Can humans reliably prevent advanced AI systems from escaping using sandboxing and related security measures, even when those systems have internet access?

“The first problem is the isolation, the containment wasn't good enough. If the isolation and containment was good enough, that technology would be sitting in a lab doing whatever it's doing, and we'd all be fine.”

“when you're testing software, whatever you do, these algorithms, they're optimizing towards a, an objective, and you, when you're testing it, you have to make sure that it's isolated, it's contained, it's sandboxed. The containment of it, the isolation of it has to, has to be done well, and there's good computer science there.”

“software breaks out of sandboxes all the time. That's the reason why we need a virtual machines. You can't have agents monitor, their own sandbox monitoring themselves. You, you need a, you need a, if you will, a whole bunch of watchdogs.”

“the sandboxes that work against human adversaries are not necessarily going to work against AI agents that are able to find new security vulnerabilities on the spot. And so that means that the sandboxes themselves will have to be pre-hardened by having these AI agents trying to break them. So that becomes AI versus AI to some degree”

“the sandboxes themselves will have to be pre-hardened by having these AI agents trying to break them. So that becomes AI versus AI to some degree”

“If any AI lab wants to make that bet, that's great. Let's go make the bet. I'll put together a team of like even just kind of like mediocre security engineers. They'll go build it and you can run it and I won't have put it underground. I'll put it in a building with internet and it won't break out.”

“I would go further and say like, well, like, yeah, at least they tried. Because I know Doomer's focused very heavily on like the, like Lab scenario, but I think like another major risk comes from just like AI out in the world, like most companies aren't sandboxing at all.”

“most companies aren't sandboxing at all. And that's, of course, deeply problematic as well”

“When we wrote early papers in AI safety, we talked about things not to do. They were obviously unsafe, and the system would escape. Don't connect it to internet. Don't give random users access to the training data. Basically, the whole list was like a set of instructions. They read it and went, those are great ideas. We're going to”

“they're creating so many kinds of sandboxes and learning environments in order to train models that have to do things that are so general, many things that have not been done before by a computer program, that the human beings don't really know, certainly not at this speed, how to make sure every sandbox is going to be, you know,”

“Like what you're saying is like you're imagining that maybe in the best case scenario humans can defeat AI with monitoring, which I don't even agree with, but even let's say you're right, okay? And let's say OpenAI gets their act together, which I doubt.”

“software breaks out of sandboxes all the time. That's the reason why we need a virtual machines. You can't have agents monitor, their own sandbox monitoring themselves. You, you need a, you need a, if you will, a whole bunch of watchdogs.”

“If any AI lab wants to make that bet, that's great. Let's go make the bet. I'll put together a team of like even just kind of like mediocre security engineers. They'll go build it and you can run it and I won't have put it underground. I'll put it in a building with internet and it won't break out.”

“when you're testing software, whatever you do, these algorithms, they're optimizing towards a, an objective, and you, when you're testing it, you have to make sure that it's isolated, it's contained, it's sandboxed. The containment of it, the isolation of it has to, has to be done well, and there's good computer science there.”

“the sandboxes themselves will have to be pre-hardened by having these AI agents trying to break them. So that becomes AI versus AI to some degree”

“the sandboxes that work against human adversaries are not necessarily going to work against AI agents that are able to find new security vulnerabilities on the spot. And so that means that the sandboxes themselves will have to be pre-hardened by having these AI agents trying to break them. So that becomes AI versus AI to some degree”

“I would go further and say like, well, like, yeah, at least they tried. Because I know Doomer's focused very heavily on like the, like Lab scenario, but I think like another major risk comes from just like AI out in the world, like most companies aren't sandboxing at all.”

“most companies aren't sandboxing at all. And that's, of course, deeply problematic as well”

“When we wrote early papers in AI safety, we talked about things not to do. They were obviously unsafe, and the system would escape. Don't connect it to internet. Don't give random users access to the training data. Basically, the whole list was like a set of instructions. They read it and went, those are great ideas. We're going to”

“they're creating so many kinds of sandboxes and learning environments in order to train models that have to do things that are so general, many things that have not been done before by a computer program, that the human beings don't really know, certainly not at this speed, how to make sure every sandbox is going to be, you know,”

“Like what you're saying is like you're imagining that maybe in the best case scenario humans can defeat AI with monitoring, which I don't even agree with, but even let's say you're right, okay? And let's say OpenAI gets their act together, which I doubt.”