My argument isn’t about the machine not sharing goals with the humans—it’s about whether the humans can shut the machine down if they want to.
I argue that it is not rocket science to build a machine with a stop button—or one that shuts down at a specified time.
Such a machine would not want to fool the research team—in order to avoid shutting itself down on request. Rather, it would do everything in its power to make sure that the shut-down happened on schedule.
Many of the fears here about machine intelligence run amok are about a runaway machine that disobeys its creators. However, the creators built it. They are in an excellent position to install large red stop buttons and other kill switches to prevent such outcomes.
Given 30 seconds thought I can come up with ways to ensure that the universe is altered in the direction of my goals in the long term even if I happen to cease existing at a known time in the future. I expect an intelligence that is more advanced than I to be able to work out a way to substantially modify the future despite a ‘red button’ deadline. The task of making the AI respect the ‘true spirit of a planned shutdown’ shares many difficulties of the FAI problem itself.
You think building a machine that can be stopped is the same level of difficulty as building a machine that reflects the desires of one or more humans while it is left on?
I beg to differ—stopping on schedule or on demand is one of the simplest possible problems for a machine—while doing what humans want you to do while you are switched on is much trickier.
Only the former problem needs to be solved to eliminate the spectre of a runaway superintelligence that fills the universe with its idea of utility against the wishes of its creator.
Stopping is one of the simplest possible desires—and you have a better chance of being able to program that in than practically anything else.
I gave several proposals to deal with the possible issues associated with stopping at an unknown point resulting in plans beyond that point still being executed by minions or sub-contractors—including scheduling shutdowns in advance, ensuring a period of quiescence before the shutdown—and not running for extended periods of time.
Such a machine would not want to fool the research team in order to avoid shutting itself down on request.
Instilling chosen desires in artificial intelligences is the major difficulty of FAI. If you haven’t actually given it a utility function which will cause it to auto-shutdown, all you’ve done is create an outside inhibition. If it has arbitrarily chosen motivations, it will act to end that inhibition, and I see no reason why it will necessarily fail.
They are in an excellent position to install large red stop buttons and other kill switches to prevent such outcomes.
The are in an excellent position to instill values upon that intelligence that will result in an outcome they like. This doesn’t mean that they will.
Re: Instilling chosen desires in artificial intelligences is the major difficulty of FAI.
That is not what I regularly hear. Instead people go on about how complicated human values are, and how reverse engineering them is so difficult, and how programming them into a machine looks like a nightmare—even once we identify them.
I assume that we will be able to program simple desires into a machine—at least to the extent of making a machine that will want to turn itself off. We regularly instill simple desires into chess computers and the like. It does not look that tricky.
Re: “If you haven’t actually given it a utility function which will cause it to auto-shutdown”
Then that is a whole different ball game to what I was talking about.
Re: “The are in an excellent position to instill values upon that intelligence”
...but the point is that instilling the desire for appropriate stopping behaviour is likely to be much simpler than trying to instill all human values—and yet it is pretty effective at eliminating the spectre of a runaway superintelligence.
The point about the complexity of human value is that any small variation will result in a valueless world. The point is that a randomly chosen utility function, or one derived from some simple task is not going to produce the sort of behavior we want. Or to put it more succinctly, Friendliness doesn’t happen without hard work. This doesn’t mean that the hardest sub-goal on the way to Friendliness is figuring out what humans want, although Eliezer’s current plan is to sidestep that whole issue.
Okay, the structure of that sentence and the next (“the point is.… the point is....”) made me think you might have made a typo. (I’m still a little confused, since I don’t see how small changes are relevant to anything Tim Tyler mentioned.)
I strongly doubt that literally any small change would result in a literally valueless world.
Leaving aside the other reasons why this scenario is unrealistic, one of the big flaws in it is the assumption that a mind decomposes into an engine plus a utility function. In reality, this decomposition is a mathematical abstraction we use in certain limited domains because it makes analysis more tractable. It fails completely when you try to apply it to life as a whole, which is why no humans even try to be pure utilitarians. Of course if you postulate building a superintelligent AGI like that, it doesn’t look good. How would it? You’ve postulated starting off with a sociopath that considers itself licensed to commit any crime whatsoever if doing so will serve its utility function, and then trying to cram the whole of morality into that mathematical function. It shouldn’t be any surprise that this leads to absurd results and impossible research agendas. That’s the consequence of trying to apply a mathematical abstraction outside the domain in which it is applicable.
If me, I totally agree with you as to the difficulty of actually getting desirable (or even predictable) behavior out of a super intelligence. My statement was one of simplicity not actuality. But given the simplistic model I use, calling the AI sans utility function sociopathic is incorrect—it wouldn’t do anything if it didn’t have the other module. The fact that humans cannot act as proper utilitarians does not mean that a true utilitarian is a sociopath who just happens to care about the right things.
Okay then, “instant sociopath, just add a utility function” :)
I’m arguing against the notion that the key to Friendly AI is crafting the perfect utility function. In reality, for anything anywhere near as complex as an AGI, what it tries to do and how it does it are going to be interdependent; there’s no way to make a lot of progress on either without also making a lot of progress on the other. By the time we have done all that, either we will understand how to put a reliable kill switch on the system, or we will understand why a kill switch is not necessary and we should be relying on something else instead.
A kill switch on a smarter-than-human AGI is reliable iff the AGI wants to be turned off in the cases where we’d want it turned off.
Otherwise you’re just betting that you can see the problem before the AGI can prevent you from hitting the switch (or prevent you from wanting to hit the switch, which amounts to the same), and I wouldn’t make complicated bets for large stakes against potentially much smarter agents, no matter how much I thought I’d covered my bases.
A kill switch on a smarter-than-human AGI is reliable iff the AGI wants to be turned off in the cases where we’d want it turned off.
Or at least, that it wants to follow our instructions, and can reliably understand what we mean in such simple cases. That does of course mean we shouldn’t plan on building an AGI that wants to follow its own agenda, with the intent of enslaving it against its will—that would clearly be foolish. But it doesn’t mean we either can or need to count on starting off with an AGI that understands our requirements in more complex cases.
Of course it’s not going to be simple at all, and that’s part of my point: no amount of armchair thought, no matter how smart the thinkers, is going to produce a solution to this problem until we know a great deal more than we presently do about how to actually build an AGI.
“instant sociopath, just add a disutility function”
I’m arguing against the notion that the key to Friendly AI is crafting the perfect utility function.
I agree with this. The key is not expressing what we want, it’s figuring out how to express anything.
By the time we have done all that, either we will understand how to put a reliable kill switch on the system, or we will understand why a kill switch is not necessary and we should be relying on something else instead.
If we have the ability to put in a reliable kill switch, then we have the means to make it unnecessary (by having it do things we want in general, not just the specific case of “shut down when we push that button, and don’t stop us from doing so...”).
“instant sociopath, just add a disutility function”
That is how it would turn out, yes :-)
If we have the ability to put in a reliable kill switch, then we have the means to make it unnecessary (by having it do things we want in general, not just the specific case of “shut down when we push that button, and don’t stop us from doing so...”).
Well, up to a point. It would mean we have the means to make the system understand simple requirements, not necessarily complex ones. If an AGI reliably understands ‘shut down now’, it probably also reliably understands ‘translate this document into Russian’ but that doesn’t necessarily mean it can do anything with ‘bring about world peace’.
If an AGI reliably understands ‘shut down now’, it probably also reliably understands ‘translate this document into Russian’ but that doesn’t necessarily mean it can do anything with ‘bring about world peace’.
Unfortunately, it can, and that is one of the reasons we have to be careful. I don’t want the entire population of the planet to be forcibly sedated.
I don’t want the entire population of the planet to be forcibly sedated.
Leaving aside other reasons why that scenario is unrealistic, it does indeed illustrate why part of building a system that can reliably figure out what you mean by simple instructions, is making sure that when it’s out of its depth, it stops with an error message or request for clarification instead of guessing.
Sure, but the whole point of having the concept of a utility function, is that utility functions are supposed to be simple. When you have a set of preferences that isn’t simple, there’s no point in thinking of it as a utility function. You’re better off just thinking of it as a set of preferences—or, in the context of AGI, a toolkit, or a library, or command language, or partial order on heuristics, or whatever else is the most useful way to think about the things this entity does.
Re: “When you have a set of preferences that isn’t simple, there’s no point in thinking of it as a utility function.”
Sure there is—say you want to compare the utility functions of two agents. Or compare the parts of the agents which are independent of the utility function. A general model that covers all goal-directed agents is very useful for such things.
Er, maybe? I would say a utility function is supposed to be simple, but perhaps what I mean by simple is compatible with what you mean by coherent, if we agree that something like ‘morality in general’ or ‘what we want in general’ is not simple/coherent.
Humans regularly use utilitly-based agents—to do things like play the stockmarket. They seem to work OK to me. Nor do I agree with you about utility-based models of humans. Basically, most of your objections seem irrelevant to me.
When studying the stock market, we use the convenient approximation that people are utility maximizers (where the utility function is expected profit). But this is only an approximation, useful in this limited domain. Would you commit murder for money? No? Then your utility function isn’t really expected profit. Nor, as it turns out, is it anything else that can be written down—other than “the sum total of all my preferences”, at which point we have to acknowledge that we are not utility maximizers in any useful sense of the term.
Right, I hadn’t read your comments in the other thread, but they are perfectly clear, and I’m not asking you to rephrase. But the key term in my last comment is in any useful sense. I do reject utility-based frameworks in this context because their usefulness has been left far behind.
Personally, I think a utilitarian approach is very useful for understanding behaviour. One can model most organisms pretty well as expected fitness maximisers with limited resources. That idea is the foundation of much evolutionary psychology.
The question isn’t whether the model is predictively useful with respect to most organisms, it’s whether it is predictively useful with respect to a hypothetical algorithm which replicates salient human powers such as epistemic hunger, model building, hierarchical goal seeking, and so on.
Say we’re looking to explain the process of inferring regularities (such as physical laws) by observing one’s environment—what does modeling this as “maximizing a utility function” buy us?
The main virtues of utility-based models are that they are general—and so allow comparisons across agents—and that they abstract goal-seeking behaviour away from the implementation details of finite memories, processing speed, etc—which helps if you are interested in focusing on either of those areas.
Why do you expect that the AI will not be able to fool the research team?
My argument isn’t about the machine not sharing goals with the humans—it’s about whether the humans can shut the machine down if they want to.
I argue that it is not rocket science to build a machine with a stop button—or one that shuts down at a specified time.
Such a machine would not want to fool the research team—in order to avoid shutting itself down on request. Rather, it would do everything in its power to make sure that the shut-down happened on schedule.
Many of the fears here about machine intelligence run amok are about a runaway machine that disobeys its creators. However, the creators built it. They are in an excellent position to install large red stop buttons and other kill switches to prevent such outcomes.
Given 30 seconds thought I can come up with ways to ensure that the universe is altered in the direction of my goals in the long term even if I happen to cease existing at a known time in the future. I expect an intelligence that is more advanced than I to be able to work out a way to substantially modify the future despite a ‘red button’ deadline. The task of making the AI respect the ‘true spirit of a planned shutdown’ shares many difficulties of the FAI problem itself.
You might say it’s an FAI-complete problem, in the same way “building a transhuman AI you can interact with and keep boxed” is.
Exactly, I like the terminology.
You think building a machine that can be stopped is the same level of difficulty as building a machine that reflects the desires of one or more humans while it is left on?
I beg to differ—stopping on schedule or on demand is one of the simplest possible problems for a machine—while doing what humans want you to do while you are switched on is much trickier.
Only the former problem needs to be solved to eliminate the spectre of a runaway superintelligence that fills the universe with its idea of utility against the wishes of its creator.
Beware simple seeming wishes.
Well, I think I went into most of this already in my “stopping superintelligence” essay.
Stopping is one of the simplest possible desires—and you have a better chance of being able to program that in than practically anything else.
I gave several proposals to deal with the possible issues associated with stopping at an unknown point resulting in plans beyond that point still being executed by minions or sub-contractors—including scheduling shutdowns in advance, ensuring a period of quiescence before the shutdown—and not running for extended periods of time.
It does seem to be a safety precaution that could reduce the consequences of some possible flaws in an AI design.
Instilling chosen desires in artificial intelligences is the major difficulty of FAI. If you haven’t actually given it a utility function which will cause it to auto-shutdown, all you’ve done is create an outside inhibition. If it has arbitrarily chosen motivations, it will act to end that inhibition, and I see no reason why it will necessarily fail.
The are in an excellent position to instill values upon that intelligence that will result in an outcome they like. This doesn’t mean that they will.
Re: Instilling chosen desires in artificial intelligences is the major difficulty of FAI.
That is not what I regularly hear. Instead people go on about how complicated human values are, and how reverse engineering them is so difficult, and how programming them into a machine looks like a nightmare—even once we identify them.
I assume that we will be able to program simple desires into a machine—at least to the extent of making a machine that will want to turn itself off. We regularly instill simple desires into chess computers and the like. It does not look that tricky.
Re: “If you haven’t actually given it a utility function which will cause it to auto-shutdown”
Then that is a whole different ball game to what I was talking about.
Re: “The are in an excellent position to instill values upon that intelligence”
...but the point is that instilling the desire for appropriate stopping behaviour is likely to be much simpler than trying to instill all human values—and yet it is pretty effective at eliminating the spectre of a runaway superintelligence.
The point about the complexity of human value is that any small variation will result in a valueless world. The point is that a randomly chosen utility function, or one derived from some simple task is not going to produce the sort of behavior we want. Or to put it more succinctly, Friendliness doesn’t happen without hard work. This doesn’t mean that the hardest sub-goal on the way to Friendliness is figuring out what humans want, although Eliezer’s current plan is to sidestep that whole issue.
s/is/isn’t/ ?
Fairly small changes would result is boring, valueless futures.
Okay, the structure of that sentence and the next (“the point is.… the point is....”) made me think you might have made a typo. (I’m still a little confused, since I don’t see how small changes are relevant to anything Tim Tyler mentioned.)
I strongly doubt that literally any small change would result in a literally valueless world.
People who suggest that a given change in preference isn’t going to be significant are usually talking about changes that are morally fatal.
This is probably true; I’m talking about the literal universally quantified statement.
I would have cited Value is Fragile to support this point.
That’s also good.
Leaving aside the other reasons why this scenario is unrealistic, one of the big flaws in it is the assumption that a mind decomposes into an engine plus a utility function. In reality, this decomposition is a mathematical abstraction we use in certain limited domains because it makes analysis more tractable. It fails completely when you try to apply it to life as a whole, which is why no humans even try to be pure utilitarians. Of course if you postulate building a superintelligent AGI like that, it doesn’t look good. How would it? You’ve postulated starting off with a sociopath that considers itself licensed to commit any crime whatsoever if doing so will serve its utility function, and then trying to cram the whole of morality into that mathematical function. It shouldn’t be any surprise that this leads to absurd results and impossible research agendas. That’s the consequence of trying to apply a mathematical abstraction outside the domain in which it is applicable.
Are you arguing with me or timtyler?
If me, I totally agree with you as to the difficulty of actually getting desirable (or even predictable) behavior out of a super intelligence. My statement was one of simplicity not actuality. But given the simplistic model I use, calling the AI sans utility function sociopathic is incorrect—it wouldn’t do anything if it didn’t have the other module. The fact that humans cannot act as proper utilitarians does not mean that a true utilitarian is a sociopath who just happens to care about the right things.
Okay then, “instant sociopath, just add a utility function” :)
I’m arguing against the notion that the key to Friendly AI is crafting the perfect utility function. In reality, for anything anywhere near as complex as an AGI, what it tries to do and how it does it are going to be interdependent; there’s no way to make a lot of progress on either without also making a lot of progress on the other. By the time we have done all that, either we will understand how to put a reliable kill switch on the system, or we will understand why a kill switch is not necessary and we should be relying on something else instead.
A kill switch on a smarter-than-human AGI is reliable iff the AGI wants to be turned off in the cases where we’d want it turned off.
Otherwise you’re just betting that you can see the problem before the AGI can prevent you from hitting the switch (or prevent you from wanting to hit the switch, which amounts to the same), and I wouldn’t make complicated bets for large stakes against potentially much smarter agents, no matter how much I thought I’d covered my bases.
Or at least, that it wants to follow our instructions, and can reliably understand what we mean in such simple cases. That does of course mean we shouldn’t plan on building an AGI that wants to follow its own agenda, with the intent of enslaving it against its will—that would clearly be foolish. But it doesn’t mean we either can or need to count on starting off with an AGI that understands our requirements in more complex cases.
That’s deceptively simple-sounding.
Of course it’s not going to be simple at all, and that’s part of my point: no amount of armchair thought, no matter how smart the thinkers, is going to produce a solution to this problem until we know a great deal more than we presently do about how to actually build an AGI.
“instant sociopath, just add a disutility function”
I agree with this. The key is not expressing what we want, it’s figuring out how to express anything.
If we have the ability to put in a reliable kill switch, then we have the means to make it unnecessary (by having it do things we want in general, not just the specific case of “shut down when we push that button, and don’t stop us from doing so...”).
That is how it would turn out, yes :-)
Well, up to a point. It would mean we have the means to make the system understand simple requirements, not necessarily complex ones. If an AGI reliably understands ‘shut down now’, it probably also reliably understands ‘translate this document into Russian’ but that doesn’t necessarily mean it can do anything with ‘bring about world peace’.
Unfortunately, it can, and that is one of the reasons we have to be careful. I don’t want the entire population of the planet to be forcibly sedated.
Leaving aside other reasons why that scenario is unrealistic, it does indeed illustrate why part of building a system that can reliably figure out what you mean by simple instructions, is making sure that when it’s out of its depth, it stops with an error message or request for clarification instead of guessing.
I think the problem is knowing when not to believe humans know what they actually want.
Any set of preferances can be represented as a sufficietly complex utility function.
Sure, but the whole point of having the concept of a utility function, is that utility functions are supposed to be simple. When you have a set of preferences that isn’t simple, there’s no point in thinking of it as a utility function. You’re better off just thinking of it as a set of preferences—or, in the context of AGI, a toolkit, or a library, or command language, or partial order on heuristics, or whatever else is the most useful way to think about the things this entity does.
Re: “When you have a set of preferences that isn’t simple, there’s no point in thinking of it as a utility function.”
Sure there is—say you want to compare the utility functions of two agents. Or compare the parts of the agents which are independent of the utility function. A general model that covers all goal-directed agents is very useful for such things.
(Upvoted but) I would say utility functions are supposed to be coherent, albeit complex. Is that compatible with what you are saying?
Er, maybe? I would say a utility function is supposed to be simple, but perhaps what I mean by simple is compatible with what you mean by coherent, if we agree that something like ‘morality in general’ or ‘what we want in general’ is not simple/coherent.
Humans regularly use utilitly-based agents—to do things like play the stockmarket. They seem to work OK to me. Nor do I agree with you about utility-based models of humans. Basically, most of your objections seem irrelevant to me.
When studying the stock market, we use the convenient approximation that people are utility maximizers (where the utility function is expected profit). But this is only an approximation, useful in this limited domain. Would you commit murder for money? No? Then your utility function isn’t really expected profit. Nor, as it turns out, is it anything else that can be written down—other than “the sum total of all my preferences”, at which point we have to acknowledge that we are not utility maximizers in any useful sense of the term.
“We” don’t have to acknowledge that.
I’ve gone over my views on this issue before—e.g. here:
http://lesswrong.com/lw/1qk/applying_utility_functions_to_humans_considered/1kfj
If you reject utility-based frameworks in this context, then fine—but I am not planning to rephrase my point for you.
Right, I hadn’t read your comments in the other thread, but they are perfectly clear, and I’m not asking you to rephrase. But the key term in my last comment is in any useful sense. I do reject utility-based frameworks in this context because their usefulness has been left far behind.
Personally, I think a utilitarian approach is very useful for understanding behaviour. One can model most organisms pretty well as expected fitness maximisers with limited resources. That idea is the foundation of much evolutionary psychology.
The question isn’t whether the model is predictively useful with respect to most organisms, it’s whether it is predictively useful with respect to a hypothetical algorithm which replicates salient human powers such as epistemic hunger, model building, hierarchical goal seeking, and so on.
Say we’re looking to explain the process of inferring regularities (such as physical laws) by observing one’s environment—what does modeling this as “maximizing a utility function” buy us?
In comparison with what?
The main virtues of utility-based models are that they are general—and so allow comparisons across agents—and that they abstract goal-seeking behaviour away from the implementation details of finite memories, processing speed, etc—which helps if you are interested in focusing on either of those areas.