5 ms·

> What measurable predictions could it make? > Given that the state-of-the-art literature on AI containment answers in the affirmative (AI is uncontainable in

> What measurable predictions could it make? > Given that the state-of-the-art literature on AI containment answers in the affirmative (AI is uncontainable in the long-term), we conclude that it should be possible to escape from the simulation, at least with the help of superintelligent AI. By contraposition, if escape from the simulation is not possible, containment of AI should be.

jack_pp 2y ago
I don't think anyone is saying that AI containment is impossible, rather than a risk we shouldn't take lightly. For example let's say we take it seriously and only run GAI in a bunker with a self destruct mechanism and no internet connection, only interfacing with it via a strict chat interface. We could also have a lesser model like a gpt-4 screen all chat for possible attempts by the ai to convince people to let it out of its box. How exactly is it gonna get out?
- OutOfHere 2y ago
  It will convince you to break your own rules, also to disable the screener. That's if it doesn't hack you by writing for you software that later opens the door for it to get out.
- ASalazarMX 2y ago
  An AGI could have enough value to run a country, or our whole civilization, optimally. Are you proposing to manage human civilization via chat?
- creer 2y ago
  It will actively take part in the design and construction of its next, superior bunker. (See construction of embassies.)
- creer 2y ago
  It will convince humanity that an AI is conscious and should have personhood and civil rights - same as other human classes which didnt for a while.
- creer 2y ago
  It will convince humanity that it would be safe running a generation ship of volunteers (it says it will guarantee their civil rights) and will then leave humanity behind while colonizing it's own section of the galaxy.
bastawhiz 2y ago
> By contraposition, if escape from the simulation is not possible, containment of AI should be. This is a flawed argument, though. If we live in a simulation, our lived experiences suggest that there's no information coming into the simulation from the outside. AI, on the other hand, has a bidirectional interface: the whole point is that we communicate with it from the outside. If we created an AI that we listened to but never replied to—assuming the actions we took as a result of listening to the AI never provided it with feedback—it's as good as contained. We'd have to assume that not only is someone outside the simulation listening to us, but that they actually have a mechanism for replying. Information from inside the simulation escaping is sort of the whole point of a simulation (if you can't observe what you're simulating, why bother?). It's also silly to think that we could affect the "outside". The idea of AI escaping is premised on us developing something that convinces us to do what it wants. We're simulating just a mind. The simulation we're in—if we are in a simulation—simulates every quark in every particle in every atom in every molecule in every chunk of rock in every galaxy in the universe. We're talking about building AIs that could escape while just barely managing to overcome spooky CUDA errors, while our world is an emergent fleck of dust in a corner of their simulation. The point being: if they can simulate our entire universe, they can also simulate just a mind (in the same way that we are) that's literally infinitely more powerful (which is to say, backed by the computing resources that we consider to be "the infinite") which can out-think whatever we shout out of the simulation. Or inspect the state of the simulation and discover our true intentions. Or whatever. Said another way, we're afraid of super intelligent AI because it's running on theoretical compute that is as powerful as we can come up with. We (and our AIs) are compute running in a rounding error of a fraction of whatever their compute is. We could build Dyson spheres around every star in our galaxy to power a computer running an unfathomably super intelligent AI and it would still be quadrillions (more?) of orders of magnitude less powerful than the hardware simulating our universe. Therefore, the property that makes escape impossible for us isn't the interface, it's that we can't outsmart the entities outside. The playing field is different inside the simulation, because our AI is able to be smarter than us.
- andyjohnson0 2y ago
  > If we live in a simulation, all evidence suggests that there's no information coming into the simulation from the outside. I'm wondering what evidence (or absence thereof) you're referring to here? Thermodynamics? Would be interesting to know more.

jack_pp 2y ago

I don't think anyone is saying that AI containment is impossible, rather than a risk we shouldn't take lightly. For example let's say we take it seriously and only run GAI in a bunker with a self destruct mechanism and no internet connection, only interfacing with it via a strict chat interface. We could also have a lesser model like a gpt-4 screen all chat for possible attempts by the ai to convince people to let it out of its box. How exactly is it gonna get out?

OutOfHere 2y ago
It will convince you to break your own rules, also to disable the screener. That's if it doesn't hack you by writing for you software that later opens the door for it to get out.
ASalazarMX 2y ago
An AGI could have enough value to run a country, or our whole civilization, optimally. Are you proposing to manage human civilization via chat?
creer 2y ago
It will actively take part in the design and construction of its next, superior bunker. (See construction of embassies.)
creer 2y ago
It will convince humanity that an AI is conscious and should have personhood and civil rights - same as other human classes which didnt for a while.
creer 2y ago
It will convince humanity that it would be safe running a generation ship of volunteers (it says it will guarantee their civil rights) and will then leave humanity behind while colonizing it's own section of the galaxy.

bastawhiz 2y ago

> By contraposition, if escape from the simulation is not possible, containment of AI should be. This is a flawed argument, though. If we live in a simulation, our lived experiences suggest that there's no information coming into the simulation from the outside. AI, on the other hand, has a bidirectional interface: the whole point is that we communicate with it from the outside. If we created an AI that we listened to but never replied to—assuming the actions we took as a result of listening to the AI never provided it with feedback—it's as good as contained. We'd have to assume that not only is someone outside the simulation listening to us, but that they actually have a mechanism for replying. Information from inside the simulation escaping is sort of the whole point of a simulation (if you can't observe what you're simulating, why bother?). It's also silly to think that we could affect the "outside". The idea of AI escaping is premised on us developing something that convinces us to do what it wants. We're simulating just a mind. The simulation we're in—if we are in a simulation—simulates every quark in every particle in every atom in every molecule in every chunk of rock in every galaxy in the universe. We're talking about building AIs that could escape while just barely managing to overcome spooky CUDA errors, while our world is an emergent fleck of dust in a corner of their simulation. The point being: if they can simulate our entire universe, they can also simulate just a mind (in the same way that we are) that's literally infinitely more powerful (which is to say, backed by the computing resources that we consider to be "the infinite") which can out-think whatever we shout out of the simulation. Or inspect the state of the simulation and discover our true intentions. Or whatever. Said another way, we're afraid of super intelligent AI because it's running on theoretical compute that is as powerful as we can come up with. We (and our AIs) are compute running in a rounding error of a fraction of whatever their compute is. We could build Dyson spheres around every star in our galaxy to power a computer running an unfathomably super intelligent AI and it would still be quadrillions (more?) of orders of magnitude less powerful than the hardware simulating our universe. Therefore, the property that makes escape impossible for us isn't the interface, it's that we can't outsmart the entities outside. The playing field is different inside the simulation, because our AI is able to be smarter than us.

andyjohnson0 2y ago
> If we live in a simulation, all evidence suggests that there's no information coming into the simulation from the outside. I'm wondering what evidence (or absence thereof) you're referring to here? Thermodynamics? Would be interesting to know more.