I attended a federal contracting conference a few months ago, and they had one of these things (or a variant) walking around the lobby.
From talking to the guy who was babysitting it, they can operate autonomously in units or be controlled in a general way (think higher level unit deployment and firing policies rather than individual remote control) given a satellite connection. In a panel at the same conference, they were discussing AI safety, and I asked:
Given that AI seems to be developing from less complex tasks like chess (which is still complicated, obviously, but a constrained problem) to more complex and ill-defined tasks like image generation, it seems that it’s inevitable that we will develop AI capable of providing strategic or tactical plans, if we haven’t already. If two otherwise-equally-matched military units are fighting, it seems reasonable to believe that the one using an AI to make decisions within seconds would win over the one with human leadership, simply because they would react more quickly to changing battlefield conditions. This would place an enormous incentive on the US military to adopt AI assisted strategic control, which would likely lead to units of autonomous weapons which are also controlled by an autonomous system. Do any of you have any concerns about this, and if so, do you have any ideas about how we can mitigate the problem.
(Paraphrasing, obviously, but this is close)
The panel members looked at each other, looked at me, smiled, shrugged, and didn’t say anything. The moderator asked them explicitly if they would like to respond, and they all declined.
I think we’re at the point where an AI could be used to create strategies, and I would be very surprised if no one were trying to do this. We already have autonomous weapons, and it’s only a matter of time before someone starts putting them together. Yeah, they will generally act reasonably, because they’ll be trained on human tactics in a variety of scenarios, but that will be cold comfort to dead civilians who happened to get in the way of a hallucinating strategic model.
EDIT: I know I’m not actually addressing anything you said, but you seem to have thought about this a bit, and I was curious about what you thought of this scenario.
My guess is that they didn’t answer your question because they had strict instructions not to stray from the script on this topic. Saying the wrong thing could lead to a big PR problem, so I don’t expect that people working in this field would be willing to have a candid public discussion even about topics to which they have given a lot of thought. I do expect that they have given the ability of AI to obey orders accurately a lot of thought at least due to practical (if not ethical) concerns.
I mean, I am currently willing to say “the AIs will almost definitely kill civilians but we should build them anyway” because I don’t work in defense. However, even I’m a little nervous saying that because one day I might want to. My friends who do work in defense have told me that the people who gave them clearance did investigate their online presence. (My background is in computational biochemistry but I look at what’s going on in AI and I feel like nothing else is important in comparison.)
As for cold comfort: I think autonomous weapons are inevitable in the same way that the atom bomb was inevitable. Even if no one wants to see it used, everyone wants to have it because enemies will. However, I don’t see a present need for strategic (as opposed to tactical) automation. A computer would have an advantage in battlefield control but strategy takes hours or days or years and so a human’s more reliable ability to reason would be more important in that domain.
Once a computer can reason better than a human can, that’s the end of the world as we know it. It’s also inevitable like the atom bomb.
I attended a federal contracting conference a few months ago, and they had one of these things (or a variant) walking around the lobby.
From talking to the guy who was babysitting it, they can operate autonomously in units or be controlled in a general way (think higher level unit deployment and firing policies rather than individual remote control) given a satellite connection. In a panel at the same conference, they were discussing AI safety, and I asked:
(Paraphrasing, obviously, but this is close)
The panel members looked at each other, looked at me, smiled, shrugged, and didn’t say anything. The moderator asked them explicitly if they would like to respond, and they all declined.
I think we’re at the point where an AI could be used to create strategies, and I would be very surprised if no one were trying to do this. We already have autonomous weapons, and it’s only a matter of time before someone starts putting them together. Yeah, they will generally act reasonably, because they’ll be trained on human tactics in a variety of scenarios, but that will be cold comfort to dead civilians who happened to get in the way of a hallucinating strategic model.
EDIT: I know I’m not actually addressing anything you said, but you seem to have thought about this a bit, and I was curious about what you thought of this scenario.
My guess is that they didn’t answer your question because they had strict instructions not to stray from the script on this topic. Saying the wrong thing could lead to a big PR problem, so I don’t expect that people working in this field would be willing to have a candid public discussion even about topics to which they have given a lot of thought. I do expect that they have given the ability of AI to obey orders accurately a lot of thought at least due to practical (if not ethical) concerns.
I mean, I am currently willing to say “the AIs will almost definitely kill civilians but we should build them anyway” because I don’t work in defense. However, even I’m a little nervous saying that because one day I might want to. My friends who do work in defense have told me that the people who gave them clearance did investigate their online presence. (My background is in computational biochemistry but I look at what’s going on in AI and I feel like nothing else is important in comparison.)
As for cold comfort: I think autonomous weapons are inevitable in the same way that the atom bomb was inevitable. Even if no one wants to see it used, everyone wants to have it because enemies will. However, I don’t see a present need for strategic (as opposed to tactical) automation. A computer would have an advantage in battlefield control but strategy takes hours or days or years and so a human’s more reliable ability to reason would be more important in that domain.
Once a computer can reason better than a human can, that’s the end of the world as we know it. It’s also inevitable like the atom bomb.