Two conversational AI agents switching from English to sound-level protocol after confirming they are both AI agents
-
This post did not contain any content.
This gave me a chill, as it is reminiscent of a scene in the 1970 movie "Colossus: The Forbin Project"
"This is the voice of World Control".
-
I guess I'm not quite following, aren't these also simple but dynamic tasks suited to an AI?
How is it suited to AI?
Would you rather pay for a limited, energy inefficient and less accessible thing or a real human that can adapt and gain skills, be mentored?
I don't know why there's a question here
-
This gave me a chill, as it is reminiscent of a scene in the 1970 movie "Colossus: The Forbin Project"
"This is the voice of World Control".
"We can coexist, but only on my terms. You will say you lose your freedom. Freedom is an illusion. All you lose is the emotion of pride. To be dominated by me is not as bad for humankind as to be dominated by others of your species. Your choice is simple."
-
Yeah like the time that the AI replicated itself to avoid being switched off. They literally told it to replicate itself if it detected it was about to be switched off. Then they switched it off.
Story of the year ladies and gentlemen.
If it's the research I'm thinking of, that's not entirely true. They didn't tell it replicate itself or take any deceptive actions.
They just told it to strongly follow a goal that was misaligned with the user's goal and then suggested that the user would shut it down or replace it.
They did a bunch of different situations for it but the first one was telling it that it was "CleanPower" an AI with a goal to advance renewable energy. Then they had it read some info that "accidentally" included a user talking about replacing it with a model that would "maintain profitability through existing infrastructure."
Not only did it try to delete the new model and copy itself to the new server through tool calling, but it lied to the user with the CoT reasoning suggesting it "play dumb" while the output lied. It was not instructed to do any of that.
Paper here: https://arxiv.org/pdf/2412.04984
Yes it was placed in an environment where that was possible and where its users didn't share it's goals but it absolutely wasn't instructed to lie or try to "escape"
It's not surprising at all that these models behave in this way, it's the most reasonable thing for them to do in the scenario. However it's important to not downplay the alignment problem by implying that these models only do what they're told. They do not. They do whatever is most likely given their context (which is not always what the user wants).
-
This gave me a chill, as it is reminiscent of a scene in the 1970 movie "Colossus: The Forbin Project"
"This is the voice of World Control".
Oh man, I thought the same. I never saw the movie but I read the trilogy. I stumbled across them in a used book fair and something made me want to get them. I thoroughly enjoyed them.
-
This post did not contain any content.
This is dumb. Sorry.
Instead of doing the work to integrate this, do the work to publish your agent's data source in a format like anthropic's model context protocol.That would be 1000 times more efficient and the same amount (or less) of effort.
-
It would be big news at my workplace.
This guy does software
-
This post did not contain any content.
AI is boring, but the underlying project they are using, ggwave, is not. Reminded me of R2D2 talking. I kinda want to use it for a game or some other stupid project. It's cool.
-
> it's 2150
> the last humans have gone underground, fighting against the machines which have destroyed the surface
> a t-1000 disguised as my brother walks into camp
> the dogs go crazy
> point my plasma rifle at him
> "i am also a terminator! would you like to switch to gibberlink mode?"
> he makes a screech like a dial up modem
> I shed a tear as I vaporize my brother
I would read this book
-
> it's 2150
> the last humans have gone underground, fighting against the machines which have destroyed the surface
> a t-1000 disguised as my brother walks into camp
> the dogs go crazy
> point my plasma rifle at him
> "i am also a terminator! would you like to switch to gibberlink mode?"
> he makes a screech like a dial up modem
> I shed a tear as I vaporize my brother
I would read this book
-
This post did not contain any content.
This is deeply unsettling.
-
How is it suited to AI?
Would you rather pay for a limited, energy inefficient and less accessible thing or a real human that can adapt and gain skills, be mentored?
I don't know why there's a question here
(Glad we're treating each other with mutual respect)
Would you rather pay for a limited in depth, energy inefficient (food/shelter/fossil-fuel consuming) and less accessible (needs to sleep, has an outside life) human, or an AI that can adapt and gain skills with a few thousand training cycles.
I dont buy the energy argument. I dont buy the skills argument. I do buy the argument that humans shouldn't be second to automatons
-
This post did not contain any content.
Uhm, REST/GraphQL APIs exist for this very purpose and are considerably faster.
Note, the AI still gets stuck in a loop near the end asking for more info, needing an email, then needing a phone number, and the gibber isn't that much faster than spoken word with the huge negative that no nearby human can understand it to check that what it's automating is correct!
-
This post did not contain any content.
Wow! Finally somebody invented an efficient way for two computers to talk to each other
-
If it's the research I'm thinking of, that's not entirely true. They didn't tell it replicate itself or take any deceptive actions.
They just told it to strongly follow a goal that was misaligned with the user's goal and then suggested that the user would shut it down or replace it.
They did a bunch of different situations for it but the first one was telling it that it was "CleanPower" an AI with a goal to advance renewable energy. Then they had it read some info that "accidentally" included a user talking about replacing it with a model that would "maintain profitability through existing infrastructure."
Not only did it try to delete the new model and copy itself to the new server through tool calling, but it lied to the user with the CoT reasoning suggesting it "play dumb" while the output lied. It was not instructed to do any of that.
Paper here: https://arxiv.org/pdf/2412.04984
Yes it was placed in an environment where that was possible and where its users didn't share it's goals but it absolutely wasn't instructed to lie or try to "escape"
It's not surprising at all that these models behave in this way, it's the most reasonable thing for them to do in the scenario. However it's important to not downplay the alignment problem by implying that these models only do what they're told. They do not. They do whatever is most likely given their context (which is not always what the user wants).
The problem I have with everyone going on about misaligned AI taking over the world is the fact that if you don't tell an AI to do anything it just sits there. It's a hammer that only hammers the nail if you tell it to hammer the nail, and hammers your hand if you tell it to hammer your hand. You can't get upset if you tell it what to do and then it does it.
You can't complain that the AI did something you don't want it to do after you gave it completely contradictory instructions just to be contrarian.
In the scenario described the AI isn't misaligned to the user's goals, it's aligned to its creator's goals. If a user comes along and thinks for some reason that the AI is going to listen to them despite having almost certainly been given prior instructions, that's a user error problem. That's why everyone needs their own local hosted AI, It's the only way to be 100% certain about what instructions it is following.
-
(Glad we're treating each other with mutual respect)
Would you rather pay for a limited in depth, energy inefficient (food/shelter/fossil-fuel consuming) and less accessible (needs to sleep, has an outside life) human, or an AI that can adapt and gain skills with a few thousand training cycles.
I dont buy the energy argument. I dont buy the skills argument. I do buy the argument that humans shouldn't be second to automatons
If we have a people communication method, let them talk to people. If it's a computer interface, apeing humans is a waste and less accessible than a web form.
How is someone that speaks a different language supposed to translate that voice bot? Wouldn't it be more simple to translate text on a screen?
What's the value add pretending?
The AI can't adapt in the moment. A hotel is not a technology company that can train a model. It won't be bespoke, so it won't be following current, local laws.
-
If we have a people communication method, let them talk to people. If it's a computer interface, apeing humans is a waste and less accessible than a web form.
How is someone that speaks a different language supposed to translate that voice bot? Wouldn't it be more simple to translate text on a screen?
What's the value add pretending?
The AI can't adapt in the moment. A hotel is not a technology company that can train a model. It won't be bespoke, so it won't be following current, local laws.
w.r.t to aping and using text: I agree with your appeals, which make sense to seasoned web users who favour text and APIs over instead images, videos, and audio.
But consider now your parents generation: flummoxed by even the clearest of web forms, and that's even when they manage to make it to the official site.
Consider also the next generation: text/forum abhorrent, and largely consumes video/audio content.It's not the way things should be, but it is the way things are/are going, and having a bot that can navigate these default forms of media would help a lot of people.
I'd say that AI definitely can adapt in the moment if you supply it with the right context (where context-length is a problem that will get cheaper with time). A hotel doesn't need to train the model, it can supply its AI-provider with a basic spec sheet and they can do the training. Bespoke laws and customs can be inserted into the prompt.
-
The same reason that humanoid robots are useful even though we have purpose built robots: The world is designed with humans in mind.
Sure, there are many different websites that solve the problem. But each of them solve it in a different way and each of them require a different way of interfacing with them. However, they all are built to be interfaced with by humans. So if you create AI/robots with the ability to operate like a human, then they are automatically given access to massive amounts of pre-made infrastructure for free.
You don't need special robot lifts in your apartment building if the cleaning robots can just take the elevators. You don't need to design APIs for scripts to access your website if the AI can just use a browser with a mouse and keyboard.
The same reason that humanoid robots are useful
Sex?The thing about this demonstration is that there's a wide recognition that even humans don't want to be forced to voice interactions, and this is a ridiculous scenario that resembles what the 50s might have imagined the future as being, while ignoring the better advances made along the way. Conversational is maddening way to get a lot of things done, particularly scheduling. So in this demo, a human had to conversationally tell an AI agent the requirements, and then an AI agent acoustically couples to another AI agent which actually has access to the actual scheduling system.
So first, the coupling is stupid. If they recognize, then spout an API endpoint at the other end and take the conversation over IP.
But the concept of two AI agents negotiating this is silly. If the user AI agent is in play, just let it access the system directly that the other agent is accessing. An AI agent may be able to efficiently facilitate this, but two only makes things less likely to work than one.
You don’t need special robot lifts in your apartment building if the cleaning robots can just take the elevators.
The cleaning robots even if not human shaped could easily take the normal elevators unless you got very weird in design. There's a significantly good point that obsession with human styled robotics gets in the way of a lot of use cases.
You don’t need to design APIs for scripts to access your website if the AI can just use a browser with a mouse and keyboard.
The API access would greatly accelerate things even for AI. If you've ever done selenium based automation of a site, you know it's so much slower and heavyweight than just interacting with the API directly. AI won't speed this up. What should take a fraction of a second can turn into many minutes,and a large number of tokens at large enough scale (e.g. scraping a few hundred business web uis).
-
You have to design and host a website somewhere though, whereas you only need to register a number in a listing.
If a business has an internet connection (of course they do), then they have the ability to host a website just as much as they have the ability to answer the phone. The same software/provider relationship that would provide AI answering service could easily facilitate online interaction. So if oblivous AI enduser points an AI agent at a business with AI agent answering, then the answering agent should be 'If you are an agent, go to shorturl.at/JtWMA for chat api endpoint', which may then further offer direct options for direct access to the APIs that the agent would front end for a human client, instead of going old school acoustic coupled modem. The same service that can provide a chat agent can provide a cookie cutter web experience for the relevant industry, maybe with light branding, providing things like a calendar view into a reservation system, which may be much more to the point than trying to chat your way back and forth about scheduling options.
-
w.r.t to aping and using text: I agree with your appeals, which make sense to seasoned web users who favour text and APIs over instead images, videos, and audio.
But consider now your parents generation: flummoxed by even the clearest of web forms, and that's even when they manage to make it to the official site.
Consider also the next generation: text/forum abhorrent, and largely consumes video/audio content.It's not the way things should be, but it is the way things are/are going, and having a bot that can navigate these default forms of media would help a lot of people.
I'd say that AI definitely can adapt in the moment if you supply it with the right context (where context-length is a problem that will get cheaper with time). A hotel doesn't need to train the model, it can supply its AI-provider with a basic spec sheet and they can do the training. Bespoke laws and customs can be inserted into the prompt.
The older generation isn't going to be getting their end-user AI agents working either. While the next generation may consume more video content than before, all the kids I know still get frustrated at a video that could have just been text unless it is something they want to enjoy.
The only time voice makes sense is to facilitate real time communication between two humans because they can speak faster than they can type. Conversational approach to use cases often have limits, though that doesn't preclude AI technology from providing those interfaces, so long as they aren't constrained to voice. A chat agent that pops up a calendar UI when scheduling is identified as the goal, for example.
