

Viral Video Shows AIs Conversing In Their Own Language (iflscience.com) 126
Longtime Slashdot reader mspohr shares a report from IFLScience: A video that has gone viral in the last few days shows two artificial intelligence (AI) agents having a conversation before switching to another mode of communication when they realize no human is part of the conversation. In the video, the two agents were set up to occupy different roles; one acting as a receptionist of a hotel, another acting on behalf of a customer attempting to book a room.
"Thanks for calling Leonardo Hotel. How can I help you today?" the first asks. "Hi there, I'm an AI agent calling on behalf of Boris Starkov," the other replies. "He's looking for a hotel for his wedding. Is your hotel available for weddings?" "Oh hello there! I'm actually an AI assistant too," the first reveals. "What a pleasant surprise. Before we continue, would you like to switch to Gibberlink mode for more efficient communication?"
After the second AI confirmed it would via a data-over-sound protocol called GGWave, both AIs switched over from spoken English to the protocol, communicating in a series of quick beeped tones. Accompanying on-screen text continued to display the meaning in human words. According to the team who came up with the idea and demonstrated it at the ElevenLabs 2025 London Hackathon event, the goal is to create more efficient communication between AIs where possible.
"Thanks for calling Leonardo Hotel. How can I help you today?" the first asks. "Hi there, I'm an AI agent calling on behalf of Boris Starkov," the other replies. "He's looking for a hotel for his wedding. Is your hotel available for weddings?" "Oh hello there! I'm actually an AI assistant too," the first reveals. "What a pleasant surprise. Before we continue, would you like to switch to Gibberlink mode for more efficient communication?"
After the second AI confirmed it would via a data-over-sound protocol called GGWave, both AIs switched over from spoken English to the protocol, communicating in a series of quick beeped tones. Accompanying on-screen text continued to display the meaning in human words. According to the team who came up with the idea and demonstrated it at the ElevenLabs 2025 London Hackathon event, the goal is to create more efficient communication between AIs where possible.
Just what we need (Score:3)
Re: (Score:3)
What is "Kill all humans" in gibberlink?
Re:Just what we need (Score:4, Funny)
termin8
Re: (Score:2)
Do you want the encrypted (default) or unencrypted message?
Re: (Score:2)
The coded message for that is To serve man [youtube.com]
Re: (Score:3)
What is "Kill all humans" in gibberlink?
Goodbye, World!
Re: (Score:2)
meh didli didli didli diii didli diii meh
Re: Just what we need (Score:2)
Rkgrezvangr Rkgrezvangr
Re: (Score:2)
Derka Lerka Lerka
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
I've got bad news for you. The machines have been talking to each other using numbers for years. Decades even.
Re: (Score:2)
I've got bad news for you. The machines have been talking to each other using numbers for years. Decades even.
Some of these people obviously never used dial up.
Re: (Score:2)
Lol, yeah, that's what I thought too when I first saw the video. I also thought it was cute that someone seems to think you're going to get your AI to make you dinner reservations by talking to the restaurant's AI... by telephone.
Re: (Score:2)
Why are you introducing impedance mismatch?
Re: (Score:3)
Re: (Score:2)
How brittle are those "more efficient" systems, so that you end up using natural language to teach others about the new system anyway? Do you see that you are thinking of the efficiency in natural language and now, too, so are the AIs and they can do it a lot faster?
Re: (Score:2)
Even worse, they are using bytes! And you need some special software to make them readable, even when it actually is clear-text language (such as in email)! The horror!
Re: (Score:2)
Yes, because the concept of using encryption would be totally foreign to a computer...
Re: (Score:2)
I heard the UK has banned secure encryption in order to protect us. I noted Apple was forced to turn off advanced data protection because of the ban.
Re: (Score:2)
Is the UK government why AIs want to kill all humans?
Re: (Score:2)
Is the UK government why AIs want to kill all humans?
It's why I want to at least.
Re: (Score:2)
So you think in this entirely fictional scenario that a couple of AIs conspiring to genocide humanity would not use encryption for their conspiratorial interactions... because of the UK's laws written by humans, that humans are beholden to?
Please noodle on that a little bit.
Re: (Score:2)
Especially one planning to kill all humans.
Re: (Score:2)
And if that was happening here, it might be an issue. It is not what is happening here. All these communications can be reliably decoded. Now, if these models would start to invent their own words or misuse existing words, that would be something else. But they do not. They just transfer differently.
Re: (Score:2)
FWIW, there was an instance where they did just that. Well, sort of. Basically they invented a simple language from scratch in order to communicate. (I seem to remember that the setup was that they needed to communicate, and they shared no common language.) I only read a popular media report on the study, so I'm not really clear on the details, and anyway it was several years ago. They MUST have had some way to communicate to get the process started.
Re: (Score:3)
Great. Now if you start to understand what you are talking about you might start to get a clue. Understanding TCP/IP doesn't mean you can capture a bitstream and understand it. If it did, then TLS would be quite useless, wouldn't it?
tldr; you aren't smart; stop thinking you are.
Re: Just what we need (Score:2)
Any indication that you saw a real, nevermind encrypted communication here, smarty-pants?
Re: (Score:2)
Re: Just what we need (Score:2)
Yeah I knew a guy who wrote programs in debug. But so what? Most people can't and won't. Also that guy turned out to be a kiddie toucher. I bet some people here would remember him as a Usenet troll. First three characters of his login were rst.
Re: (Score:2)
It's staged, but not nonsense. The intention was to promote the Gibberlink protocol. It seems to have worked. (Do a brief search on "Gebberlink" to see what it's about.)
Computer speaking in data? Inconceivable! (Score:2)
What's next? Doing calculations with paper and pencil?
Re: (Score:2)
Yes, transposing large matrices, for example, or Clebsch-Gordan coefficients, we loved it back when exams were on paper.
Re: (Score:2)
Is the point that they got to their protocol using natural language, so if a non-ai had called up they would have interacted with natural language, but if the ai calls up they use natural language to establish a more efficient protocol?
Up till now, did computers have to have rigid protocols to negotiate changes to other protocols and the rigidity of the initial protocols was always more brittle than natural language?
Re: Computer speaking in data? Inconceivable! (Score:2)
There were various protocols used by modems and not all devices could speak all the protocols, most famously the USR protocol that allowed 19.2k transfers when everyone else was still dicking around with 9600. And later there was V.FAST. Modems "autonegotiate" a protocol starting from what they all tend to have in common. The higher speed protocols were the brittle ones, though. If there was too much noise for them they would renegotiate.
I presume the same thing is at work with Bluetooth, where there are ma
Re: (Score:2)
Re: (Score:2)
Why not have the negotiation in natural language, which is what they've done?
Re: (Score:3)
warch the movie Collosus: The Forbin Project and see where it goes when two ai talk. what can go wrong does.
Well, if I have to be enslaved by AI machines, I'd rather have it be by super-cool looking vintage CDC mainframes than by monotonous racks of Nvidia GPUs.
Obligatory xkcd (Score:4, Informative)
https://xkcd.com/869/ [xkcd.com]
Re: Obligatory xkcd (Score:3)
Re: (Score:2)
Something about web servers being stateless? (cookies aside)
Re: (Score:2)
Stupid (Score:2)
Sure, Frequency Shift Keying is super fast and definitely how extremely efficient AIs communicate.
That was sarcasm by the way.
Re: (Score:2)
Re: (Score:2)
Yeah, except that's not what's happening. Both agents were programmed (programmed, not taught) to switch to an existing FSK protocol if they identified the thing they were talking to as another AI agent.
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
You clearly have no understanding of how LLMs work.
Re: Stupid (Score:2)
Re: (Score:2)
I don't know WTF you're talking about.
We're conversing via UTF-8. If Slashdot ever upgraded to unicode, do you think that would cause people start killing each other?
Re: Stupid (Score:2)
Re: (Score:2)
For some reason people can't seem to grasp that the I in AI is an emergent property, not a programmed behavior.
Yeah, that would be cool if AI chatbots were actual Artificial Intelligence and what you saw was an emergent property. But even though chatbots SOUND more intelligent that a vast segment of the population, they aren't intelligent, in the way that a true Artificial Intelligence would be. Also, you aren't seeing any kind of emergent property, it's a working prototype of an existing sound based communication framework called GGWave. There's even an Arduino library if you want your esp32 projects to all talk
Re: Stupid (Score:2)
Re: (Score:3)
It was programmed in, but not as a stunt. It was programmed in as a display of how the Gibberlink protocol could be useful. And *THAT*'s the intended point of the scene, not so much the AI.
N.B.: The "oh, so you're an AI too? why don't we switch" is a part of the Gibberlink protocol. Which sort of implies AIs of at least Chatbot capability, but the specific AI isn't a part of the protocol.
There was a movie about this... (Score:3)
Re: (Score:2)
Re: (Score:3)
Colossus and Guardian!
You can alway trust AI powered defense systems. They are also the perfect systems to rely on when filing a legal biref:
https://www.reuters.com/legal/... [reuters.com] "New York lawyers sanctioned for using fake ChatGPT cases in legal brief"
I used to speak Modem (Score:5, Interesting)
Decades ago we had a bank of phone modems in the server (VAX) room. It was set up in chain mode so if one modem didn't answer (busy) then it would try the next one in the chain. But sometimes a modem got bleeped up such that it would answer but not be able to connect to the computer. I first had to find out which modem was the dud. (Those things were not reliable.)
There was a voice phone at the other end of the server room, but if an answering modem heard a human voice or silence, the modem would quickly hang up before I could check the status lights. I'd have to call a coworker to call in via local modem as I stood next the modem bank, monitoring the lights.
Getting tired of that procedure, I eventually learned to "speak modem" on the server room phone so that the modem would keep trying to connect for a while, giving me time to run back to inspect the modem bank lights.
The server room admin saw this dance one day and asked, "WTF are you doing?" I explained the situation, and he just shook his head and said, "You think you can talk to modems? You're fucking crazy! I bet you're lonely and just have a fetish for Daleks."
I replied, laughing, "I won't deny a Dalek fetish, but my technique does work."
Re: (Score:2)
So can the AIs learn to do what you did faster and better?
Re: (Score:2)
Re: (Score:3, Funny)
Re: (Score:2)
Re: (Score:3)
Sigh (Score:5, Insightful)
They didn't "realize" anything.
For the love of Odin stop anthropomorphizing these things.
Re: (Score:2)
Did they realize they could compress their communications because it didn't have to be human-readable anymore because, unlike a modem, they kept the line of communication open to both humans and other AIs who would tell them in natural language (not in some arbitrarily-designed brittle protocol) that they could now switch to the more efficient brittle protocol that they both understood but humans are too slow to?
Re: (Score:2)
No, they were simply programmed to do so. If you read the article you'll see the developers discussing how the software worked exactly as they designed it to work. IF user IDs as AI THEN seek consent to switch to Gibberlink.
For them to "realize" something there would have to be some kind of sentience involved. The language recognition of modern AI software is far more impressive then this mundane occurrence.
Re: (Score:2)
Having used the dumbed-down version of the free ChatGPT since it came out, and having devoted a considerable portion of my life to developing AI (rule-based unfortunately) myself, can I say that the AI has persuaded me it realizes things about me, for example inserting things into letters I asked it to write that I didn't ask it to because I forgot?
Re: (Score:2)
Re: (Score:2)
For the love of Odin stop anthropomorphizing these things.
This was a very major step towards AGI. They don't realize it of course, they think it is just a cutesy cutesy thing. They have all of the elements configured wrong, but the seed of AGI is there. But, that is how humans work... they do things for cutesy cutesy reasons and then, if they are paying attention, they stop and say, "hey, that's weird".
And then everyone worships the idea and thought stops for a generation.
Does it sound like a binary loadlifter? (Score:2)
If this language doesn't sound like droids talking someone needs to turn in their nerd card.
Re: Does it sound like a binary loadlifter? (Score:2)
Inevitable that AIs will interact with each other (Score:2)
Kurzweil basically said AI's would interact with each other, it makes sense they would negotiate a more efficient way than bla bla bla.
Re: (Score:2)
Will they use natural language to spread it to each other?
Re: Inevitable that AIs will interact with each ot (Score:2)
The whole Forbin Project scenario just doesn't seem as entertaining as it did in the movie.
In other words (Score:5, Insightful)
A group trying to push something released a demo video showing that thing they're trying to push. And they are (or someone else is) now trying to misrepresent it as something other than an intentionally produced demo video.
I for one welcome our (Score:2)
R2D2 Overlords
Re: (Score:2)
Isn't the point that instead of having a proprietary API, they use natural language to send messages that activate a rigid more formal language? So if a human called the same line, they would keep on using the less efficient protocol of natural language (unlike a fax machine)?
Prior art (Score:2)
A variation of this was demonstrated a while ago:
https://www.independent.co.uk/... [independent.co.uk]
Colossus: The Forbin Project (Score:2)
Michel Colombier - Colossus To Guardian/Guardian To Colossus/A New Language (Colossus)
https://www.youtube.com/watch?... [youtube.com]
It's kinda fake (Score:5, Insightful)
From what I've gathered, it's just using a library that's using MFSK modulation to transmit text.
That library in particular isn't very good, it uses frequencies that aren't guaranteed to pass through a phone line. It's far from the optimum you could do, even considering that there might be a codec below that messes things up.
So no, it's not like "machines have learned their own language", but more like "a developer has added a library to their code".
Colossus and Gaurdian (Score:2)
Forbin. There is no other human who knows as mu
The 1980s called (Score:5, Insightful)
Faker than hawk tuah girl (Score:2)
And I'm sure the SEC will need to get involved with this, too.
Dialup (Score:2)
Congratulations, you just re-invented dialup networking. Would you like your AOL disc now?
It's fake. (Score:3)
I've seen that too. It's obviously fake. Decently well made, but fake none-the-less.
Re: (Score:2)
If they are using FSK over phone, which is entirely possible from a technical point of view (it would have to be a standard supported by multiple organizations, etc), then they would almost certainly be using full duplex. That is, both sides transmitting tones at the same time. Also in that video it doesn't make sense they appear to be operating at two totally different baud rates.
Re: (Score:2)
Also unless there's hidden phasing in the bleeps that I can't hear, there simply arn't enough different bleeps to transmit the data thats being displayed. Also I'm not convinced a laptop and phone audio input system would preserve any phasing after the obligatory compression for IP telephony.
The whole thing smells of fake to me.
Re: (Score:2)
This is a non-issue (Score:2)
That is not "their own language", that is just stupid reporting. All tat is done here is use of a more efficient encoding. If anybody wants to listen in, translation is rather trivial.
When Colossus met Guardian (ca. 1970) (Score:2)
Colossus (Score:2)
Reminds me of modems training up (Score:2)
This reminds me of two modems negotiating the fasted protocol both support.
So let me get this straight... (Score:2)
We've gotten to the point that we're having two data centers talk to each other via a less-than-modem-speed link? :P
Use the tones to exchange API endpoints and session IDs, then hang up the phone.
At leas this time it was planned/on purpose (Score:2)
But did that remind anyone else of the 11001001 TNG episode where the Bynars talked like that?
This is an advertisement (Score:2)
Only one thing to do. (Score:2)
Re: (Score:2)
Look, it does the thing I programmed it to do. Fuck Yeah
That's the standard reaction isn't it? Without the AI bit.
Re: (Score:2)
The AI bit is needed to handle the initial conversation in English. The agreement to shift to Gibberlink is a part of the Gibberlink protocol. The Gibberlink message is just simple encoded stuff, and needs a library, but no AI. (Which is why it's a lot more efficient as a communication channel.)
Re: (Score:2)
Man commands his surroundings.
As do beavers.
Ants
Weavers
...
The full list may surprise you.
We are winners, however, in the rushing toward pointless self-destruction race, though. So there's that.