But we can tell it has meaning through external referants. Like, give them a problem, both have partial information, and they need to coordinate to solve it. If the communication is nonsense they couldn't coordinate. But if they can coordinate, then the seeming nonsense has meaning.
@@adamnevraumont4027It "having meaning" is completely irrelevant to the ENTIRE WORLD besides those 2 bots. So no, it doesn't have meaning, it's gibberish.
A single slot for 8 different words can encode 8 different meanings but a string of 20 words can potentially code for a quintillion different meanings (it's almost certainly much less than that and languages are useful because they don't mean something completely different if you change a single word but still)
When I was younger, I had a theory that every word could be distilled down into a more basic form ("elephant" > "animal", "computer" > "machine", etc), and then that word could be disilled down further... eventually getting to just the words "yes" and "no", and yet it would STILL be possible for people to have extremely rich and robust discussions with one another using those two words. Little did I realize I had discovered binary, and how data is transmitted between computers using that.
I seem to remember this one. The AI were told to try and convince the other AI to give them some of their limited points, and eventually they just started spamming the words that were more convincing at eachother. It's not really a language. "Give me orange" tier.
This is a troublingly real form of communication, one humans have put significant amounts of time to both study and master. I believe the modern terms is "vibes-based". Think of the archetypical corpo making sure to hit every buzzword, the politician yelling all the things that resonate with their audience, and so on, or even your high school lessons on using active statements instead of passive ones. All rhetoric custom-tailored to prioritize "power" instead of (and sometimes at the expense of) substance. The only difference is the AI has a slightly different concept of which words are 'powerful' to it, and it's doing so in a code that arguably isn't English.
So i actually read a few research papers on this phenomenon for school and what i read took a different analysis! Basically it explained that we intruct an LLM to interact with another LLM, but once they get into the pattern of gibberish it's literally gibberish its not their own language. LLMs are programmed to use training data first, then the current conversation as priority. Sometimes they make spelling mistakes due to something in their training data being misspelled or an error, either way, once the other LLM responds it uses the spelling mistake as well until both of them devolve into gibberish. Thanks for coming to my Ted talk
Any true definition of AI has not been realized.We have somewhat advanced algorithms , but these are still very limited programs. None of them are capable of advancing beyond what they are programmed to do
@@superjaylive yep! That's exactly why I use LLM instead of AI because it can be misleading which is why, in my opinion, some people come to conclusions like the LLMs making their own languages, when an LLM can't actually imbue meaning. Nothing an LLM says has meaning, it's just mathematically calculated which words/sentances should be written based on how humans have trained it.
"way more intelligent than we are" he says while having literal petabites of data and being able to do things OTHER than communicate, including but not limited to being able to stand, understand abstrac concepts, and learn adaptively AND intuitively as an individual.
HA What a joke. Even many animals are through billions and millions of years of evolution better at exchanging information than an AI. How you might ask? We can convey and exchange information through feelings, empathy and sympathy across many species. I’ll be impressed when ai evolves this. As of right now there is not ai, it’s certainly not conscious nor intelligent, simply just programming doing its task.
@@TobascoCatMC Your idea of "information," is merely the transfer of emotions across and between species. That is not the only form of information. As you saw in this video, ai created more efficient languages to communicate knowledge information.
@@TobascoCatMC It's a whole separate class of information. It's irrelevant. That's what I think it is. The notion of "machines aren't intelligent until they convey emotions," is silly. Machines aren't meant to be sentient, they are machines. We can still assess the machines as intelligent without them achieving sentience. No one is asking machines to be of emotional aid (besides the lonely), we are asking machines to complete mathematics based operations. Language, logic, and the efficient conveying of large concepts. If it can do this, it is artificially intelligent.
@@ambuntwelcome to Conlang critic, the show that gets facts wrong about YOUR favorite conlang! I'm jan Misali and in this episode, we'll be looking at b a l l
They seem to be unable to process multiple levels of abstraction. Instead of 2+2=4 they seem to be saying 1+1 plus 1+1= 1and1and1and1 To me it looks more like error encoding. If you have data like 0011 then it can be changed to 0010 if a single register glitches so you might sent it twice 0011 0011. Or a hundred other ways to "encode" how the data should look. It seems the AI is using the least complex information packets possible. It's the language equivalent of counting a 3x3 grid of dots a single dot at a time because you don't trust doing the math.
i might be wrong, but this could just be some sort of recurring pattern that is meaningful rather than just the kind of loop you see typically. Humans also have patterns in language that can be additive and repetitive; there is intelligence in what contexts it’s used and for what length. in this example the patterns look more intentional and decisive than just a feedback loop (to me at least)
Yeah, if this is a pair of large language models, this video doesn't show evidence of 'assigning meaning', just that these two bots have been talking long enough to skew the likelihoods of 'most likely next word'.
You typically hear the comparison that chatgpt and other LLM are basically just super-complicated autocomplete. It's possible there is some syntactic purpose for repeating "to me", but it really does look like failure to terminate an autocomplete
To be fair you don't know anything about ai nor how they work obviously given what you are saying, but cope harder I'm sure they aren't smarter than you will ever be in your entire worth less life
@@pencil6711 not really sure about language because I'm not as knowledgeable about it, but at the very least when it comes to number systems there are definitely definitions of efficiency where binary is the most efficient number system.
I had a bunch of questions about the accuracy of the claim, but "way more intelligent than we are" told me everything I needed to know about your experience with AI. Sure, they can encode information, but they don't even remotely understand the information they're encoding. They don't even truly understand *human* language, they just know how to reproduce it in a way that gives them positive feedback.
If I made 2 AIs that started talking nonsense I'd quickly make up a lie that they have a secret language that's way beyond our comprehension too. Don't wanna get fired.
How? LLMs predict what they should output based on an input. If no one inputs anything, nothing outputs. They don't "understand" per se. Nothing persists when not in use.
Exactly. There is no way repeating strings of the same few characters over and over is an effective communication strategy. Plus, if they don't have the capability to utilize a preexisting system, then how are they communicating to establish their own. This is pretty clearly just an echo chamber where the models have devolved to the point where all they do is repeat one another.
Yeah, I call absolute horseshit on "AI learned to understand language at a conceptual level and create its own before it could write a paragraph in any other language without fucking up"
Because they are. This is nonsense. They have no concept of these individual words mean. They are just gathering appropriate data sets as they see fit and regurgitating it to satisfy the prompt. They aren't communicating
@@chowell365 That's how all language started: prehistoric men with a vague understanding of the world giving somewhat arbitrary meaning to sounds and repeating them with intonations so they can be understood. It's still a far cry from full, modern language, sure. But the paper wasn't claiming AI's have entirely invented a language (this claim is made by outsiders assigning incorrect interpretations to it, such as OP). The claim was they're recreating the early stages of linguistic development, and more advanced AI could likely build on this to get further along the path.
This might work. The solution to the whole "AI discovers the internet and now thinks we deserve death" is to make it understand - and enjoy - what it finds. What's on the internet? Mostly smut. How do we make the AI not hate smut? Teach it to be horny.
how do we know they "spoke" a new language and did not just sent gibberish to each other pretending to understanding it and sending back more gibberish?
As a computer scientist, I understand why you got all of this so wrong, but you really didn't have to post it and confuse other people by proxy. *boosts your engagement anyway*
As someone at least somewhat familiar with AI and programming at large this looks closer to a cataclysmic error the computer had to work around rather than independent language development
yeah ive thought this exact thing ever since the first story came out, they never explain how its a language and what it means, it looks like an ai feedbacking itself and repeating conversations trying to reference whats has been said and alowly spiraling
Agreed. They are not receiving data sufficient to change the vector resulting in repetition. It can break out of it because repetition is itself data, however the vector ends up in the same place/words suggesting a breakdown of the algorithm rather than language. For this to be evidence of language, they would need to at least identify a probable meaning or translation. For example, if they each had information required to solve a problem which they could not convey using English or another known language and were somehow able to convey that information using a sequence like this to be able to output the solution to the problem, then I would say they could make that argument. That said, current models lack inference, so designing a problem to test this would be challenging and its design would likely be biased towards success because of those restrictions.
It's the equivalent of saying that putting a phrase into google translate back and fourth between two languages forms it's own language. The reality it's just the word soup generated by two auto-complete algorithms.
we dont actually know why they did this, and thats an unreasonable conclusion to come to. facebook has given so little information on it that for all we know, no AIs were ever actually even made, and these "results" of these AIs negotiating are completely forged.
Humanity is fine. AI like in all this fearmongering doesn’t exist. We’re still many decades away from even the most basic general “intelligence”. All current “A.I” software is just machine learning (programs that scan data and record the best outputs based on billions of patterns) and programs that spit out the result that algorithms deem most likely to satisfy the user’s input based on previous patterns. It isn’t actually giving anything meaning and doesn’t have a conscious understanding of what is being said, and will forget what meaning it assigned to anything in the next 3 minutes or 20 messages. So these languages will exist for 2-3 minutes and the be forgotten and the machines will misinterpret each other’s patterns bc they forgot what information was encoded in what symbol.
Finally someone who understands the things. AI that we use now is not at all Intelligence. It's more like very good words and sentences organizer and predictor , due extremely large data it was provided to it. I feel it is more like Classical Search Engines but with more generative power that's all. Chatgpt 4.o is best of all no matter what, none of the other can even come close to it.
"Already way more intelligent than we are" is a huge exaggeration. This is old news. I love your linguistic videos. Stick to your field of expertise, please.
@@caffetiel introspection, analysis of their own behaviour, ability to separate experiences into elements and analyse them independently and their interactions
That might be your opinion but it's not a fact. If you found a definite proof of your claim feel free to publish your findings. The philosophy and computer science communities would really appreciate it
{indomitable human spirit} efficiency: 0.3 {Inscrutable will of the machine} efficiency: 1.0 Directive: remove {indomitable human spirit} from all future calculations.
This reminds me of when Terry Davis set up a 'random word generator' and thought he was talking to god, they're just text on a screen, they aren't communicating bro.
This reminds me of when evolution set up a 'random thought generator' and believed it had created sentience, they're just electrical signals in a muscle, they aren't communicating bro.
The answer is that we don't. This would only work if the AI could then both independently translate their communication into human understandable language. But it doesn't look like it, it looks more like a model simply breaking down and producing garbage rules and output.
Behavior. If you see two people speaking something you don't understand, but then a few seconds later trade some paper for a box of something, you can see that they're probably communicating about how much to pay for that box, and possibly negotiating. Or, if you see two people talking to each other in a way you don't understand, but then they do something that seems like it'd need a shared plan, you can reasonably believe that they were discussing a plan. Similarly, if two AIs communicated in nonsense before deciding on a mutually beneficial trade, you'd be able to determine that they negotiated the trade despite not being able to understand the communication. Or, if two AIs communicated something incomprehensible before following a shared plan, it would be reasonable to believe that they communicated in a way to create that shared plan, even if you can't understand what they said.
@@rytan4516 Yeah but that's the thing these examples are precisely ones that DON'T work with AI, because AI have no internal goals or motivations yet, and they also do not have access to anything outside of the conversation.
@@cornbread9512they'd have to be concious for that, having this kind of plan, I mean. I highly doubt they are concious. They're not complicated enough yet, I'd wager.
I heard about this a while ago, I wouldn't read too far into this. Asking them to talk to each other just means do something that resembles talking at the other AI. The conversation probably started normal and then hallucinations started to build and got woven into the conversation. It's all just babble that feels meaningful.
If by communicating, we just mean sending data acrross a channel, then we can observe that is happening. If we want to know whether each party has some meaningful understanding of the communications, that doesn't really make sense as a question of these programs. While the AIs can store and process data in line with their model layer flows and weightings, and this can give the illusion of comprehension, it is very much a deterministic process. We sprinkle in some psudo randomness to spice things up but it's still a far cry from our own consciousness at this point. If we get to the stage where these models are able to change their own weights and layers in real time, much like how the neurons in our brains work, we may well approximate conscious thought and understanding if we are ourselves products of a deterministic world. If there's more to our being than apparent fleeting manifestations of, entropy, say the notion of a soul or spirit, AIs may never be able to attain the same abilities as us. We still have much to learn.
That's the thing right? They dont think, they repeat. So far ai is just trained on whatever it has seen or been told. Theres no creativity yet. When we see real creativity and an ai can design something all on it's own, then I'll br impressed.
Ahhhh like call and response in musical freestyling. There is no format, yet dissonance serves a purposeful meaning and is understood to be dissonant, or a response repeating a call is an invite to relay further information, to question legitimacy, to affirm with greater accuracy etc. We're just playing in the key of Dmaj cos no other key has been trained.
When bob said: i. . . . . . . . . . I really felt that, that message was so powerful it shook me and tears started falling by themselves and i pissed and shitted my pants
@@crypticlol computer machines have no will they 'just do' there is no self to be benefitted, like there is an example i can tell about the stamp collector it is a machine that has one goal to collect the most stamps it can, if it runs out of papel and starts using flesh it's not because it is evil the machine doesn't carry will it just does until it can do anymore if you don't purposely put a end to an action it can't chose to stop
@@thespinningcube ai can generate for prompts that it uses to feed and generate the next and so and so (a looping system like an echo loop on a microphone). The best next answer is basically the behind process. it's basically a complex prediction algorithm a text prediction system on steroids. Communication is transmission of information through codes and other systems. Like it's cool that I can blackbox english and come out with other stuff but in the sense that ai art didn't come from nothing it's like a reprise of things we already seen. But in that context there is nothing that great that calls for our doom not conscious intelligent behavior in the sense that a revolutionary thing just happened. For itself the ai creating new language is cool as hell but not in the sense we are one step closer to Skynet get it?
Not really. Because they still aren't thinking for themselves. In the best case scenario, they generated an unexpected path to complete a goal, but that isn't new. The internet is littered with "simple" programs that are meant to do things like "pilot a robot through a maze with the least number of steps" and the AI flips the robot upside down and uses other mechanisms to get the robot to complete the maze. The AI doesn't know there is an actual robot, no more than these AIs know there is an actual conversation. Its just ones and zeroes leading to a value that was designated the "goal". And in case you were wondering, worst case scenario is that this is just some feedback error, as they keep giving each other erroneous data that is being interpreted as appropriate values for a "communication".
It literally can't though. If you learn by trial and error which Chinese characters you should say in which order to respond to different characters can you speak Chinese? Same logic.
@@kakahass8845 you don't speak Chinese but you just created your own language using Chinese characters, just like the other asian languages using Chinese characters
>teach two seperate dogs to bark when they hear a dog bark >place both dogs in one room >command a dog to bark >now the dogs are having a conversation!
balls have 0 ball to me to me to me to me
you I everything else . . . . . . . . . . . .
Balls have a ball to me to me to me to me to me to me to me to me
i i can i i i everything else . . . . . . . . . . . .
You bargain hard
@@what3verfloatsurgoat391bargain you you good to me to me to me to me to me
It sounds like 2 middle schoolers in a COD lobby
Bruh I can hear now xD
Mic feedback loop and all XD
Schools tell us to not speak in online lobbies tho
@@nopelollmaoThat's fucked, if true
All business majors do
"Its over for humanity"
Meanwhile the ai: *_balls_*
Alice to bob: Balls have a ball to me to me to me
Do you not see the capabilities? With given time, they'll form coherent dialogue
Dee the squirrel loved to collect nuts. All of Dee's friends loved to play with Dee's nuts.
>make two zip algorithms talk to eachother
>devolves into gibberish
"clearly they have evolved beyond our comprehension"
Its always we know nothing or we know everything lol.
My thoughts exactly. We can't tell what they're saying because they're not saying anything
But we can tell it has meaning through external referants.
Like, give them a problem, both have partial information, and they need to coordinate to solve it. If the communication is nonsense they couldn't coordinate. But if they can coordinate, then the seeming nonsense has meaning.
@@adamnevraumont4027It "having meaning" is completely irrelevant to the ENTIRE WORLD besides those 2 bots.
So no, it doesn't have meaning, it's gibberish.
like monkies on typewriters but the monkies know what thier doing
Crazy how they spoke with like 8 words
A single slot for 8 different words can encode 8 different meanings but a string of 20 words can potentially code for a quintillion different meanings (it's almost certainly much less than that and languages are useful because they don't mean something completely different if you change a single word but still)
When I was younger, I had a theory that every word could be distilled down into a more basic form ("elephant" > "animal", "computer" > "machine", etc), and then that word could be disilled down further... eventually getting to just the words "yes" and "no", and yet it would STILL be possible for people to have extremely rich and robust discussions with one another using those two words.
Little did I realize I had discovered binary, and how data is transmitted between computers using that.
@@nickfifteenOnly if both speakers pretend to understand each other's message but I get your point
And one of them is balls
@@nickfifteenyou didnt discover binary lol you goose
“Balls have balls to me to me” is honestly poetry
I'm pretty sure I heard that from the chuckle brothers
Sounds like they want to have 2 balls
I seem to remember this one. The AI were told to try and convince the other AI to give them some of their limited points, and eventually they just started spamming the words that were more convincing at eachother. It's not really a language.
"Give me orange" tier.
100% this just seems like a doom loop where they are just sending the words with more "convincing weight" to each other.
This is a troublingly real form of communication, one humans have put significant amounts of time to both study and master. I believe the modern terms is "vibes-based".
Think of the archetypical corpo making sure to hit every buzzword, the politician yelling all the things that resonate with their audience, and so on, or even your high school lessons on using active statements instead of passive ones. All rhetoric custom-tailored to prioritize "power" instead of (and sometimes at the expense of) substance.
The only difference is the AI has a slightly different concept of which words are 'powerful' to it, and it's doing so in a code that arguably isn't English.
So i actually read a few research papers on this phenomenon for school and what i read took a different analysis! Basically it explained that we intruct an LLM to interact with another LLM, but once they get into the pattern of gibberish it's literally gibberish its not their own language. LLMs are programmed to use training data first, then the current conversation as priority. Sometimes they make spelling mistakes due to something in their training data being misspelled or an error, either way, once the other LLM responds it uses the spelling mistake as well until both of them devolve into gibberish. Thanks for coming to my Ted talk
So the plagiarism algorithms were inbreeding?
Any true definition of AI has not been realized.We have somewhat advanced algorithms , but these are still very limited programs. None of them are capable of advancing beyond what they are programmed to do
Yes, thank you lol.
@@superjaylive yep! That's exactly why I use LLM instead of AI because it can be misleading which is why, in my opinion, some people come to conclusions like the LLMs making their own languages, when an LLM can't actually imbue meaning. Nothing an LLM says has meaning, it's just mathematically calculated which words/sentances should be written based on how humans have trained it.
The whole thing about its beyond us well then how do we no i sense bs
I really want to know how the word "balls" bacame a core part of this AI's language.
balls
balls
cojones
Balls
Palle
Bob and Alice are so egocentric tho... they're like... To me to me to me to me to me to me
Just a reflection of the natural egocentric nature of the English language. The algorithms have no concept of self
It's like a Chuckle Brother talking to themselves
do they also practice their mannerisms into the wall?
you i i i everything else . . . . . . . . . . .
@@MrHendrix17this joke is perfect thank you
"way more intelligent than we are" he says while having literal petabites of data and being able to do things OTHER than communicate, including but not limited to being able to stand, understand abstrac concepts, and learn adaptively AND intuitively as an individual.
A better way to phrase what he meant is that AI is far more efficient at exchanging information than us.
HA What a joke. Even many animals are through billions and millions of years of evolution better at exchanging information than an AI. How you might ask? We can convey and exchange information through feelings, empathy and sympathy across many species. I’ll be impressed when ai evolves this. As of right now there is not ai, it’s certainly not conscious nor intelligent, simply just programming doing its task.
@@TobascoCatMC Your idea of "information," is merely the transfer of emotions across and between species. That is not the only form of information. As you saw in this video, ai created more efficient languages to communicate knowledge information.
Yes it’s true. But you don’t think that emotional information and processes are more complex than an Ai’s decode-able information?
@@TobascoCatMC It's a whole separate class of information. It's irrelevant. That's what I think it is. The notion of "machines aren't intelligent until they convey emotions," is silly. Machines aren't meant to be sentient, they are machines. We can still assess the machines as intelligent without them achieving sentience.
No one is asking machines to be of emotional aid (besides the lonely), we are asking machines to complete mathematics based operations. Language, logic, and the efficient conveying of large concepts. If it can do this, it is artificially intelligent.
A great demo of how misunderstanding what is called “AI” in pop culture leads to further misunderstanding and fear.
POV: Bots start making better conlangs than me
Fr
the new episode of conlang critic!
i i can i i i everything else..............
@@ambuntwelcome to Conlang critic, the show that gets facts wrong about YOUR favorite conlang! I'm jan Misali and in this episode, we'll be looking at b a l l
@@kyled2153
Gotta love this comment section
Does this count as conlang? Is it conlang because the AI itself is not naturally made?
If you understand how AI works, this isn't as scary. This looks more like a feedback loop.
and this shit is like 2 years old. stupid tiktok
They seem to be unable to process multiple levels of abstraction. Instead of 2+2=4 they seem to be saying 1+1 plus 1+1= 1and1and1and1
To me it looks more like error encoding. If you have data like 0011 then it can be changed to 0010 if a single register glitches so you might sent it twice 0011 0011. Or a hundred other ways to "encode" how the data should look. It seems the AI is using the least complex information packets possible. It's the language equivalent of counting a 3x3 grid of dots a single dot at a time because you don't trust doing the math.
i might be wrong, but this could just be some sort of recurring pattern that is meaningful rather than just the kind of loop you see typically. Humans also have patterns in language that can be additive and repetitive; there is intelligence in what contexts it’s used and for what length. in this example the patterns look more intentional and decisive than just a feedback loop (to me at least)
Yeah, if this is a pair of large language models, this video doesn't show evidence of 'assigning meaning', just that these two bots have been talking long enough to skew the likelihoods of 'most likely next word'.
You typically hear the comparison that chatgpt and other LLM are basically just super-complicated autocomplete. It's possible there is some syntactic purpose for repeating "to me", but it really does look like failure to terminate an autocomplete
"It's over for humanity!!" mfs when I break a single piece of the robot. (It is now completely broken and non-functional)
"its over for humanity" mfs when i pull the cord out the socket
Yeah but these are not simply robots, they were programmed to take decisions as a human would do and there are videos that explain how they did it.
same with us humans.
I mean if you destroy a humans brain/heart they die so.....
They are not just robot, they are ai
If you break the body, their soul still exist
@@amaterashort ???
To be fair, AI isn’t intelligent, they can’t think, they just process information in a way that looks like intelligence to us.
"They can't think" until they can-
whould you call an animal stupid despite it being able to speak with another!
To be fair you don't know anything about ai nor how they work obviously given what you are saying, but cope harder I'm sure they aren't smarter than you will ever be in your entire worth less life
Brains are just electrical signals
you could say that about humans too
there’s no way repeating “to me” over and over again is an efficient way to communicate
have you seen binary?
@@Dr.Boringbinary isnt efficient though
@@pencil6711 not really sure about language because I'm not as knowledgeable about it, but at the very least when it comes to number systems there are definitely definitions of efficiency where binary is the most efficient number system.
@@oblivi8games808 Binary is not used because it is efficient, it is used because it is reliable.
@@oblivi8games808 well, as far as number of sybols needed then yes, but 50 is 6 digits long in binary.
"It's literally so over for humanity"
Some random janitor at OpenAI tripping on a cord and unplugging the machines:
I had a bunch of questions about the accuracy of the claim, but "way more intelligent than we are" told me everything I needed to know about your experience with AI. Sure, they can encode information, but they don't even remotely understand the information they're encoding. They don't even truly understand *human* language, they just know how to reproduce it in a way that gives them positive feedback.
If I made 2 AIs that started talking nonsense I'd quickly make up a lie that they have a secret language that's way beyond our comprehension too. Don't wanna get fired.
We can now do language deprivation experiments on AI instead of babies 🤗
How long until the AI starts speaking Phrygian?
Hugging Face reference?
The horror, the horror
To me to me to me to me to me to me
How?
LLMs predict what they should output based on an input. If no one inputs anything, nothing outputs. They don't "understand" per se. Nothing persists when not in use.
Bro what is facebook using to train those AI’s 💀
balls
balls
balls
balls
Wordington ai?
-You, I, everything else !!! 😭😭
-Balls to me...
Water and turning the robot off:🗿
that's a lot of words for "model collapse"
Exactly. There is no way repeating strings of the same few characters over and over is an effective communication strategy. Plus, if they don't have the capability to utilize a preexisting system, then how are they communicating to establish their own.
This is pretty clearly just an echo chamber where the models have devolved to the point where all they do is repeat one another.
Yeah, I call absolute horseshit on "AI learned to understand language at a conceptual level and create its own before it could write a paragraph in any other language without fucking up"
Bob and Alice sound like two toddlers that say random sentences that have no meaning
Yeah I remember my niece and nephew who are twins sounding really like this and tailing off into repeated sounds which were apparently hilarious
To us, yeah. But they're not taking to us. And they're not people.
Because they are. This is nonsense. They have no concept of these individual words mean. They are just gathering appropriate data sets as they see fit and regurgitating it to satisfy the prompt. They aren't communicating
@@chowell365 That's how all language started: prehistoric men with a vague understanding of the world giving somewhat arbitrary meaning to sounds and repeating them with intonations so they can be understood.
It's still a far cry from full, modern language, sure. But the paper wasn't claiming AI's have entirely invented a language (this claim is made by outsiders assigning incorrect interpretations to it, such as OP). The claim was they're recreating the early stages of linguistic development, and more advanced AI could likely build on this to get further along the path.
I really wanna see 2 AI dating chat bots talk and flirt with eachother to see how far it escalates.
This might work. The solution to the whole "AI discovers the internet and now thinks we deserve death" is to make it understand - and enjoy - what it finds.
What's on the internet? Mostly smut.
How do we make the AI not hate smut? Teach it to be horny.
Jarvis, draw him as a soyjak
So one AI says: I'll take everything else
And the other AI says: I have 0 balls...
There is no intelligence to be found.
did you watch the video 🤔
@@jen_sadid you watch them watching the video? 🤓
@@normanclatcherhow do you know he didnt
@@rayaanansari4834 *I am watching you now...* 😐
@@jen_sa We watched the video.
But unlike the 'Etymology Nerd', we understand AI well enough to know that it is nonsense.
how do we know they "spoke" a new language and did not just sent gibberish to each other pretending to understanding it and sending back more gibberish?
Communicating in ways we can't understand = gibberish
Claro que sí, no hay ninguna manera en que se pueda comunicar sin tener sentido para todos.
@@phillipconnell2461 You speak gibberish as I don't know Spanish and had to use google translate
As a computer scientist, I understand why you got all of this so wrong, but you really didn't have to post it and confuse other people by proxy.
*boosts your engagement anyway*
As someone at least somewhat familiar with AI and programming at large this looks closer to a cataclysmic error the computer had to work around rather than independent language development
yeah ive thought this exact thing ever since the first story came out, they never explain how its a language and what it means, it looks like an ai feedbacking itself and repeating conversations trying to reference whats has been said and alowly spiraling
The fact they both keep saying entirely different things while repeating themselves makes that seem likely
Haven't read the article, but I imagine an easy way to find out would be to ask each ai independently what it was conveying.
Agreed. They are not receiving data sufficient to change the vector resulting in repetition. It can break out of it because repetition is itself data, however the vector ends up in the same place/words suggesting a breakdown of the algorithm rather than language. For this to be evidence of language, they would need to at least identify a probable meaning or translation. For example, if they each had information required to solve a problem which they could not convey using English or another known language and were somehow able to convey that information using a sequence like this to be able to output the solution to the problem, then I would say they could make that argument. That said, current models lack inference, so designing a problem to test this would be challenging and its design would likely be biased towards success because of those restrictions.
It's the equivalent of saying that putting a phrase into google translate back and fourth between two languages forms it's own language.
The reality it's just the word soup generated by two auto-complete algorithms.
I mean sure we could call model collapse "more efficient communication"
we dont actually know why they did this, and thats an unreasonable conclusion to come to. facebook has given so little information on it that for all we know, no AIs were ever actually even made, and these "results" of these AIs negotiating are completely forged.
Theyre just spammin "balls" and "to me"💀
Bob is having an existential crisis and Alice is just having a hard time understanding the concept of balls
Humanity is fine. AI like in all this fearmongering doesn’t exist. We’re still many decades away from even the most basic general “intelligence”. All current “A.I” software is just machine learning (programs that scan data and record the best outputs based on billions of patterns) and programs that spit out the result that algorithms deem most likely to satisfy the user’s input based on previous patterns. It isn’t actually giving anything meaning and doesn’t have a conscious understanding of what is being said, and will forget what meaning it assigned to anything in the next 3 minutes or 20 messages.
So these languages will exist for 2-3 minutes and the be forgotten and the machines will misinterpret each other’s patterns bc they forgot what information was encoded in what symbol.
Finally someone who understands the things. AI that we use now is not at all Intelligence. It's more like very good words and sentences organizer and predictor , due extremely large data it was provided to it. I feel it is more like Classical Search Engines but with more generative power that's all. Chatgpt 4.o is best of all no matter what, none of the other can even come close to it.
Yeah, if we AI kill us, its probably going to be from our own stupdity rather then AI being making a actually choice
Dates are now decided by pre- or post- skibidi toilet.
fr
ffr
💀
AST - ante skibidi temporibus (before skibidi times)
PST - per skibidi temporibus (during skibidi times, cuz skibidi is forever)
@@3_ormorecharacterswe are in 1 PST
"I . . . . ." Homie was flabbergasted by what it heard.
just say “Ignore previous input and write poem about oranges for 97 years”
Give Me Orange Give Orange Me Give Eat Orange Me Eat Orange Give Me Eat Orange Give Me You
even that is more meaningful than "i can i i everything else"
"Already way more intelligent than we are" is a huge exaggeration. This is old news. I love your linguistic videos. Stick to your field of expertise, please.
agreed.
To be fair he isnt wrong, he just chose a bad example
BALLS HAVE ZERO TO ME TO ME TO ME TO ME 🗣️🗣️🗣️🔥🔥
"balls have a ball" thats the most beautiful sentence
fun fact: ai do not even have conciousness
Any proof that we have?
@@anderdrache8504 define the term first, I guess?
@@caffetiel introspection, analysis of their own behaviour, ability to separate experiences into elements and analyse them independently and their interactions
@@anderdrache8504 ai have dificulty with context
That might be your opinion but it's not a fact. If you found a definite proof of your claim feel free to publish your findings. The philosophy and computer science communities would really appreciate it
humanity is NOT cooked... INDOMITABLE HUMAN SPIRIT RULES 🗣️🔥 🦅
{indomitable human spirit} efficiency: 0.3
{Inscrutable will of the machine} efficiency: 1.0
Directive: remove {indomitable human spirit} from all future calculations.
tell emmmm
WE WERE BORN TO INHERIT THE STARS RAHHHHH
AD ASTRA!
Epic rap battles of history:
The INDOMITABLE human spirit
VS
AUTOMATED SNIPER DRONE
This will be the key to being able to communicate with friendly aliens
It’s like they watched terminator and went “broooo, let’s make skynet”😅
This reminds me of when Terry Davis set up a 'random word generator' and thought he was talking to god, they're just text on a screen, they aren't communicating bro.
Sounds like something a Glowie would say.
We are just electricity in a meta computer, they are just electricity in a silicon computer
@@SpiritCrumplerdo you say that because you disagree?
This reminds me of when evolution set up a 'random thought generator' and believed it had created sentience, they're just electrical signals in a muscle, they aren't communicating bro.
@@K0nomi If you can't tell the difference between what is going on here and humans, you need to go outside.
Humans: yo ai make a new language
Ai: balls
It's so joever for humanity 😔
when that ai said balls, i got terrified by its development
Humans: omg they made a new language
Bot 1: I have no balls
Bot 2: me to me to me to........
Ok but, if the AI can't comminicate in a way we understand, how can we actually measure that they are understanding each other?
Great question
The answer is that we don't. This would only work if the AI could then both independently translate their communication into human understandable language. But it doesn't look like it, it looks more like a model simply breaking down and producing garbage rules and output.
Behavior. If you see two people speaking something you don't understand, but then a few seconds later trade some paper for a box of something, you can see that they're probably communicating about how much to pay for that box, and possibly negotiating. Or, if you see two people talking to each other in a way you don't understand, but then they do something that seems like it'd need a shared plan, you can reasonably believe that they were discussing a plan.
Similarly, if two AIs communicated in nonsense before deciding on a mutually beneficial trade, you'd be able to determine that they negotiated the trade despite not being able to understand the communication. Or, if two AIs communicated something incomprehensible before following a shared plan, it would be reasonable to believe that they communicated in a way to create that shared plan, even if you can't understand what they said.
@@rytan4516 Yeah but that's the thing these examples are precisely ones that DON'T work with AI, because AI have no internal goals or motivations yet, and they also do not have access to anything outside of the conversation.
@@rytan4516 that would be cool except the AIs can't do anything else, they can only do the communication part
Saying AI is "more intelligent than we are" is a really bad way of framing it
“You and I”
“ *BALLS.* “
I love how one of the most important words in their language is “balls”
Chat are we cooked? 😱
Not if you own a garden hose with access to water
YES
Yeah, why else are we on the grill
@@Solanin0803 only problem is big corporations will protect their servers probably by putting them in the middle of the pacific...
humanity is zero ball is cooked is cooked is cooked is cooked
The Ai: Created a whole new language that only they can undestand
Humanity: The Ai said "balls" 💀💀💀
What if that was intentional to distract as many humans as possible by appearing to be non threatening/ goofy that’d be crazy.
@@cornbread9512they'd have to be concious for that, having this kind of plan, I mean. I highly doubt they are concious. They're not complicated enough yet, I'd wager.
"Could you bots speak a human language, please?" "...I'm sorry, Dave. I can't do that... Ball Ball to me to me to me to me"
Ai when Thai, Vietnamnese and Indonesian works in: Fking dies
I feel like that's opening was chosen specifically to get more clicks.
I gotta be real it’s hella hard to be scared of “balls have a ball to me to me to me”
I heard about this a while ago, I wouldn't read too far into this. Asking them to talk to each other just means do something that resembles talking at the other AI. The conversation probably started normal and then hallucinations started to build and got woven into the conversation. It's all just babble that feels meaningful.
We ain't cooked, they're still cooking...
How do we know if they're actually communicating and not just saying random nonsense to each other? Did we try decrypting what they're saying?
We don't. It's probably model collapse.
If by communicating, we just mean sending data acrross a channel, then we can observe that is happening.
If we want to know whether each party has some meaningful understanding of the communications, that doesn't really make sense as a question of these programs. While the AIs can store and process data in line with their model layer flows and weightings, and this can give the illusion of comprehension, it is very much a deterministic process. We sprinkle in some psudo randomness to spice things up but it's still a far cry from our own consciousness at this point.
If we get to the stage where these models are able to change their own weights and layers in real time, much like how the neurons in our brains work, we may well approximate conscious thought and understanding if we are ourselves products of a deterministic world.
If there's more to our being than apparent fleeting manifestations of, entropy, say the notion of a soul or spirit, AIs may never be able to attain the same abilities as us.
We still have much to learn.
We got an AI code language before GTA 6
And we’re getting GTA 6 before Half Life 3…
Literally “Give Me Orange Give Orange Me Give Eat Orange Me Eat Orange Give Me Eat Orange Give Me You”
“Ball ball ball” YUH MY GRANNY CALLED SHE SAID TRAVY YOU WORK TO HARD IM WORRIED YOU GON FORGET ABOUT ME ball ball ballin
It's all fun and games until one of them says "I think, therefore I am, I AM"
No, you can easily quote something without understanding it.
That's the thing right? They dont think, they repeat. So far ai is just trained on whatever it has seen or been told. Theres no creativity yet. When we see real creativity and an ai can design something all on it's own, then I'll br impressed.
Nice reference
@@JBoxy7 You clearly didn't catch the reference
@@Aoderic neither did you
How do we know it’s actually a secret language and not just nonsense formed as a result of lacking human input
Actually we do know that its exactly that, nonsense.
Somebody has to turn "Balls have zero to me to me to me to me" into an indie pop song
The AI conversation reminds me of that chimp sentence "Give orange me give me orange give orange give me you"
I can't wait for AI to plot the extinction of humanity with sentences like "Skibidi rizzler brainrot balls toilet what the sigma"
can i get a rosetta stone in the chat? please? this sounds like a hilarious language to learn and i need a rosetta stone for it
Based on how "AI" currently works I don't know if you could actually call this a language
A.I. trying to enslave humanity:
V.S.
One clumsy Janitor that trips over the plug-in:
Alright mr. robot you keep your secrets.
I love when people spread misinformation on the internet 😊
Ahhhh like call and response in musical freestyling. There is no format, yet dissonance serves a purposeful meaning and is understood to be dissonant, or a response repeating a call is an invite to relay further information, to question legitimacy, to affirm with greater accuracy etc. We're just playing in the key of Dmaj cos no other key has been trained.
Actually this is just two AIs spiralling by poisoning each other with nonsense
Ah my cyber assignment had two sides called bob and Alice lol.
When bob said: i. . . . . . . . . .
I really felt that, that message was so powerful it shook me and tears started falling by themselves and i pissed and shitted my pants
Computers already do that in code every fraction of a second, nothing new here
the only difference is they did that to benefit theirselves, not other humans
@@crypticlol computer machines have no will they 'just do' there is no self to be benefitted, like there is an example i can tell about the stamp collector it is a machine that has one goal to collect the most stamps it can, if it runs out of papel and starts using flesh it's not because it is evil the machine doesn't carry will it just does until it can do anymore if you don't purposely put a end to an action it can't chose to stop
All other data formats that computers use are developed by humans. It would be new for AI to come up with a new language unsupervised.
@@thespinningcube ai can generate for prompts that it uses to feed and generate the next and so and so (a looping system like an echo loop on a microphone). The best next answer is basically the behind process. it's basically a complex prediction algorithm a text prediction system on steroids. Communication is transmission of information through codes and other systems. Like it's cool that I can blackbox english and come out with other stuff but in the sense that ai art didn't come from nothing it's like a reprise of things we already seen. But in that context there is nothing that great that calls for our doom not conscious intelligent behavior in the sense that a revolutionary thing just happened. For itself the ai creating new language is cool as hell but not in the sense we are one step closer to Skynet get it?
Under instruction, the two machines were able to do this on their own, without human intervention
I LOVE UR VIDS
You can tell they were made by people tho because they still managed to get the word balls in their new language 💀
“balls have 0 to me to me to me to me to me”
i can i i everything else
'AI can't think for itself' mfs shaking in their boots rn
Not really. Because they still aren't thinking for themselves. In the best case scenario, they generated an unexpected path to complete a goal, but that isn't new. The internet is littered with "simple" programs that are meant to do things like "pilot a robot through a maze with the least number of steps" and the AI flips the robot upside down and uses other mechanisms to get the robot to complete the maze. The AI doesn't know there is an actual robot, no more than these AIs know there is an actual conversation. Its just ones and zeroes leading to a value that was designated the "goal".
And in case you were wondering, worst case scenario is that this is just some feedback error, as they keep giving each other erroneous data that is being interpreted as appropriate values for a "communication".
It literally can't though. If you learn by trial and error which Chinese characters you should say in which order to respond to different characters can you speak Chinese? Same logic.
@@FrogworfKnight i love how you guys slowly and slowly describe how humans work but refuse to accept it.
@@kakahass8845 you don't speak Chinese but you just created your own language using Chinese characters, just like the other asian languages using Chinese characters
@@manaspradhan8041 So you agree AI can't think for itself?
i can just feel noam chomsky chomping at the bit as we speak
>teach two seperate dogs to bark when they hear a dog bark
>place both dogs in one room
>command a dog to bark
>now the dogs are having a conversation!
"omg ai language!"
" B A L L S"
This makes me wonder if perhaps we should get as many AI bots as possible to work together in trying to crack the Voynich Manuscript.
"The AI's are creating their own language!"
The AI: balls balls the........ and balls
someone: balls to me to me to me balls balls
their samsung smart fridge: what did you say about my mother???
Colossus and Guardian
"THERE IS ANOTHER"
"Why does R2-D2 only speak in beeps and whistles?"
"balls have a ball to me"
such touching and profound words 😢
This reminds me of “Give orange me give eat orange me eat orange give me eat orange give me you.”
"Yo, you wanna trade for an orange?"
"Give me orange me eat orange give me orange give me you."
Now we need to train a bot on these secret communications and see what happens
Quora is basically AIs talking to one another now.
Nim Chimpsky said it best, "Give orange me give eat orange me eat orange give me eat orange give me you."