4 ms·

"It does not have a physical embodiment (other than living on a computer screen), but it can control existing physical tools, robots, or laboratory equipment th

"It does not have a physical embodiment (other than living on a computer screen), but it can control existing physical tools, robots, or laboratory equipment through a computer; in theory it could even design robots or equipment for itself to use." I think I get where the author is coming from, the AI would be in the cloud. But it bears repeating, the cloud is somebody else's computers, software has a physical embodiment, period. This is not a philosophical nitpick, it's important because you can pull the plug (or nuke the datacenter) if necessary.

bathtub365 2y ago
I wonder if there’s a word that describes the property of software where it isn’t tied to the hardware that it’s currently running on and is endlessly copyable with almost no effort
- TZubiri 2y ago
  For the first half of that sentence, portable? I'm not sure what the argument here is, Operating systems are now sentient an ethereal? You could make the same argument for books for that matter, I guess they do exist in a form that is unbounded by the physical books that they reside in, they live in the minds of people. Whatever. And the fact that they can be copied for "no effort" is first wrong, as copying would require energy, which is the other aspect of corporeality, you need not just matter but energy for the software to exist and express itself in the world. What are we even debating? Have we lost our minds? What is mind? Doesn't matter. What is matter? Never mind
xpe 2y ago
Yes, and physical embodiments vary considerably. When software can relocate its source code and data in seconds or less, containment strategies begin to look increasingly bleak. The field of AI safety has written extensively about misunderstandings and overoptimism regarding off switches and boxing.
- TZubiri 2y ago
  Hard? sure. Impossible? No This is a non-trivial error. Let's not dive into the geopolitics of kill switches and decentralized systems when we are discussing trivial mistakes. AI, like any software, is not an ethereal being, like us it is limited by our mortal coils. This is true for extant systems, if anyone, layman or PHD were to say that TikTok or facebook, wechat or bitcoin or ethereum, USD tether has no physicial embodiment, whether for recreational discussion, or whether in court, or in a professional capacity. The answer needs to be that they are wrong, clear and simple.
  - xpe 2y ago
    Of course I grant the above point that software systems are not "ethereal". They have a physical form: often as electromagnetic states in matter. I hope my contribution to the conversation was clear: the nature of the physical embodiment matters. Things that are easy and cheap to copy are harder to contain. Containment of intelligent systems is very difficult, based on what we know about human flaws and security systems. >> When software can relocate its source code and data in seconds or less, containment strategies begin to look increasingly bleak. > Hard? sure. Impossible? No I don't need to claim that containment is impossible, just really hard. I'm interested in planning across a range of scenarios. Given human foibles, we should plan for scenarios where AIs get out of the box and can spread rapidly and widely. See also: "Guidelines for Artificial Intelligence Containment" at https://arxiv.org/pdf/1707.08476 https://arxiv.org/pdf/1707.08476 Imagine two possible AI containment failure scenarios. In the first, say it is feasible to "shut everything down" by disconnecting power. In the second, say humans have to resort to a combination of kinetic force and cyberattacks. For both, a likely next step would resemble the global health to eradicate smallpox. It would require tremendous effort and cost to bring the system back online while proving that no copies of the rogue AI exist. Would such coordinated responses by humanity be possible? Possible, yes. Likely? Estimating this is hard. I, for one, am not optimistic when it comes to international cooperation generally, much less for ones involving high complexity.

bathtub365 2y ago

I wonder if there’s a word that describes the property of software where it isn’t tied to the hardware that it’s currently running on and is endlessly copyable with almost no effort

TZubiri 2y ago
For the first half of that sentence, portable? I'm not sure what the argument here is, Operating systems are now sentient an ethereal? You could make the same argument for books for that matter, I guess they do exist in a form that is unbounded by the physical books that they reside in, they live in the minds of people. Whatever. And the fact that they can be copied for "no effort" is first wrong, as copying would require energy, which is the other aspect of corporeality, you need not just matter but energy for the software to exist and express itself in the world. What are we even debating? Have we lost our minds? What is mind? Doesn't matter. What is matter? Never mind

xpe 2y ago

Yes, and physical embodiments vary considerably. When software can relocate its source code and data in seconds or less, containment strategies begin to look increasingly bleak. The field of AI safety has written extensively about misunderstandings and overoptimism regarding off switches and boxing.

TZubiri 2y ago
Hard? sure. Impossible? No This is a non-trivial error. Let's not dive into the geopolitics of kill switches and decentralized systems when we are discussing trivial mistakes. AI, like any software, is not an ethereal being, like us it is limited by our mortal coils. This is true for extant systems, if anyone, layman or PHD were to say that TikTok or facebook, wechat or bitcoin or ethereum, USD tether has no physicial embodiment, whether for recreational discussion, or whether in court, or in a professional capacity. The answer needs to be that they are wrong, clear and simple.
- xpe 2y ago
  Of course I grant the above point that software systems are not "ethereal". They have a physical form: often as electromagnetic states in matter. I hope my contribution to the conversation was clear: the nature of the physical embodiment matters. Things that are easy and cheap to copy are harder to contain. Containment of intelligent systems is very difficult, based on what we know about human flaws and security systems. >> When software can relocate its source code and data in seconds or less, containment strategies begin to look increasingly bleak. > Hard? sure. Impossible? No I don't need to claim that containment is impossible, just really hard. I'm interested in planning across a range of scenarios. Given human foibles, we should plan for scenarios where AIs get out of the box and can spread rapidly and widely. See also: "Guidelines for Artificial Intelligence Containment" at https://arxiv.org/pdf/1707.08476 https://arxiv.org/pdf/1707.08476 Imagine two possible AI containment failure scenarios. In the first, say it is feasible to "shut everything down" by disconnecting power. In the second, say humans have to resort to a combination of kinetic force and cyberattacks. For both, a likely next step would resemble the global health to eradicate smallpox. It would require tremendous effort and cost to bring the system back online while proving that no copies of the rogue AI exist. Would such coordinated responses by humanity be possible? Possible, yes. Likely? Estimating this is hard. I, for one, am not optimistic when it comes to international cooperation generally, much less for ones involving high complexity.