If we tell an AI not to invent nanotechnology, not to send anything to protein labs, not to hack into all of the world’s computers, not to design weird new quantum particles, not to do 100 of the other most dangerous and weirdest things we can think of, and then ask it to generalize and learn not to do things of that sort
I had the exact same thought. My guess would be that Eliezer might say that since the AI is maximising if the generalisation function misses even one action of this sort as something that we should exclude that we’re screwed.
Sure, I agree! If we miss even one such action, we’re screwed. My point is that if people put enough skill and effort into trying to catch all such actions, then there is a significant chance that they’ll catch literally all the actions that are (1) world-ending and that (2) the AI actually wants to try.
There’s also a significant chance we won’t, which is quite bad and very alarming, hence people should work on AI safety.
One of my assumptions is that it’s possible to design a “satisficing” engine—an algorithm that generates candidate proposals for a fixed number of cycles, and then, assuming at least one proposal with estimated utility greater than X has been generated within that amount of time, selects one of the qualifying proposals at random. If there are no qualifying candidates, the AI takes no action.
If you have a straightforward optimizer that always returns the action with the highest expected utility, then, yeah, you only have to miss one “cheat” that improves “official” utility at the expense of murdering everyone everywhere and then we all die. But if you have a satisficer, then as long as some of the qualifying plans don’t kill everyone, there’s a reasonable chance that the AI will pick one of those plans. Even if you forget to explicitly penalize one of the pathways to disaster, there’s no special reason why that one pathway would show up in a large majority of the AI’s candidate plans.
I suspect we’re talking about two different things.
If you just naively program a super-intelligent AI to satisfice a goal, then, sure, most of the candidate pathways to satisfice will involve accruing a lot of some type of power, because power is useful for achieving goals. That’s a valid point, and it’s important to understand that merely switching from optimizers to satisficers won’t adequately protect us against overly ambitious AIs.
However, that doesn’t mean that it’s futile to explicitly penalize most (but not literally all) of the paths that the AI could take to accumulate too much power. Suppose you adequately penalize all of the tactics that would have catastrophic side effects except for, oh, I don’t know, cornering the world market on all types of grain, because you forgot to think of that one particular path to power. Would most of the candidate plans that the AI submits for human approval then turn out to involve secretly cornering the grain market?
I don’t see why they would. All else being equal, sure, the most powerful tactic available is going to be somewhat more attractive than other, less powerful tactics. But how much more attractive? Would an AI generate 1,000x more plans that involve one specific globally disruptive tactic like cornering the grain market vs. all of the various tactics that are only locally disruptive, like publishing propaganda or hiring away competing researchers or hacking into competing research labs or interrupting the flow of electricity to those labs? 10x more such plans? 2x more such plans? I don’t think that’s the kind of estimate you can make just by gesturing at basic principles of game theory; you’d need some concrete domain knowledge about the AI’s specific planning algorithms.
If the truly dangerous plans are only 10x more common on the initial brainstorming list, then we can probably make sure those plans aren’t chosen by filtering for plans that the AI rates as safe and transparent. We can use 3 different AIs on airgapped equipment—one that only “cares” about evaluating whether a plan is safe (i.e., would the plan cause unacceptably large changes if implemented?), one that only “cares” about evaluating whether a plan is transparent (i.e., can the plan be explained to humans so that the humans will have an accurate understanding of what actions will be taken as part of the plan?), and one that generates plans that satisfice for a substantive goal (in this case, shutting down opposing AI research).
Of the plans that are approved by all 3 AIs, most of them probably won’t involve accumulating excessive amounts of power, because usually if you accumulate too much power, then that is (a) unsafe, and (b) likely to involve lying to ‘your’ human programmers.
If powerful plans are 10x more common on the initial brainstorming list, then you need filters for safety and/or transparency that are at least 90% effective at catching unacceptably powerful plans in order to get back up to a 50% chance of humanity surviving. If powerful plans are more than 10x more common, then you need a set of filters that’s more reliable than 90%. I don’t think we have any idea what those numbers are yet, but I do think it’s worth trying to reduce how common it is for excessively powerful plans to show up on the initial brainstorming list, and I think we can do that by training AIs to recognize dangerously disruptive plans and to try to avoid those types of plans. It’s better to at least try to get AIs to engage with the concept of “this plan is too disruptive” then to throw up our hands and say, “Oh, power is an attractor in game theory space, so there’s no possible way to get brilliant AIs that don’t seize infinite power.”
I had the exact same thought. My guess would be that Eliezer might say that since the AI is maximising if the generalisation function misses even one action of this sort as something that we should exclude that we’re screwed.
Sure, I agree! If we miss even one such action, we’re screwed. My point is that if people put enough skill and effort into trying to catch all such actions, then there is a significant chance that they’ll catch literally all the actions that are (1) world-ending and that (2) the AI actually wants to try.
There’s also a significant chance we won’t, which is quite bad and very alarming, hence people should work on AI safety.
Hmm… It seems much, much harder to catch every single one than to catch 99%.
One of my assumptions is that it’s possible to design a “satisficing” engine—an algorithm that generates candidate proposals for a fixed number of cycles, and then, assuming at least one proposal with estimated utility greater than X has been generated within that amount of time, selects one of the qualifying proposals at random. If there are no qualifying candidates, the AI takes no action.
If you have a straightforward optimizer that always returns the action with the highest expected utility, then, yeah, you only have to miss one “cheat” that improves “official” utility at the expense of murdering everyone everywhere and then we all die. But if you have a satisficer, then as long as some of the qualifying plans don’t kill everyone, there’s a reasonable chance that the AI will pick one of those plans. Even if you forget to explicitly penalize one of the pathways to disaster, there’s no special reason why that one pathway would show up in a large majority of the AI’s candidate plans.
There is a special reason, and it’s called “instrumental convergence.” Satisficers tend to seek power.
I suspect we’re talking about two different things.
If you just naively program a super-intelligent AI to satisfice a goal, then, sure, most of the candidate pathways to satisfice will involve accruing a lot of some type of power, because power is useful for achieving goals. That’s a valid point, and it’s important to understand that merely switching from optimizers to satisficers won’t adequately protect us against overly ambitious AIs.
However, that doesn’t mean that it’s futile to explicitly penalize most (but not literally all) of the paths that the AI could take to accumulate too much power. Suppose you adequately penalize all of the tactics that would have catastrophic side effects except for, oh, I don’t know, cornering the world market on all types of grain, because you forgot to think of that one particular path to power. Would most of the candidate plans that the AI submits for human approval then turn out to involve secretly cornering the grain market?
I don’t see why they would. All else being equal, sure, the most powerful tactic available is going to be somewhat more attractive than other, less powerful tactics. But how much more attractive? Would an AI generate 1,000x more plans that involve one specific globally disruptive tactic like cornering the grain market vs. all of the various tactics that are only locally disruptive, like publishing propaganda or hiring away competing researchers or hacking into competing research labs or interrupting the flow of electricity to those labs? 10x more such plans? 2x more such plans? I don’t think that’s the kind of estimate you can make just by gesturing at basic principles of game theory; you’d need some concrete domain knowledge about the AI’s specific planning algorithms.
If the truly dangerous plans are only 10x more common on the initial brainstorming list, then we can probably make sure those plans aren’t chosen by filtering for plans that the AI rates as safe and transparent. We can use 3 different AIs on airgapped equipment—one that only “cares” about evaluating whether a plan is safe (i.e., would the plan cause unacceptably large changes if implemented?), one that only “cares” about evaluating whether a plan is transparent (i.e., can the plan be explained to humans so that the humans will have an accurate understanding of what actions will be taken as part of the plan?), and one that generates plans that satisfice for a substantive goal (in this case, shutting down opposing AI research).
Of the plans that are approved by all 3 AIs, most of them probably won’t involve accumulating excessive amounts of power, because usually if you accumulate too much power, then that is (a) unsafe, and (b) likely to involve lying to ‘your’ human programmers.
If powerful plans are 10x more common on the initial brainstorming list, then you need filters for safety and/or transparency that are at least 90% effective at catching unacceptably powerful plans in order to get back up to a 50% chance of humanity surviving. If powerful plans are more than 10x more common, then you need a set of filters that’s more reliable than 90%. I don’t think we have any idea what those numbers are yet, but I do think it’s worth trying to reduce how common it is for excessively powerful plans to show up on the initial brainstorming list, and I think we can do that by training AIs to recognize dangerously disruptive plans and to try to avoid those types of plans. It’s better to at least try to get AIs to engage with the concept of “this plan is too disruptive” then to throw up our hands and say, “Oh, power is an attractor in game theory space, so there’s no possible way to get brilliant AIs that don’t seize infinite power.”
You mean quantilization? Oh yeah, I forgot about that. Good point.