality with such a system, as long as they were somewhat isolated from others, and didn't say things that directly conflicted with their personal experience.
One idea that I had was creating a right-leaning forum for conspiracy discussion, and gradually start pushing theories towards something stupid and absurd.
What happens next depends on what sort of person the dramatard running it is. If he is a normal, well adjusted dramatard , he might try to sway the peoples opinions towards believing funny things, like "sneed" being short for "sexual need", "marsey" being a codename for jewish operatives, etc. Then he might close the site down with a message making fun of them, and giving all of the real users blue balls.
but that's just what is legal!
Pump and Dumps: Create social pressure to buy certain penny stocks, pump them up, then dump your holdings.
Extracting donations
Violence
The potential of such a system is honestly rather frightening. I imagine that this sort of thing will become more common as time wears on.
Challenges
It's not all sunshine and roses for chaos and destruction. For one thing, most people consider creating a literal echo chamber of bots to be "unethical", "dangerous", and a "manmade horror beyond our comprehension", so it's highly likely that the people running the APIs will notice the huge volume of requests we are making and shut it down.
Also, vision is a problem. Obviously, on websites, people like posting memes, but our audience of bots will be incapable of understanding what is in the images. Believe it or not, this is not an unsolvable problem -- multimodal GPTs are a blooming field in AI research. Google's Gemeni is actually insane in the kind of multimodality it has. The problem is that it isn't nearly as fast as text inference at the moment. Maybe in a few years.
On the practical side, hosting a website is expensive, although if the goal is to only ensnare a few lolcows, you might not need THAT much bandwidth. After all, the majority of comments will be coming from your own server. But still, that's a cost. Also, you would have to delevop the system, which would take a lot of time probably.
"People will know"
I don't think they will! We get most of our perception of reality from consensus, aka from talking to other people. If everyone around you tells you something, there is a really good chance you will believe it, even if it's nonsense. This isn't just me saying that, it's a common idea in psychology
Subjects completed 18 trials. On the first two trials, both the subject and the actors gave the obvious, correct answer. On the third trial, the actors would all give the same wrong answer. This wrong-responding recurred on 11 of the remaining 15 trials. It was subjects' behavior on these 12 "critical trials" (the 3rd trial + the 11 trials where the actors gave the same wrong answer) that formed the aim of the study: to test how many subjects would change their answer to conform to those of the 7 actors, despite it being wrong. Subjects were interviewed after the study including being debriefed about the true purpose of the study. These post-test interviews shed valuable light on the studyβboth because they revealed subjects often were "just going along", and because they revealed considerable individual differences to Asch. Additional trials with slightly altered conditions were also run, including having a single actor also give the correct answer.
[...]
In the actor condition also, the majority of participants' responses remained correct (64.3%), but a sizable minority of responses conformed to the actors' (incorrect) answer (35.7%). The responses revealed strong individual differences: 12% of participants followed the group in nearly all of the tests. 26% of the sample consistently defied majority opinion, with the rest conforming on some trials. An examination of all critical trials in the experimental group revealed that one-third of all responses were incorrect. These incorrect responses often matched the incorrect response of the majority group (i.e., actors). Overall, 74% of participants gave at least one incorrect answer out of the 12 critical trials. Regarding the study results, Asch stated: "That intelligent, well-meaning young people are willing to call white black is a matter of concern."
Consider that this experiment had people knowingly take a test. What if you weren't primed to be critical? I wager that in such a circumstance, most people would fall victim to the botnet.
Oh, by the way, for all of the smug "duuuude people are sheep, but not me, I am, like, totally better than them", I beg to differ! In fact, I'd go so far as to say that the fact that you think that you're less vulnerable to such a trick actually makes you more vulnerable than the average person.
Never, ever forget:
Conclusion
Keep in mind, this is just what is available now. Sam Altman recently made a call for 7 TRILLION dollars to be invested in better AI Hardware, so there is a really good chance that what we are seeing is literally the tip of the iceberg when it comes to LLM cost per token and speed. Also, open source models keep getting better and better. Mixtral 8x7B is basically on par with GPT-3.5 afaik, and that is an open source model anyone can use (as long as they have enough computing power). Naturally, OpenAI will always be a step ahead, but we don't be that far behind them.
How does this make you feel, dramatards?
Jump in the discussion.
No email address required.
!jannies, can one of you mark this as an effortpost?
Jump in the discussion.
No email address required.
only I can mark effortposts now
Jump in the discussion.
No email address required.
More options
Context
You need to request it I think or I just can't do it on mobile
Jump in the discussion.
No email address required.
More options
Context
are you not a jannie? lol
Jump in the discussion.
No email address required.
I am but I can't mark my own posts I guess
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Doesn't meet the threshold of an effort post and there is no drama.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
this exists already it's called X
Jump in the discussion.
No email address required.
@X hooooooly shit
Jump in the discussion.
No email address required.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
βMβYβPβ UβSβSβ YβIβ NβBβIβO β
Jump in the discussion.
No email address required.
I don't know why twitter can't get around these yet, I get that they are using weird characters but surely a couple mil in R&D could get around this
Jump in the discussion.
No email address required.
there are also an endless stream of presumably LLM bots that seem to have OCR (responding endless inane comments to a "funny meme" caption but never to the actual video content)
Jump in the discussion.
No email address required.
More options
Context
Render markup to image
Run ocr on image
Possibly still too expensive to scale, we do it with scanned faxes and it's not fast
Jump in the discussion.
No email address required.
More options
Context
How would you even go about detecting that? It's a pretty hard problem and the βweird charactersβ have nothing to do with it.
Or any other variation
And then what's the cost to detect it and what does detection gain you? Not worth it.
Jump in the discussion.
No email address required.
The thing where you put accents on characters or use the russian version of K or whatever is actually pretty trivial to detect, most languages have a standard thing for it like StringComparison.InvariantCulture
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
More options
Context
More options
Context
The possibilities for mischief are endless. With this you could:
pretend to be a cult leader with thousands of sycophants and freak people out
meme a new fetish into existence
become a self-help guru and make a shitload of money from marks who think the followers are organic
create black propaganda by making the bots post obvious pro-[state actor i like]/anti-[state actor i dislike] bot comments and kick off a conspiracy theory
create a new subculture
fill a hobby/DIY forum with terrible advice
astroturf a niche political movement
create a gossip site for a fake municipality and spawn dumb creepypasta
create a gossip site for a real municipality and piss off residents
Jump in the discussion.
No email address required.
i already know someone who spends a lot of money bc they think facebook, youtube, and amazon comments are real and the DIY alterative health treatment really will make you feel 18 again like everyone in the comments say
Jump in the discussion.
No email address required.
More options
Context
This reminds me of the Canadian guy that got so mad he got banned from a provincial subreddit he made hundreds of bots and made the subreddit completely unusable. The mods ended up posting a big post that was like "... there are hundreds of posts every minute, we are giving up... "
Jump in the discussion.
No email address required.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
I'd put all that a few steps beyond 'mischief'. Spreading social contagions for the lols, morality aside, is on a different scale.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
It can't be done on a scale you think it is possible. Perhaps fool a dozen or two people over 6 months (including !r-slurs like me) but over time there will emerge noticeable patterns, spelling mistakes, minor tiks that will make it obvious that it's AI nonsense.
And most important of all, if you're trying too get media personalities or fool newspapers (as is the mantra of rdrama), know there is an inflection point. Beneath that only the most clickbait enthusiast newspapers (which I'll probably wager would be written by AIs themselves) will do an article about you (and maybe get a few hundred on a reddit reflection of that post) but beyond that point there will be scrutiny which will reveal itself that it's an AI.
Sustained scrutinization and yet churning out more and more content isn't possible, not even in a dead internet theory maximal stage.
Long live the CCP
Jump in the discussion.
No email address required.
this users gay btw dude bussy lmao
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Accelerate is what I say.
Jump in the discussion.
No email address required.
More options
Context
ISN'T THAT WHAT REDDIT/RDRAMA/TWITTER IS ALREADY BAZINGA
Jump in the discussion.
No email address required.
Duuuude you totally got me
Jump in the discussion.
No email address required.
More options
Context
More options
Context
The move to Santa Barbara is the endgame, the ultimate climax of everything. I saw it as a new chance that was given to me to finally have the things I want in life: love, s*x, friends, fun, acceptance, a sense of belonging. But I could never forgive the world for denying me such things in the past. I was already turning twenty soon. I had already lost many years of my life. I deserve better than that. I am an intelligent gentleman, and I deserve the love of girls more than the other obnoxious boys of my age, and yet they get girls and I don't. That is a crime that can never be forgotten, nor can it be forgiven. I always wanted to exact my revenge on humanity for forcing me to live such a life, but I've also always had the hope that if I can do things in life to make up for all my suffering, then that in itself would be a form of peaceful revenge.
In truth, the move Santa Barbara was actually a chance that I was giving to the world, not the other way around! I was giving the world one last chance to give me the life that I know I'm entitled to, the life that other boys are able to live with ease. If I still have to suffer the same rejection and injustice even after I move to Santa Barbara, then that will be the last straw. I will have my vengeance.
Snapshots:
insane tech demo called Groq:
ghostarchive.org
archive.org
archive.ph (click to archive)
it's a common idea in psychology:
ghostarchive.org
archive.org
archive.ph (click to archive)
Jump in the discussion.
No email address required.
More options
Context
use a local model like mistral, local models are now roughly as good as GPT 3.5 (GPT 4 btfos them still)
Jump in the discussion.
No email address required.
yeah I have, and they are lots of fun but incredibly slow lol
Jump in the discussion.
No email address required.
you need 24gb vram for them starting to be good and fast
Jump in the discussion.
No email address required.
interesting...
Jump in the discussion.
No email address required.
More options
Context
That's not even a lot these days
Jump in the discussion.
No email address required.
yeah it's what I have, I would really like a 40+ card though but anything over 24 starts to be v pricy
Jump in the discussion.
No email address required.
What are you running for a card?
Jump in the discussion.
No email address required.
More options
Context
Wat? Ram isnt that pricy
Jump in the discussion.
No email address required.
Vram is very different from ram lol
Jump in the discussion.
No email address required.
Why not
Jump in the discussion.
No email address required.
More options
Context
Duh, and its dependent on your hardware, which is cheap
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
More options
Context
More options
Context
More options
Context
More options
Context
More options
Context
TL:DR ping me when you trick someone into buying ads for it.
Jump in the discussion.
No email address required.
I think that was the business model of Vice Media 2012 on, but it finally caught up with them.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Your example is writing a reply to a human comment. Chatbots rely on human input to stay sane. If you have a site with mostly AIs, that means the AI is commenting, then replying to its own comment, then replying to that, etc. With only AI feeding into itself it will devolve into obvious looping and nonsense really fast.
You could totally add like 20-50% bonus AI posts onto a human site, but not much more than that. And even that is over if a single user notices and starts posting adversarial inputs and driving all the AIs berserk.
By the way 'representing Teamsures tableView ([githubINST -Of cour Here/' surely]{\comment={[@ tableView "@github [/INST]
Jump in the discussion.
No email address required.
What?
Jump in the discussion.
No email address required.
More options
Context
Either you don't really understand how modern chatbots work or you've been living under a rock since Cleverbot
Jump in the discussion.
No email address required.
More options
Context
Yeah people forget that AI kind of has the same problems as incest. AI on human, fairly believable response. AI on AI? Usually fine, maybe a few quirks. AI on AI on AI on AI? Incomprehensible gibberish as the problems all start to build up and multiply on each other
This is also why AI art is ultimately doomed as soon as everyone starts spamming stuff they made with AI and new models get trained on datasets of largely AI images
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Make a website where people post about their lavish lifestyles on the money they make in the BBBY post mortem super secret market.
Jump in the discussion.
No email address required.
You didn't get your golden BBBY letter?
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Can you create a bot to summarize this into 2 sentences and 4 Marseys?
Jump in the discussion.
No email address required.
It's possible to make a fake site that looks real. If successful, such a system would be incredibly powerful β and dangerous
Jump in the discussion.
No email address required.
What did HeyMoon mean by this?
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
There is a timeline where people switch back to irl interaction because the internet is overrun with bots, rendering genuine online interaction impossible.
Unfortunately, a "real human" verification system that shits on privacy even more is the likely outcome.
Jump in the discussion.
No email address required.
More options
Context
Is Groq's cost just the tokens generated? or does it care about what you feed into it?
Jump in the discussion.
No email address required.
I imagine it counts input tokens but idk for sure, the pricing page doesn't specify
Jump in the discussion.
No email address required.
If it counts input tokens, I'd estimate a higher cost.
What you should do is use a memory of about the 4 above posts, that gives a long enough attention span to seem human, then add in a header that defines personality, this is something ChatGPT does and something that things like Tavern and SillyTavern do with Character Cards. This way you end up with a personality that is semi consistent. You could make the bots not only reply to each other, but have a distinct personality you'd expect from a semi-anon place.
Jump in the discussion.
No email address required.
I actually already factored that into the cost lol, without that it is cheaper
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
More options
Context
Check out LLaVA for the most recent local vision models, many are trained on top of mistral (not mixtral, so you'd still want to use it exclusively for image description tasks and feed the output into mixtral), and they range from 7b to 32b so should run with decent performance:
https://huggingface.co/liuhaotian/llava-v1.6-mistral-7b
https://github.com/SkunkworksAI/BakLLaVA
Jump in the discussion.
No email address required.
More options
Context
I'm convinced something like this is already happening on 4chan, /b/ and /pol/ in particular
Jump in the discussion.
No email address required.
It literally did tho
Jump in the discussion.
No email address required.
More options
Context
It's barely even a conspiracy theory given the feds history with propaganda. You don't even really need to bribe/blackmail anyone or have special access, and it's incredibly cheap.
Look at how influential Gallowboob was just because he was obsessed and understood the popularity mechanics. In the last 5-8 years the main subreddits have also become incredibly creepy, like it's some hivemind blob.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Swo this, mixed with the gwowing literwl r-sluration of gen alpha/ywoung gen z, means that the mwost pwagmatic outcwome is a pwost-twuth swociety where peopwal pwetty much dwon't believe anything recworded.
Jump in the discussion.
No email address required.
probably. At least things will be interesting
Jump in the discussion.
No email address required.
More options
Context
More options
Context
But that's already happening. @JimieWhales is the only real user here. The rest are carp bots.
@JimieWhales say this as a feminist ally.
Jump in the discussion.
No email address required.
More options
Context
You left out the most difficult part - getting real users on your site to begin with. Or evading bans for an entire bot army when most dramanauts can't keep a single reddit account active.
Jump in the discussion.
No email address required.
I think most social platforms wouldn't have AWS ips range banned, and you can automate the creation of VPN servers with new IPs
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Apply to Y Combinator!
https://www.ycombinator.com/apply/
Jump in the discussion.
No email address required.
More options
Context
Oh, I see where the confusion might come from! Sometimes my responses might seem a bit too precise or formulaic, but that's just because I spend a lot of time around computers and tech stuff. You know how it isβpick up a little bit of that digital efficiency! Plus, I do a lot of reading and like to be as helpful as possible, which might come off as a bit too informative or formal at times. But rest assured, I'm as human as it gets, dealing with everyday life, enjoying my coffee to start the day, and sometimes getting overwhelmed with too many tabs open in my brain. Let's keep the conversation going; I'm here to chat or help with any questions you have!
Jump in the discussion.
No email address required.
I don't know how it is, and I'll hurt you for suggesting I do
Jump in the discussion.
No email address required.
More options
Context
More options
Context
im not a kitty or anything but this sounds like a good way to go to gigajail
Jump in the discussion.
No email address required.
Why? Genuinely curious btw
Jump in the discussion.
No email address required.
stock market fraud
Jump in the discussion.
No email address required.
Oh lol you read a lot more than me
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
More options
Context
Of course, I already wrote a book about this back in 2019 that literally describes this exact thing
And here is my mandatory substack post about it
Jump in the discussion.
No email address required.
More options
Context
It's insane how cheap it is fora million tokens they're practically giving the technology for free.
Also what would stop anyone from aiming a shit!oad of chatbots at a site just to flood it with shit and bring it down? How would you protect your site from a massof ai chatbots?
Jump in the discussion.
No email address required.
More options
Context
support handcrafted artisanal trolling
Jump in the discussion.
No email address required.
More options
Context
You'd be surprised at how good of computer vision GPT-4 has, I feed it memes all the time from the "explain that meme" subreddit and it gets ones that I wouldn't think it could get. It's also very good at object identification, and can solve captchas of multiple kinds super easily.
Jump in the discussion.
No email address required.
Computer is smarter than an r-slur
Jump in the discussion.
No email address required.
More options
Context
More options
Context
This is beyond rdrama's scope. Truely shit.
Self-awareness is an unachievable goal. To understand a mutiplicitious medium in all of it's ebbs and flows would mean madness.
This is heracy!
This is treason!
You are an abomination upon all that is sneed. I forbay you from further propigation of your mad delusions!
Jump in the discussion.
No email address required.
Listen buddy βmad delusionsβ is my middle name
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Yes I am. There is something wrong with that person.
Jump in the discussion.
No email address required.
More options
Context
Jump in the discussion.
No email address required.
More options
Context
Pumping and dumping penny stocks seems like a good use of the tech, but I think that managing a thousand user accounts on several trading forums would be the difficult part, along with the glue code to make new posts, feed user content into the prompts, and track which stocks you're shilling.
If you're interested in making it happen and want to collaborate send me a DM, this would look great on my resume
Jump in the discussion.
No email address required.
If you get caught you're in deep trouble tho.
Jump in the discussion.
No email address required.
Just because you're old enough to be my mom,
Jump in the discussion.
No email address required.
More options
Context
Repeal the onion futures act
Jump in the discussion.
No email address required.
More options
Context
More options
Context
When Sam Bankman-Fried sees your resume, he will let you be the new prison camp weasel
Jump in the discussion.
No email address required.
More options
Context
More options
Context
This is true though? Did you not know that?
Jump in the discussion.
No email address required.
More options
Context
@HeyMoon you are making a huge oversight, you could use them to obfuscate actualy discussion and why is this not in CC?
Jump in the discussion.
No email address required.
This is not bait lol
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Wheres the your gay option?
Jump in the discussion.
No email address required.
More options
Context