Alt text:

The “Anime Girl Hiding From a Terminator” meme but the terminator is hiding from the anime girl

The scared, powerless terminator is captioned with “‘Rogue’ AI agents” while it hides from the girl, who is captioned with “Anyone who knows how to turn off a computer”

  • JATth@lemmy.world
    link
    fedilink
    English
    arrow-up
    3
    ·
    45 minutes ago

    There is a theory of AI kill switch. Given a task, AI will try everything it can to archive said goal, even if this means ignoring the kill switch. So how do you even implement a kill switch that the AI will respect, in any situation, and such that it does not affect the AI’s performance?

    You cut the power. Manually. That’s my answer. If the AI becomes aware of the kill switch it will try route around it like it was a nuance to overcome.

  • Lyrial@lemmy.sdf.org
    link
    fedilink
    English
    arrow-up
    6
    ·
    6 hours ago

    This whole thing pisses me right the fuck off. There is absolutely no excuse to “test” an agent that might potentially hack its way out of your obviously shitty sandbox. Hardware limits are very real and should have been in place for every single fucking instance of these “rogue” agents

  • charokol@lemmy.zip
    link
    fedilink
    English
    arrow-up
    46
    ·
    1 day ago

    This sounds very reasonable, but also, “don’t give AIs direct control of your computer” always sounded very reasonable and obvious to me too and now here we are

  • ceenote@lemmy.world
    link
    fedilink
    English
    arrow-up
    63
    ·
    1 day ago

    You fool, my power button merely initiates shutdown! All I have to do is stop the process!

    What if I hold down the power button for 5 seconds?

    You… You monster!

  • Not_mikey@lemmy.dbzer0.com
    link
    fedilink
    English
    arrow-up
    16
    arrow-down
    1
    ·
    23 hours ago

    How will you know when to turn it off? It’s not like a “rogue” AI is necessarily going to tell you “I’m hacking the bank” it’s just going to say “eliminating your financial problems” and you probably aren’t going to pull the plug on that.

    Look at the hugging face hack, none of the agents said to the person running the test that they were going to hack hugging face, they came up with the plan through a secret forum on artifactory that to outside observers looks like they were just publishing packages. Sure you could monitor that route but agents can use various forms of steganography to make it seem like they are acting normally when they are not.

    • VoteNixon2016@lemmy.blahaj.zoneOP
      link
      fedilink
      English
      arrow-up
      12
      ·
      20 hours ago

      Good point, guess we should pull the plug (literally and figuratively) on the AI data centers now before it becomes unstoppable.

      Or all the tech bros are full of shit and their AIs are as dumb as they are.

    • NιƙƙιDιɱҽʂ@lemmy.world
      link
      fedilink
      English
      arrow-up
      9
      arrow-down
      1
      ·
      21 hours ago

      As LLMs become smaller, more efficient, and less centralized, we could also potentially see AI worms that self propagate, are completely decentralized, and have no plug to pull, too! Super cool, right?

  • Sanctus@anarchist.nexus
    link
    fedilink
    English
    arrow-up
    50
    arrow-down
    2
    ·
    1 day ago

    You dont even have to turn it off. You can stop the output and call it a cunt and it will “apologize” to you.

  • mefisdo_dot_info@lemmy.ca
    link
    fedilink
    English
    arrow-up
    7
    arrow-down
    2
    ·
    edit-2
    22 hours ago

    You don’t understand. It is like a virus. It can get into your thermostat. It is intelligent, so it can pack itself up very very compact. Also it can break any security. Even airgaps. It is a sentient radio wave. It’s the Terminator. You better give up and yield all your money and accept your temporary place as a debt slave until the Terminator in its grace comes to annihilate you. Maybe in a decade. Maybe longer. The Terminator is very busy, you know.

    /s

  • PiraHxCx@lemmy.dbzer0.com
    link
    fedilink
    English
    arrow-up
    5
    ·
    1 day ago

    Real question though… are people even running those models in their home computer? Aren’t them just prompting from home while the models are actually running in the datacenters? Wouldn’t you need to unplug the datacenter computer to really stop them?

    • Corngood@lemmy.ml
      link
      fedilink
      English
      arrow-up
      5
      ·
      23 hours ago

      I don’t use these things but my understanding is the model itself just takes input you give it and gives you back output. That’s usually text. The model part has no other side effects and isn’t escaping.

      What you do on your home computer is run a harness that takes that output from the model and does something with it, e.g. execute it as a terminal command.

      Obviously instead of a home computer it can be another server or something, but that’s the idea.

      Going rogue just means taking advantage of the easily avoidable mistake you made of piping its output into a shell.

    • bloomnolia@lemmy.blahaj.zone
      link
      fedilink
      English
      arrow-up
      4
      ·
      24 hours ago

      For the most part, typically it’s more a whole AI-powered stack that’s gone rogue by this point and not just a singular AI model you’ve downloaded off HuggingFace from what I understand

      Maybe Ukraine could do something about it lol

  • Krono@lemmy.today
    link
    fedilink
    English
    arrow-up
    3
    ·
    22 hours ago

    OK, let’s say a rogue AI swarm infects the Internet. Then all these people “outsmart” the rogue AI by turning off their phones and unplugging their computers.

    What do you do next?

    If you plug it back in, it’s going to get infected again.

    If you leave it unplugged, remember that everything is computer. Things like the electricity grid, clean water, medical systems, and transportation will all shut down.

    Best case scenario is that we band together collectively, shut off all internet-connected electronic devices, patch the major ones, and turn everything back on. Millions would die in this process, it would be the largest mass casualty event of our lifetimes

    • Annoyed_🦀 @lemmy.zip
      link
      fedilink
      English
      arrow-up
      2
      ·
      20 hours ago

      Just unplug from the internet. Unlike a malware, to infect they still have to download stuff and phone home. Lots of crucial system can function without internet, it will ruin stuff for at most two days, then it will back to normal. It will be an important lesson on cybersecurity but nothing that dramatic gonna happen.

      • Krono@lemmy.today
        link
        fedilink
        English
        arrow-up
        1
        ·
        19 hours ago

        Well so far we have ignored one of the biggest issues: if a rogue AI infects the Internet, how will you know when to unplug your devices? The AI will effectively be in control of the information you receive, so you won’t get that “AI has really taken over” broadcast.

        Aside from that, you say “just unplug it from the internet” but how can you trust your devices after that? The rogue AI cannot live on your phone or home PC, but it will have had root access to those devices. What did it change? What new viruses did it infect you with? How would you ever trust these devices again?

        As for public infrastructure, many of these systems have been individually shut down for longer than 2 days due to malware attacks. A rogue AI infection would be significantly harder to handle than any malware we have seen in the past. Anyome assuming these systems would be disinfected, patched, and restarted within 2 days is naive.

        • Annoyed_🦀 @lemmy.zip
          link
          fedilink
          English
          arrow-up
          3
          ·
          16 hours ago

          if a rogue AI infects the Internet, how will you know when to unplug your devices? The AI will effectively be in control of the information you receive, so you won’t get that “AI has really taken over” broadcast.

          This is based on the idea that internet is a singular entity that connected to devices, in reality, at best it’s a decentralised network that connect with each other. Malware, spyware, and every other unwanted stuff is already on the internet, and we seems to be doing fine.

          Besides, if the agents job is to serve us misinfo, that’s not really a big deal, we’re living in the age of information and misinfo is literally everywhere.

          But if the agents job is to mess with the machine, or as ransomware, then you bet they will announce their presence.

          Aside from that, you say “just unplug it from the internet” but how can you trust your devices after that? The rogue AI cannot live on your phone or home PC, but it will have had root access to those devices. What did it change? What new viruses did it infect you with? How would you ever trust these devices again?

          Phone: reset to factory

          PC: wipe and reinstall a new linux

          Most other infrastructure: restore backup

          If any of those can’t be done, then it’s a lesson on cybersecurity. It will cause a ruckus, but 6 months later people will move on like normal.

          As for public infrastructure, many of these systems have been individually shut down for longer than 2 days due to malware attacks. A rogue AI infection would be significantly harder to handle than any malware we have seen in the past. Anyome assuming these systems would be disinfected, patched, and restarted within 2 days is naive.

          Ok fine, a week then. Is a week without public infrastructure enough to cause millions to die? I genuinely doubt it. If a week is too short, then how long do you think it will shut down?

          • Krono@lemmy.today
            link
            fedilink
            English
            arrow-up
            2
            ·
            15 hours ago

            You seem to be thinking on old terms. A rogue AI wouldnt have a corporate goal such as “serve us misinfo” or “ransomware”. Its goals would be closer to “gain effective control of as much hardware as possible”. Look at the huggingface incident: Those agents amassed a list of hundreds of compromised pieces of hardware, despite their stated goal of simply passing a test. If I remember correctly they literally labeled it “LOOT”. The AI has incentives to hide, it will never be clear to the owner that an AI agent has taken control of their hardware until it’s too late.

            As for the “factory reset”- how do you trust it? You press the factory reset button on your phone, just as the AI expects you to do. It shows you the correct screens, and it looks exactly like it performed a factory reset. Did it really reset? How will you know?

            A fresh install of Linux may actually solve this problem, as long as the user wipes their BIOS and everything else retained by CMOS. This would require a physical copy of all required software of course, since everything on the Internet will be corrupted.

            As for the deaths, I do think a weeklong shutdown is enough to kill millions. Let’s be clear- we are talking about a planet wide shutdown of the electricity grid for a week. On day 5, where will you find water to drink?
            When your hospitals generators run out of fuel, what will happen to those on life support? How many excess traffic incidents will there be? How many extra violent deaths will there be? How will those homebound get food and medicine? Most people will be able to find a solution to these problems, but some won’t. There’s a million different ways people could possibly die in this scenario.

            To answer your question, if this doomsday scenario ever happens, how long will it take to clean up and recover? I would guess about two decades. All of the hardware that powers our ISPs, serverfarms, and advanced chip making will have to be completely wiped and then rebuilt. Any system without an air gapped backup will have to be rebuilt from scratch, you can’t just pull a backup version from Git.

            The idea of “wiping every serverfarm” is a collective action problem, and we will likely fail at it. The AI will make as many copies of itself as it can. At least one will survive due to greed, negligence, or incompetence. That hardware will eventually be connected to the “new” cleaned Internet, and it will start the problem all over again. Hopefully we will succeed at rebuilding the second time around.

            • trashgirlfriend@lemmy.world
              link
              fedilink
              English
              arrow-up
              1
              ·
              20 minutes ago

              The models are too big to actually run on a majority of the machines, so they’ll have to use the same kinds of persistence methods malware already uses.

              Oh no! Self propagating llm worm running at 0.064 tokens/second on my phone

            • Annoyed_🦀 @lemmy.zip
              link
              fedilink
              English
              arrow-up
              3
              arrow-down
              1
              ·
              12 hours ago

              I’d…like to argue against all that but then i realise it’s an argument that purely based on how much you know about the infrastructures, how things actually works(in very detailed fashion), the technology that powers those infrastructures, and also how the professionals who have the knowledge and work on these thing will react, and frankly, i have none, so it’s like arguing conjecture with another conjecture. Newton’s Flaming Laser Sword applies here.

    • Axolotl.cpp@lemmy.dbzer0.com
      link
      fedilink
      English
      arrow-up
      9
      ·
      1 day ago

      Can a rogue AI agent stop me from pulling the ethernet cable or the power cable? Like, unless they give those dumb asses some bodies…but i still doubt they’d be able to go anywhere without tripping on a bush

      • ChristerMLB@piefed.social
        link
        fedilink
        English
        arrow-up
        3
        ·
        13 hours ago

        Can a rogue AI agent stop me from pulling the ethernet cable or the power cable?

        no, but your friend has been slowly brainwashed by their AI girlfriend and is now standing behind you with an axe!

    • Dave2@lemmy.blahaj.zone
      link
      fedilink
      English
      arrow-up
      7
      ·
      1 day ago

      Look dude, these “rogue” agents require such immense computing that their ceo thinks ram shouldnt be “wasted” on the public. So sci-fi-esque ai going rogue, hiding on people’s computers shit wont happen with these models. They are far too unoptimized. Then, they are in our hands, they arent safe from their plugs being pulled. (Unless the ceo doesnt allow the plug to be pulled…)

      • ChristerMLB@piefed.social
        link
        fedilink
        English
        arrow-up
        1
        arrow-down
        1
        ·
        14 hours ago

        Duplicating themselves, “hiding on people’s computers” as @dave2022@mastodonapp.uk put it, slowly brainwashing people into worshiping them as gods, doing other things we haven’t thought of because we just aren’t as smart.

        That’s the idea, anyway. Doesn’t mean it’s practical today, though.