We seem to agree that for an AI to talk itself out of a confinement (like in the AI box experiment), the AI would have to understand what humans mean and want.
As far as I understand your position, you believe that it is difficult to make an AI care to do what humans want, apart from situations where it is temporarily instrumentally useful to do what humans want.
Do you agree that for such an AI to do what humans want, in order to deceive them, humans would have to succeed at either encoding the capability to understand what humans want, or succeed at encoding the capability to make itself capable of understanding what humans want?
My question, do you believe there to be a conceptual difference between encoding capabilities, what an AI can do, and goals, what an AI will do? As far as I understand, capabilities and goals are both encodings of how humans want an AI to behave.
In other words, humans intend an AI to be intelligent and use its intelligence in a certain way. And in order to be an existential risk, humans need to succeed making and AI behave intelligently but fail at making it use its intelligence in a way that does not kill everyone.
Your summaries of my views here are correct, given that we’re talking about a superintelligence.
My question, do you believe there to be a conceptual difference between encoding capabilities, what an AI can do, and goals, what an AI will do? As far as I understand, capabilities and goals are both encodings of how humans want an AI to behave.
Well, there’s obviously a difference; ‘what an AI can do’ and ‘what an AI will do’ mean two different things. I agree with you that this difference isn’t a particularly profound one, and the argument shouldn’t rest on it.
What the argument rests on is, I believe, that it’s easier to put a system into a positive feedback loop that helps it better model its environment and/or itself, than it is to put a system into a positive feedback loop that helps it better pursue a specific set of highly complex goals we have in mind (but don’t know how to fully formalize).
If the AI incorrectly models some feature of itself or its environment, reality will bite back. But if it doesn’t value our well-being, how do we make reality bite back and change the AI’s course? How do we give our morality teeth?
Whatever goals it initially tries to pursue, it will fail in those goals more often the less accurate its models are of its circumstances; so if we have successfully programmed it to do increasingly well at any difficult goal at all (even if it’s not the goal we intended it to be good at), then it doesn’t take a large leap of the imagination to see how it could receive feedback from its environment about how well it’s doing at modeling states of affairs. ‘Modeling states of affairs well’ is not a highly specific goal, it’s instrumental to nearly all goals, and it’s easy to measure how well you’re doing at it if you’re entangled with anything about your environment at all, e.g., your proximity to a reward button.
(And when a system gets very good at modeling itself, its environment, and the interactions between the two, such that it can predict what changes its behaviors are likely to effect and choose its behaviors accordingly, then we call its behavior ‘intelligent’.)
This stands in stark contrast to the difficulty of setting up a positive feedback loop that will allow an AGI to approximate our True Values with increasing fidelity. We understand how accurately modeling something works; we understand the basic principles of intelligence. We don’t understand the basic principles of moral value, and we don’t even have a firm grasp about how to go about finding out the answer to moral questions. Presumably our values are encoded in some way in our brains, such that there is some possible feedback loop we could use to guide an AGI gradually toward Friendliness. But how do we figure out in advance what that feedback loop needs to look like, without asking the superintelligence? (We can’t ask the superintelligence what algorithm to use to make it start becoming Friendly, because to the extent it isn’t already Friendly it isn’t a trustworthy source of information. This is in addition to the seed/intelligence distinction I noted above.)
If we slightly screw up the AGI’s utility function, it will still need to to succeed at modeling things accurately in order to do anything complicated at all. But it will not need to succeed at optimally caring about what humans care about in order to do anything complicated at all.
...put a system into a positive feedback loop that helps it better model its environment and/or itself...
This can be understood as both a capability and as a goal. What humans mean an AI to do is to undergo recursive self-improvement. What humans mean an AI to be capable of is to undergo recursive self-improvement.
I am only trying to clarify the situation here. Please correct me if you think that above is wrong.
If the AI incorrectly models some feature of itself or its environment, reality will bite back. But if it doesn’t value our well-being, how do we make reality bite back and change the AI’s course?
I do not disagree with the orthogonality thesis insofar as an AI can have goals that interfere with human values in a catastrophic way, possibly leading to human extinction.
...if we have successfully programmed it to do increasingly well at any difficult goal at all (even if it’s not the goal we intended it to be good at), then it doesn’t take a large leap of the imagination to see how it could receive feedback from its environment about how well it’s doing at modeling states of affairs.
I believe here is where we start to disagree. I do not understand how the “improvement” part of recursive self-improvement can be independent of properties such as the coherence and specificity of the goal the AI is supposed to achieve.
Either you have a perfectly specified goal, such as “maximizing paperclips”, where it is clear what “maximization” means, and what the properties of “paperclips” are, or there is some amount of uncertainty about what it means to achieve the goal of “maximizing paperclips”.
Consider the programmers forgot to encode what shape the paperclips are supposed to have. How do you suppose would that influence the behavior of the AI. Would it just choose some shape at random, or would it conclude that shape is not part of its goal? If the former, where would the decision to randomly choose a shape come from? If the latter, what would it mean to maximize shapeless objects?
I am just trying to understand what kind of AI you have in mind.
‘Modeling states of affairs well’ is not a highly specific goal, it’s instrumental to nearly all goals,...
This is a clearer point of disagreement.
An AI needs to be able to draw clear lines where exploration ends and exploitation starts. For example, an AI that thinks about every decision for a year would never get anything done.
An AI also needs to discount low probability possibilities, as to not be vulnerable to internal or external Pascal’s mugging scenarios.
These are problems that humans need to solve and encode in order for an AI to be a danger.
But these problems are in essence confinements, or bounds on how an AI is going to behave.
How likely is an AI then going to take over the world, or look for dangerous aliens, in order to make sure that neither aliens nor humans obstruct it from achieving its goal?
Similarly, how likely is such an AI to convert all resources into computronium in order to be better able to model states of affairs well?
This stands in stark contrast to the difficulty of setting up a positive feedback loop that will allow an AGI to approximate our True Values with increasing fidelity.
I understand this. And given your assumptions about how an AI will affect the whole world in a powerful way, it makes sense to make sure that it does so in a way that preserves human values.
I have previously compared this to uncontrollable self-replicating nanobots. Given that you cannot confine the speed or scope of their self-replication, only the nature of the transformation that they cause, you will have to make sure that they transform the world into a paradise rather than grey goo.
or there is some amount of uncertainty about what it means to achieve the goal of “maximizing paperclips
“uncertainty” is in your human understanding of the program, not in the actual program. A program doesn’t go “I don’t know what I’m supposed to do next”, it follows instructions step-by-step.
If the latter, what would it mean to maximize shapeless objects?
It would mean exactly what it’s programmed to mean, without any uncertainty in it at all.
This can be understood as both a capability and as a goal.
Yes. To divide it more finely, it could be a terminal goal, or an instrumental goal; it could be a goal of the AI, or a goal of the human; it could be a goal the human would reflectively endorse, or a goal the human would reflectively reject but is inadvertently promoting anyway.
I believe here is where we start to disagree. I do not understand how the “improvement” part of recursive self-improvement can be independent of properties such as the coherence and specificity of the goal the AI is supposed to achieve.
I agree that, at a given time, the AI must have a determinate goal. (Though the encoding of that goal may be extremely complicated and unintentional. And it may need to be time-indexed.) I’m not dogmatically set on the idea that a self-improving AGI is easy to program; at this point it wouldn’t shock me if it took over 100 years to finish making the thing. What you’re alluding to are the variety of ways we could fail to construct a self-improving AGI at all. Obviously there are plenty of ways to fail to make an AGI that can improve its own ability to track things about its environment in a domain-general way, without bursting into flames at any point. If there weren’t plenty of ways to fail, we’d have already succeeded.
Our main difference in focus is that I’m worried about what happens if we do succeed in building a self-improving AGI that doesn’t randomly melt down. Conditioned on our succeeding in the next few centuries in making a machine that actually optimizes for anything at all, and that optimizes for its own ability to generally represent its environment in a way that helps it in whatever else it’s optimizing for, we should currently expect humans to go extinct as a result. Even if the odds of our succeeding in the next few centuries were small, it would be worth thinking about how to make that extinction event less likely. (Though they aren’t small.)
I gather that you think that making an artificial process behave in any particular way at all (i.e., optimizing for something), while recursively doing surgery on its own source code in the radical way MIRI is interested in, is very tough. My concern is that, no matter how true that is, it doesn’t entail that if we succeed at that tough task, we’ll therefore have made much progress on other important tough tasks, like Friendliness. It does give us more time to work on Friendliness, but if we convince ourselves that intelligence explosion is a completely pie-in-the-sky possibility, then we won’t use that time effectively.
I also gather that you have a hard time imagining our screwing up on a goal architecture without simply breaking the AGI. Perhaps by ‘screwing up’ you’re imagining failing to close a set of parentheses. But I think you should be at least as worried about philosophical, as opposed to technical, errors. A huge worry isn’t just that we’ll fail to make the AI we intended; it’s that our intentions while we’re coding the thing will fail to align with the long-term interests of ourselves, much less of the human race.
But these problems are in essence confinements, or bounds on how an AI is going to behave.
How likely is an AI then going to take over the world, or look for dangerous aliens, in order to make sure that neither aliens nor humans obstruct it from achieving its goal?
We agree that it’s possible to ‘bind’ a superintelligence. (By this you don’t mean boxing it; you just mean programming it to behave in some ways as opposed to others.) But if the bindings fall short of Friendliness, while enabling superintelligence to arise at all, then a serious risk remains. Is your thought that Friendliness is probably an easier ‘binding’ to figure out how to code than are, say, resisting Pascal’s mugging, or having consistent arithmetical reasoning?
Our main difference in focus is that I’m worried about what happens if we do succeed in building a self-improving AGI that doesn’t randomly melt down.
I am trying to understand if the kind of AI, that is underlying the scenario that you have in mind, is a possible and likely outcome of human AI research.
As far as I am aware, as a layman, goals and capabilities are intrinsically tied together. How could a chess computer be capable of winning against humans at chess without the terminal goal of achieving a checkmate?
Coherent and specific goals are necessary to (1) decide which actions are instrumental useful (2) judge the success of self-improvement. If the given goal is logically incoherent, or too vague for the AI to be able to tell apart success from failure, would it work at all?
If I understand your position correctly, you would expect a chess playing general AI, one that does not know about checkmate, instead of “winning at chess”, to improve against such goals as “modeling states of affairs well” or “make sure nothing intervenes chess playing”. You believe that these goals do not have to be programmed by humans, because they are emergent goals, an instrumental consequence of being general intelligent.
These universal instrumental goals, these “AI drives”, seem to be a major reason for why you believe it to be important to make the AI care about behaving correctly. You believe that these AI drives are a given, and the only way to prevent an AI from being an existential risk is to channel these drives, is to focus this power on protecting and amplifying human values.
My perception is that these drives that you imagine are not special and will be as difficult to get “right” than any other goal. I think that the idea that humans not only want to make an AI exhibit such drives, but also succeed at making such drives emerge, is a very unlikely outcome.
As far as I am aware, here is what you believe an AI to want:
It will want to self-improve
It will want to be rational
It will try to preserve their utility functions
It will try to prevent counterfeit utility
It will be self-protective It will want to acquire resources and use them efficiently
What AIs that humans would ever want to create would require all of these drives, and how easy will it be for humans to make an AI exhibit these drives compared to making an AI that can do what humans want without these drives?
Take mathematics. What are the difficulties associated with making an AI better than humans at mathematics, and will an AI need these drives in order to do so?
Humans did not evolve to play chess or do mathematics. Yet it is considerably more difficult to design a chess AI than an AI that is capable of discovering interesting and useful mathematics.
I believe that the difficulty is due to the fact that it is much easier to formalize what it means to play chess than doing mathematics. The difference between chess and mathematics is that chess has a specific terminal goal in the form of a clear definition of what constitutes winning. Although mathematics has unambiguous rules, there is no specific terminal goal and no clear definition of what constitutes winning.
The progress of the capability of artificial intelligence is not only related to whether humans have evolved for a certain skill or to how much computational resources it requires but also to how difficult it is to formalize the skill, its rules and what it means to succeed.
In the light of this, how difficult would it be to program the drives that you imagine, versus just making an AI win against humans at a given activity without exhibiting these drives?
All these drives are very vague ideas, not like “winning at chess”, but more like “being better at mathematics than Terence Tao”.
The point I am trying to make is that these drives constitute additional complexity, rather than being simple ideas that you can just assume, and from which you can reason about the behavior of an AI.
It is this context that the “dumb superintelligence” argument tries to highlight. It is likely incredibly hard to make these drives emerge in a seed AI. They implicitly presuppose that humans succeed at encoding intricate ideas about what “winning” means in all those cases required to overpower humans, but not in the case of e.g. winning at chess or doing mathematics. I like to analogize such a scenario to the creation of a generally intelligent autonomous car that works perfectly well at not destroying itself in a crash but which somehow manages to maximize the number of people to run over.
I agree that if you believe that it is much easier to create a seed AI to exhibit the drives that you imagine, than it is to make a seed AI use its initial resources to figure out how to solve a specific problem, then we agree about AI risks.
How could a chess computer be capable of winning against humans at chess without the terminal goal of achieving a checkmate?
Humans are capable of winning at chess without the terminal goal of doing so. Nor were humans designed by evolution specifically for chess. Why should we expect a general superintelligence to have intelligence that generalizes less easily than a human’s does?
If the given goal is logically incoherent, or too vague for the AI to be able to tell apart success from failure, would it work at all?
You keep coming back to this ‘logically incoherent goals’ and ‘vague goals’ idea. Honestly, I don’t have the slightest idea what you mean by those things. A goal that can’t motivate one to do anything ain’t a goal; it’s decor, it’s noise. ‘Goals’ are just the outcomes systems tend to produce, especially systems too complex to be easily modeled as, say, physical or chemical processes. Certainly it’s possible for goals to be incredibly complicated, or to vary over time. But there’s no such thing as a ‘logically incoherent outcome’. So what’s relevant to our purposes is whether failing to make a powerful optimization process human-friendly will also consistently stop the process from optimizing for anything whatsoever.
I think that the idea that humans not only want to make an AI exhibit such drives, but also succeed at making such drives emerge, is a very unlikely outcome.
Conditioned on a self-modifying AGI (say, an AGI that can quine its source code, edit it, then run the edited program and repeat the process) achieving domain-general situation-manipulating abilities (i.e., intelligence), analogous to humans’ but to a far greater degree, which of the AI drives do you think are likely to be present, and which absent? ‘It wants to self-improve’ is taken as a given, because that’s the hypothetical we’re trying to assess. Now, should we expect such a machine to be indifferent to its own survival and to the use of environmental resources?
The point I am trying to make is that these drives constitute additional complexity, rather than being simple ideas that you can just assume
Sometimes a more complex phenomenon is the implication of a simpler hypothesis. A much narrower set of goals will have intelligence-but-not-resource-acquisition as instrumental than will have both as instrumental, because it’s unlikely to hit upon a goal that requires large reasoning abilities but does not call for many material resources.
It is likely incredibly hard to make these drives emerge in a seed AI.
You haven’t given arguments suggesting that here. At most, you’ve given arguments against expecting a seed AI to be easy to invent. Be careful to note, to yourself and others, when you switch between the claims ‘a superintelligence is too hard to make’ and ‘if we made a superintelligence it would probably be safe’.
You keep coming back to this ‘logically incoherent goals’ and ‘vague goals’ idea. Honestly, I don’t have the slightest idea what you mean by those things.
Well, I’m not sure what XXD means by them, but…
G1 (“Everything is painted red”) seems like a perfectly coherent goal. A system optimizing G1 paints things red, hires people to paint things red, makes money to hire people to paint things red, invents superior paint-distribution technologies to deposit a layer of red paint over things, etc.
G2 (“Everything is painted blue”) similarly seems like a coherent goal.
G3 (G1 AND G2) seems like an incoherent goal. A system with that goal… well, I’m not really sure what it does.
A system’s goals have to be some event that can be brought about. In our world, ‘2+2=4’ and ‘2+2=5’ are not goals; ‘everything is painted red and not-red’ may not be a goal for similar reasons. When we’re talking about an artificial intelligence’s preferences, we’re talking about the things it tends to optimize for, not the things it ‘has in mind’ or the things it believes are its preferences.
This is part of what makes the terminology misleading, and is also why we don’t ask ‘can a superintelligence be irrational?‘. Irrationality is dissonance between my experienced-‘goals’ (and/or, perhaps, reflective-second-order-‘goals’) and my what-events-I-produce-‘goals’; but we don’t care about the superintelligence’s phenomenology. We only care about what events it tends to produce.
Tabooing ‘goal’ and just talking about the events a process-that-models-its-environment-and-directs-the-future tends to produce would, I think, undermine a lot of XiXiDu’s intuitions about goals being complex explicit objects you have to painstakingly code in. The only thing that makes it more useful to model a superintelligence as having ‘goals’ than modeling a blue-minimizing robot as having ‘goals’ is that the superintelligence responds to environmental variation in a vastly more complicated way. (Because, in order to be even a mediocre programmer, its model-of-the-world-that-determines-action has to be more complicated than a simple camcorder feed.)
we’re talking about the things it tends to optimize for, not the things it ‘has in mind’
Oh. Well, in that case, all right. If there exists some X a system S is in fact optimizing for, and what we mean by “S’s goals” is X, regardless of what target S “has in mind”, then sure, I agree that systems never have vague or logically incoherent goals.
just talking about the events a process-that-models-its-environment-and-directs-the-future tends to produce
Well, wait. Where did “models its environment” come from? If we’re talking about the things S optimizes its environment for, not the things S “has in mind”, then it would seem that whether S models its environment or not is entirely irrelevant to the conversation.
In fact, given how you’ve defined “goal” here, I’m not sure why we’re talking about intelligence at all. If that is what we mean by “goal” then intelligence has nothing to do with goals, or optimizing for goals. Volcanoes have goals, in that sense. Protons have goals.
“Since I am so uncertain of Kasparov’s moves, what is the empirical content of my belief that ‘Kasparov is a highly intelligent chess player’? What real-world experience does my belief tell me to anticipate? [...]
“The empirical content of my belief is the testable, falsifiable prediction that the final chess position will occupy the class of chess positions that are wins for Kasparov, rather than drawn games or wins for Mr. G. [...] The degree to which I think Kasparov is a ‘better player’ is reflected in the amount of probability mass I concentrate into the ‘Kasparov wins’ class of outcomes, versus the ‘drawn game’ and ‘Mr. G wins’ class of outcomes.”
“When I think you’re a powerful intelligence, and I think I know something about your preferences, then I’ll predict that you’ll steer reality into regions that are higher in your preference ordering. [...]
“Ah, but how do you know a mind’s preference ordering? Suppose you flip a coin 30 times and it comes up with some random-looking string—how do you know this wasn’t because a mind wanted it to produce that string?
“This, in turn, is reminiscent of the Minimum Message Length formulation of Occam’s Razor: if you send me a message telling me what a mind wants and how powerful it is, then this should enable you to compress your description of future events and observations, so that the total message is shorter. Otherwise there is no predictive benefit to viewing a system as an optimization process. This criterion tells us when to take the intentional stance.
“(3) Actually, you need to fit another criterion to take the intentional stance—there can’t be a better description that averts the need to talk about optimization. This is an epistemic criterion more than a physical one—a sufficiently powerful mind might have no need to take the intentional stance toward a human, because it could just model the regularity of our brains like moving parts in a machine.
“(4) If you have a coin that always comes up heads, there’s no need to say “The coin always wants to come up heads” because you can just say “the coin always comes up heads”. Optimization will beat alternative mechanical explanations when our ability to perturb a system defeats our ability to predict its interim steps in detail, but not our ability to predict a narrow final outcome. (Again, note that this is an epistemic criterion.)
“(5) Suppose you believe a mind exists, but you don’t know its preferences? Then you use some of your evidence to infer the mind’s preference ordering, and then use the inferred preferences to infer the mind’s power, then use those two beliefs to testably predict future outcomes. The total gain in predictive accuracy should exceed the complexity-cost of supposing that ‘there’s a mind of unknown preferences around’, the initial hypothesis.”
Notice that throughout this discussion, what matters is the mind’s effect on its environment, not any internal experience of the mind. Unconscious preferences are just as relevant to this method as are conscious preferences, and both are examples of the intentional stance. Note also that you can’t really measure the rationality of a system you’re modeling in this way; any evidence you raise for ‘irrationality’ could just as easily be used as evidence that the system has more complicated preferences than you initially thought, or that they’re encoded in a more distributed way than you had previously hypothesized.
My take-away from this is that there are two ways we generally think about minds on LessWrong: Rational Choice Theory, on which all minds are equally rational and strange or irregular behaviors are seen as evidence of strange preferences; and what we might call the Ideal Self Theory, on which minds’ revealed preferences can differ from their ‘true self’ preferences, resulting in irrationality. One way of unpacking my idealized values is that they’re the rational-choice-theory preferences I would exhibit if my conscious desires exhibited perfect control over my consciously controllable behavior, and those desires were the desires my ideal self would reflectively prefer, where my ideal self is the best trade-off between preserving my current psychology and enhancing that psychology’s understanding of itself and its environment.
We care about ideal selves when we think about humans, because we value our conscious, ‘felt’ desires (especially when they are stable under reflection) more than our unconscious dispositions. So we want to bring our actual behavior (and thus our rational-choice-theory preferences, the ‘preferences’ we talk about when we speak of an AI) more in line with our phenomenological longings and their idealized enhancements. But since we don’t care about making non-person AIs more self-actualized, but just care about how they tend to guide their environment, we generally just assume that they’re rational. Thus if an AI behaves in a crazy way (e.g., alternating between destroying and creating paperclips depending on what day of the week it is), it’s not because it’s a sane rational ghost trapped by crazy constraints. It’s because the AI has crazy core preferences.
Where did “models its environment” come from?
If we’re talking about the things S optimizes its environment for, not the things S “has in mind”, then it would seem that whether S models its environment or not is entirely irrelevant to the conversation.
Yes, in principle. But in practice, a system that doesn’t have internal states that track the world around it in a reliable and useable way won’t be able to optimize very well for anything particularly unlikely across a diverse set of environments. In other words, it won’t be very intelligent. To clarify, this is an empirical claim I’m making about what it takes to be particularly intelligent in our universe; it’s not part of the definition for ‘intelligent’.
a system that doesn’t have internal states that track the world around it in a reliable and useable way won’t be able to optimize very well for anything particularly unlikely across a diverse set of environments
Yes, that seems plausible.
I would say rather that modeling one’s environment is an effective tool for consistently optimizing for some specific unlikely thing X across a range of environments, so optimizers that do so will be more successful at optimizing for X, all else being equal, but it more or less amounts to the same thing.
But… so what?
I mean, it also seems plausible that optimizers that explicitly represent X as a goal will be more successful at consistently optimizing for X, all else being equal… but that doesn’t stop you from asserting that explicit representation of X is irrelevant to whether a system has X as its goal.
So why isn’t modeling the environment equally irrelevant? Both features, on your account, are optional enhancements an optimizer might or might not display.
It keeps seeming like all the stuff you quote and say before your last two paragraphs ought to provide an answer that question, but after reading it several times I can’t see what answer it might be providing. Perhaps your argument is just going over my head, in which case I apologize for wasting your time by getting into a conversation I’m not equipped for..
Maybe it will help to keep in mind that this is one small branch of my conversation with Alexander Kruel. Alexander’s two main objections to funding Friendly Artificial Intelligence research are that (1) advanced intelligence is very complicated and difficult to make, and (2) getting a thing to pursue a determinate goal at all is extraordinarily difficult. So a superintelligence will never be invented, or at least not for the foreseeable future; so we shouldn’t think about SI-related existential risks. (This is my steel-manning of his view. The way he actually argues seems to instead be predicated on inventing SI being tied to perfecting Friendliness Theory, but I haven’t heard a consistent argument for why that should be so.)
Both of these views, I believe, are predicated on a misunderstanding of how simple and disjunctive ‘intelligence’ and ‘goal’ are, for present purposes. So I’ve mainly been working on tabooing and demystifying those concepts. Intelligence is simply a disposition to efficiently convert a wide variety of circumstances into some set of specific complex events. Goals are simply the circumstances that occur more often when a given intelligence is around. These are both very general and disjunctive ideas, in stark contrast to Friendliness; so it will be difficult to argue that a superintelligence simply can’t be made, and difficult too to argue that optimizing for intelligence requires one to have a good grasp on Friendliness Theory.
Because I’m trying to taboo the idea of superintelligence, and explain what it is about seed AI that will allow it to start recursively improving its own intelligence, I’ve been talking a lot about the important role modeling plays in high-level intelligent processes. Recognizing what a simple idea modeling is, and how far it gets one toward superintelligence once one has domain-general modeling proficiency, helps a great deal with greasing the intuition pump ‘Explosive AGI is a simple, disjunctive event, a low-hanging fruit, relative to Friendliness.’ Demystifying unpacking makes things seem less improbable and convoluted.
I mean, it also seems plausible that optimizers that explicitly represent X as a goal will be more successful at consistently optimizing for X, all else being equal… but that doesn’t stop you from asserting that explicit representation of X is irrelevant to whether a system has X as its goal.
I think this is a map/territory confusion. I’m not denying that superintelligences will have a map of their own preferences; at a bare minimum, they need to know what they want in order to prevent themselves from accidentally changing their own preferences. But this map won’t be the AI’s preferences—those may be a very complicated causal process bound up with, say, certain environmental factors surrounding the AI, or oscillating with time, or who-knows-what.
There may not be a sharp line between the ‘preference’ part of the AI and the ‘non-preference’ part. Since any superintelligence will be exemplary at reasoning with uncertainty and fuzzy categories, I don’t think that will be a serious obstacle.
Does that help explain why I’m coming from? If not, maybe I’m missing the thread unifying your comments.
I suppose it helps, if only in that it establishes that much of what you’re saying to me is actually being addressed indirectly to somebody else, so it ought not surprise me that I can’t quite connect much of it to anything I’ve said. Thanks for clarifying your intent.
For my own part, I’m certainly not functioning here as Alex’s proxy; while I don’t consider explosive intelligence growth as much of a foregone conclusion as many folks here do, I also don’t consider Alex’s passionate rejection of the possibility justified, and have had extended discussions on related subjects with him myself in past years. So most of what you write in response to Alex’s positions is largely talking right past me.
(Which is not to say that you ought not be doing it. If this is in effect a private argument between you and Alex that I’ve stuck my nose into, let me know and I’ll apologize and leave y’all to it in peace.)
Anyway, I certainly agree that a system might have a representation of its goals that is distinct from the mechanisms that cause it to pursue those goals. I have one of those, myself. (Indeed, several.) But if a system is capable of affecting its pursuit of its goals (for example, if it is capable of correcting the effects of a state-change that would, uncorrected, have led to value drift), it is not merely interacting with maps. It is also interacting with the territory… that is, it is modifying the mechanisms that cause it to pursue those goals… in order to bring that territory into line with its pre-existing map.
And in order to do that, it must have such a mechanism, and that mechanism must be consistently isomorphic to its representations of its goals.
Right. I’m not saying that there aren’t things about the AI that make it behave the way it does; what the AI optimizes for is a deterministic result of its properties plus environment. I’m just saying that something about the environment might be necessary for it to have the sorts of preferences we can most usefully model it as having; and/or there may be multiple equally good candidates for the parts of the AI that are its values, or their encoding. If we reify preferences in an uncautious way, we’ll start thinking of the AI’s ‘desires’ too much as its first-person-experienced urges, as opposed to just thinking of them as the effect the local system we’re talking about tends to have on the global system.
So, all right. Cconsider two systems, S1 and S2, both of which happen to be constructed in such a way that right now, they are maximizing the number of things in their environment that appear blue to human observers, by going around painting everything blue.
Suppose we add to the global system a button that alters all human brains so that everything appears blue to us, and we find that S1 presses the button and stops painting, and S2 ignores the button and goes on painting.
Suppose that similarly, across a wide range of global system changes, we find that S1 consistently chooses the action that maximizes the number of things in its environment that appear blue to human observers, while S2 consistently goes on painting.
I agree with you that if I reify S2′s preferences in an uncautious way, I might start thinkng of S2 as “wanting to paint things blue” or “wanting everything to be blue” or “enjoying painting things blue” or as having various other similar internal states that might simply not exist, and that I do better to say it has a particular effect on the global system. S2 simply paints things blue; whether it has the goal of painting things blue or not, I have no idea.
I am far less comfortable saying that S1 has no goals, precisely because of how flexibly and consistently it is revising its actions so as to consistently create a state-change across wide ranges of environments. To use Dennett’s terminology, I am more willing to adopt an intentional stance with respect to S1 than S2.
If I’ve understood your position correctly, you’re saying that I’m unjustified in making that distinction… that to the extent that we can say that S1 and S2 have “goals,” the word “goals” simply refer to the state changes they create in the world. Initially they both have the goal of painting things blue, but S1′s goals keep changing: first it paints things blue, then it presses a button, then it does other things. And, sure, I can make up some story like “S1 maximizes the number of things in its environment that appear blue to human observers, while S2 just paints stuff blue” and that story might even have predictive power, but I ought not fall into the trap of reifying some actual thing that corresponds to those notional “goals”.
I think you’re switching back and forth between a Rational Choice Theory ‘preference’ and an Ideal Self Theory ‘preference’. To disambiguate, I’ll call the former R-preferences and the latter I-preferences. My R-preferences—the preferences you’d infer I had from my behaviors if you treated me as a rational agent—are extremely convoluted, indeed they need to be strongly time-indexed to maintain consistency. My I-preferences are the things I experience a desire for, whether or not that desire impacts my behavior. (Or they’re the things I would, with sufficient reflective insight and understanding into my situation, experience a desire for.)
We have no direct evidence from your story addressing whether S1 or S2 have I-preferences at all. Are they sentient? Do they create models of their own cognitive states? Perhaps we have a little more evidence that S1 has I-preferences than that S2 does, but only by assuming that a system whose goals require more intelligence or theory-of-mind will have a phenomenology more similar to a human’s. I wouldn’t be surprised if that assumption turns out to break down in some important ways, as we explore more of mind-space.
But my main point was that it doesn’t much matter what S1 or S2′s I-preferences are, if all we’re concerned about is what effect they’ll have on their environment. Then we should think about their R-preferences, and bracket exactly what psychological mechanism is resulting in their behavior, and how that psychological mechanism relates to itself.
I’ve said that R-preferences are theoretical constructs that happen to be useful a lot of the time for modeling complex behavior; I’m not sure whether I-preferences are closer to nature’s joints.
Initially they both have the goal of painting things blue, but S1′s goals keep changing: first it paints things blue, then it presses a button, then it does other things.
S1′s instrumental goals may keep changing, because its circumstances are changing. But I don’t think its terminal goals are changing. The only reason to model it as having two completely incommensurate goal sets at different times would be if there were no simple terminal goal that could explain the change in instrumental behavior.
I don’t think I’m switching back and forth between I-preferences and R-preferences.
I don’t think I’m talking about I-preferences at all, nor that I ever have been.
I completely agree with you that they don’t matter for our purposes here, so if I am talking about them, I am very very confused. (Which is certainly possible.)
But I don’t think that R-preferences (preferences, goals, etc.) can sensibly be equated with the actual effects a local system has on a global system. If they could, we could talk equally sensibly about earthquakes having R-preferences (preferences, goals, etc.), and I don’t think it’s sensible to talk that way.
R-preferences (preferences, goals, etc.) are, rather, internal states of a system S.
If S is a competent optimizer (or “rational agent,” if you prefer) with R-preferences (preferences, goals, etc.) P, the existence of P will cause S to behave in ways that cause isomorphic effects (E) on a global system, so we can use observations of E as evidence of P (positing that S is a competent optimizer) or as evidence that S is a competent optimizer (positing the existence of P) or a little of both.
But however we slice it, P is not the same thing as E, E is merely evidence of P’s existence. We can infer P’s existence in other ways as well, even if we never observe E… indeed, even if E never gets produced. And the presence or absence of a given P in S is something we can be mistaken about; there’s a fact of the matter.
I think you disagree with the above paragraph, because you describe R-preferences (preferences, goals, etc.) as theoretical constructs rather than parts of the system, which suggests that there is no fact of the matter… a different theoretical approach might never include P, and it would not be mistaken, it would just be a different theoretical approach.
I also think that because way back at the beginning of this exchange when I suggested “paint everything red AND paint everything blue” was an example of an incoherent goal (R-preference, preference, P), your reply was that it wasn’t a goal at all, since that state can’t actually exist in the world. Which suggests that you don’t see goals as internal states of optimizers and that you do equate P with E.
This is what I’ve been disputing from the beginning.
But to be honest, I’m not sure whether you disagree or not, as I’m not sure we have yet succeeded in actually engaging with one another’s ideas in this exchange.
But I don’t think that R-preferences (preferences, goals, etc.) can sensibly be equated with the actual effects a local system has on a global system. If they could, we could talk equally sensibly about earthquakes having R-preferences (preferences, goals, etc.), and I don’t think it’s sensible to talk that way.
You can treat earthquakes and thunderstorms and even individual particles as having ‘preferences’. It’s just not very useful to do so, because we can give an equally simple explanation for what effects things like earthquakes tend to have that is more transparent about the physical mechanism at work. The intentional strategy is a heuristic for black-boxing physical processes that are too complicated to usefully describe in their physical dynamics, but that can be discussed in terms of the complicated outcomes they tend to promote.
(I’d frame it: We’re exploiting the fact that humans are intuitively dualistic by taking the non-physical modeling device of humans (theory of mind, etc.) and appropriating this mental language and concept-web for all sorts of systems whose nuts and bolts we want to bracket. Slightly regimented mental concepts and terms are useful, not because they apply to all the systems we’re talking about in the same way they were originally applied to humans, but because they’re vague in ways that map onto the things we’re uncertain about or indifferent to.)
‘X wants to do Y’ means that the specific features of X tend to result in Y when its causal influence is relatively large and direct. But, for clarity’s sake, we adopt the convention of only dropping into want-speak when a system is too complicated for us to easily grasp in mechanistic terms why it’s having these complex effects, yet when we can predict that, whatever the mechanism happens to be, it is the sort of mechanism that has those particular complex effects.
Thus we speak of evolution as an optimization process, as though it had a ‘preference ordering’ in the intuitively human (i.e., I-preference) sense, even though in the phenomenological sense it’s just as mindless as an earthquake. We do this because black-boxing the physical mechanisms and just focusing on the likely outcomes is often predictively useful here, and because the outcomes are complicated and specific. This is useful for AIs because we care about the AI’s consequences and not its subjectivity (hence we focused on R-preference), and because AIs are optimization processes of even greater complex specificity in mechanism and outcome than evolution (hence we adopted the intentional stance of ‘preference’-talk in the first place).
R-preferences (preferences, goals, etc.) are, rather, internal states of a system S.
I agree this is often the case, because when we define ‘what is this system capable of?’ we often hold the system fixed while examining possible worlds where the environment varies in all kinds of ways. But if the possible worlds we care about all have a certain environmental feature in common—say, because we know in reality that the environmental condition obtains, and we’re trying to figure out all the ways the AI might in fact behave given different values for the variables we don’t know about with confidence—then we may, in effect, include something about the environment ‘in the AI’ for the purposes of assessing its optimization power and/or preference ordering.
For instance, we might model the AI as having the preference ‘surround the Sun with a dyson sphere’ rather than ‘conditioned on there being a Sun, surround it with a dyson sphere’; if we do the former, then the fact that that is the system’s preference depends in part on the actual existence of the Sun. Does that mean the Sun is a part of the AI’s preference encoding? Is the Sun a component of the AI? I don’t think these questions are important or interesting, so I don’t want us to be too committed to reifying AI preferences. They’re just a useful shorthand for the expected outcomes of the AI’s distinguishing features having a more large and direct causal impact on things.
‘X wants to do Y’ means that the specific features of X tend to result in Y when its causal influence is relatively large and direct. But, for clarity’s sake, we adopt the convention of only dropping into want-speak when a system is too complicated for us to easily grasp in mechanistic terms why it’s having these complex effects
Yes, agreed, for some fuzzy notion of “easily grasp” and “too complicated.” That is, there’s a sense in which thunderstorms are too complicated for me to describe in mechanistic terms why they’re having the effects they have… I certainly can’t predict those effects. But there’s also a sense in which I can describe (and even predict) the effects of a thunderstorm that feels simple, whereas I can’t do the same thing for a human being without invoking “want-speak”/intentional stance.
I’m not sure any of this is [i]justified[/i], but I agree that it is what we do… this is how we speak, and we draw these distinctions. So far, so good.
if the possible worlds we care about all have a certain environmental feature in common [..] we may, in effect, include something about the environment ‘in the AI’
I’m not really sure what you mean by “in the AI” here, but I guess I agree that the boundary between an agent and its environment is always a fuzzy one. So, OK, I suppose we can include things about the environment “in the AI” if we choose. (I can similarly choose to include things about the environment “in myself.”) So far, so good.
we might model the AI as having the preference ‘surround the Sun with a dyson sphere’ rather than ‘conditioned on there being a Sun, surround it with a dyson sphere’; if we do the former, then the fact that that is the system’s preference depends in part on the actual existence of the Sun.
Here is where you lose me again… once again you talk as though there’s simply no fact of the matter as to which preference the AI has, merely our choice as to how we model it.
But it seems to me that there are observations I can make which would provide evidence one way or the other. For example, if it has the preference ‘surround the Sun with a dyson sphere,’ then in an environment lacking the Sun I would expect it to first seek to create the Sun… how else can it implement its preferences? Whereas if it has the preference ‘conditioned on there being a Sun, surround it with a dyson sphere’; in an environment lacking the Sun I would not expect it to create the Sun.
So does the AI seek create the Sun in such an environment, or not? Surely that doesn’t depend on how I choose to model it. The AI’s preference is whatever it is, and controls its behavior. Of course, as you say, if the real world always includes a sun, then I might not be able to tell which preference the AI has. (Then again I might… the test I describe above isn’t the only test I can perform, just the first one I thought of, and other tests might not depend on the Sun’s absence.)
But whether I can tell or not doesn’t affect whether the AI has the preference or not.
if we do the former, then the fact that that is the system’s preference depends in part on the actual existence of the Sun
Again, no. Regardless of how we model it, the system’s preference is what it is, and we can study the system (e.g., see whether it creates the Sun) to develop more accurate models of its preferences.
Does that mean the Sun is a part of the AI’s preference encoding? Is the Sun a component of the AI? I don’t think these questions are important or interesting
I agree. But I do think the question of what the AI (or, more generally, an optimizing agent) will do in various situations is interesting, and it seems to be that you’re consistently eliding over that question in ways I find puzzling.
A system’s goals have to be some event that can be brought about.
This sounds like a potentially confusing level of simplification; a goal should be regarded as at least a way of comparing possible events.
When we’re talking about an artificial intelligence’s preferences, we’re talking about the things it tends to optimize for, not the things it ‘has in mind’ or the things it believes are its preferences.
Its behavior is what makes its goal important. But in a system designed to follow an explicitly specified goal, it does make sense to talk of its goal apart from its behavior. Even though its behavior will reflect its goal, the goal itself will reflect itself better.
If the goal is implemented as a part of the system, other parts of the system can store some information about the goal, certain summaries or inferences based on it. This information can be thought of as beliefs about the goal. And if the goal is not “logically transparent”, that is its specification is such that making concrete conclusions about what it states in particular cases is computationally expensive, then the system never knows what its goal says explicitly, it only ever has beliefs about particular aspects of the goal.
But in a system designed to follow an explicitly specified goal, it does make sense to talk of its goal apart from its behavior. Even though its behavior will reflect its goal, the goal itself will reflect itself better.
Perhaps, but I suspect that for most possible AIs there won’t always be a fact of the matter about where its preference is encoded. The blue-minimizing robot is a good example. If we treat it as a perfectly rational agent, then we might say that it has temporally stable preferences that are very complicated and conditional; or we might say that its preferences change at various times, and are partly encoded, for instance, in the properties of the color-inverting lens on its camera. An AGI’s response to environmental fluctuation will probably be vastly more complicated than a blue-minimizer’s, but the same sorts of problems arise in modeling it.
I think it’s more useful to think of rational-choice-theory-style preferences as useful theoretical constructs—like a system’s center of gravity, or its coherently extrapolated volition—than as real objects in the machine’s hardware or software. This sidesteps the problem of haggling over which exact preferences a system has, how those preferences are distributed over the environment, how to decide between causally redundant encodings which is ‘really’ the preference encoding, etc. See my response to Dave.
“Goal” is a natural idea for describing AIs with limited resources: these AIs won’t be able to make optimal decisions, and their decisions can’t be easily summarized in terms of some goal, but unlike the blue-minimizing robot they have a fixed preference ordering that doesn’t gradually drift away from what it was originally, and eventually they tend to get better at following it.
For example, if a goal is encrypted, and it takes a huge amount of computation to decrypt it, system’s behavior prior to that point won’t depend on the goal, but it’s going to work on decrypting it and eventually will follow it. This encrypted goal is probably more predictive of long-term consequences than anything else in the details of the original design, but it also doesn’t predict its behavior during the first stage (and if there is only a small probability that all resources in the universe will allow decrypting the goal, it’s probable that system’s behavior will never depend on the goal). Similarly, even if there is no explicit goal, as in the case of humans, it might be possible to work with an idealized goal that, like the encrypted goal, can’t be easily evaluated, and so won’t influence behavior for a long time.
My point is that there are natural examples where goals and the character of behavior don’t resemble each other, so that each can’t be easily inferred from the other, while both can be observed as aspects of the system. It’s useful to distinguish these ideas.
I agree preferences aren’t reducible to actual behavior. But I think they are reducible to dispositions to behave, i.e., behavior across counterfactual worlds. If a system prefers a specific event Z, that means that, across counterfactual environments you could have put it in, the future would on average have had more Z the more its specific distinguishing features had a large and direct causal impact on the world.
The examples I used seem to apply to “dispositions” to behave, in the same way (I wasn’t making this distinction). There are settings where the goal can’t be clearly inferred from behavior, or collection of hypothetical behaviors in response to various environments, at least if we keep environments relatively close to what might naturally occur, even as in those settings the goal can be observed “directly” (defined as an idealization based in AI’s design).
An AI with encypted goal (i.e. the AI itself doesn’t know the goal in explicit form, but the goal can be abstractly defined as the result of decryption) won’t behave in accordance with it in any environment that doesn’t magically let it decrypt its goal quickly, there is no tendency to push the events towards what the encrypted goal specifies, until the goal is decrypted (which might be never with high probability).
I don’t think a sufficiently well-encrypted ‘preference’ should be counted as a preference for present purposes. In principle, you can treat any physical chunk of matter as an ‘encrypted preference’, because if the AI just were a key of exactly the right shape, then it could physically interact with the lock in question to acquire a new optimization target. But if neither the AI nor anything very similar to the AI in nearby possible worlds actually acts as a key of the requisite sort, then we should treat the parts of the world that a distant AI could interact with to acquire a preference as, in our world, mere window dressing.
Perhaps if we actually built a bunch of AIs, and one of them was just like the others except where others of its kind had a preference module, it had a copy of The Wind in the Willows, we would speak of this new AI as having an ‘encrypted preference’ consisting of a book, with no easy way to treat that book as a decision criterion like its brother- and sister-AIs do for their homologous components. But I don’t see any reason right now to make our real-world usage of the word ‘preference’ correspond to that possible world’s usage. It’s too many levels of abstraction away from what we should be worried about, which are the actual real-world effects different AI architectures would have.
Evolution was able to come up with cats. Cats are immensely complex objects. Evolution did not intend to create cats. Now consider you wanted to create an expected utility maximizer to accomplish something similar, except that it would be goal-directed, think ahead, and jump fitness gaps. Further suppose that you wanted your AI to create qucks, instead of cats. How would it do this?
Given that your AI is not supposed to search design space at random, but rather look for something particular, you would have to define what exactly qucks are. The problem is that defining what a quck is, is the hardest part. And since nobody has any idea what a quck is, nobody can design a quck creator.
The point is that thinking about the optimization of optimization is misleading, as most of the difficulty is with defining what to optimize, rather than figuring out how to optimize it. In other words, the efficiency of e.g. the scientific method depends critically on being able to formulate a specific hypothesis.
Trying to create an optimization optimizer would be akin to creating an autonomous car to find the shortest route between Gotham City and Atlantis. The problem is not how to get your AI to calculate a route, or optimize how to calculate such a route, but rather that the problem is not well-defined. You have no idea what it means to travel between two fictional cities. Which in turn means that you have no idea what optimization even means in this context, let alone meta-level optimization.
The problem is, you don’t have to program the bit that says “now make yourself more intelligent.” You only have to program the bit that says “here’s how to make a new copy of yourself, and here’s how to prove it shares your goals without running out of math.”
And the bit that says “Try things until something works, then figure out why it worked.” AKA modeling.
The AI isn’t actually an intelligence optimizer. But it notes that when it takes certain actions, it is better able to model the world, which in turn allows it to make more paperclips (or whatever). So it’ll take those actions more often.
Humans are capable of winning at chess without the terminal goal of doing so. Nor were humans designed by evolution specifically for chess. Why should we expect a general superintelligence to have intelligence than generalizes less easily than a human’s does?
Biological evolution is not the full picture here. Humans were programmed to be capable of winning at chess, and to care to do so, by cultural evolution, education, and environmental feedback in the form of incentives given by other people challenging them to play.
I don’t know how this works. But I do not dispute the danger of neuromorphic AIs, as you know from a comment elsewhere.
Do you suggest that from the expected behavior of neuromorphic AIs it is possible to draw conclusions about the behavior of what you call a ‘seed AI’? Would such a seed AI, as would be the case with neuromorphic AIs, be constantly programmed by environmental feedback?
You keep coming back to this ‘logically incoherent goals’ and ‘vague goals’ idea. Honestly, I don’t have the slightest idea what you mean by those things.
What I mean is that if you program a perfect scientist but give this perfect scientist a hypothesis that does not make any predictions, then it will not be able to unfold its power.
Conditioned on a self-modifying AGI...the hypothetical we’re trying to assess.
I believe that I already wrote that I do not dispute that the idea you seem to have in mind is a risk by definition. If such an AI is likely, then we are likely going extinct if we fail at making it care about human values.
You haven’t given arguments suggesting that here.
I feel uncomfortable to say this, but I do not see that the burden of proof is on me to show that it takes deliberate and intentional effort to make an AI exhibit those drives, as long that is not part of your very definition. I find the current argument in favor of AI drives to be thoroughly unconvincing.
Be careful to note, to yourself and others, when you switch between the claims ‘a superintelligence is too hard to make’ and ‘if we made a superintelligence it would probably be safe’.
The former has always been one of the arguments in favor of the latter in the posts I wrote on my blog.
(Note: I’m also a layman, so my non-expert opinions necessarily come with a large salt side-dish)
My guess here is that most of the “AI Drives” to self-improve, be rational, retaining it’s goal structure, etc. are considered necessary for a functional learning/self-improving algorithm. If the program cannot recognize and make rules for new patterns observed in data, make sound inferences based on known information or keep after it’s objective it will not be much of an AGI at all; it will not even be able to function as well as a modern targeted advertising program.
The rest, such as self-preservation, are justified as being logical requirements of the task. Rather than having self-preservation as a terminal value, the paperclip maximizer will value it’s own existence as an optimal means of proliferating paperclips. It makes intuitive sense that those sorts of ‘drives’ would emerge from most-any goal, but then again my intuition is not necessarily very useful for these sorts of questions.
This point might also be a source of confusion;
The progress of the capability of artificial intelligence is not only related to whether humans have evolved for a certain skill or to how much computational resources it requires but also to how difficult it is to formalize the skill, its rules and what it means to succeed.
In the light of this, how difficult would it be to program the drives that you imagine, versus just making an AI win against humans at a given activity without exhibiting these drives?
As Dr Valiant (great name or the greatest name?) classifies things in Probably Approximately Correct, Winning Chess would be a ‘theoryful’ task while Discovering (Interesting) Mathematical Proofs would be a ‘theoryless’ one. In essence, the theoryful has simple and well established rules for the process which could be programmed optimally in advance with little-to-no modification needed afterwards while the theoryless is complex and messy enough that an imperfect (Probably Approximately Correct) learning process would have to be employed to suss out all the rules.
Now obviously the program will benefit from labeling in it’s training data for what is and is not an “interesting” mathematical proof, otherwise it can just screw around with computationally-cheap arithmetic proofs (1 + 1 = 2, 1.1 + 1 = 2.1, 1.2 + 1 = 2.2, etc.) until the heat death of the universe. Less obviously, as the hidden tank example shows, insufficient labeling or bad labels will lead to other unintended results.
So applying that back to Friendliness; despite attempts to construct a Fun Theory, human value is currently (and may well forever remain) theoryless. A learning process whose goal is to maximize human value is going to have to be both well constructed and have very good labels initially to not be Unfriendly. Of course, it could very well correct itself later on, that is in fact at the core of a PAC algorithm, but then we get into questions of FOOM-ing and labels of human value in the environment which I am not equipped to deal with.
Is your thought that Friendliness is probably an easier ‘binding’ to figure out how to code than are, say, resisting Pascal’s mugging, or having consistent arithmetical reasoning?
To explain what I have in mind, consider Ben Goertzel’s example of how to test for general intelligence:
...when a robot can enrol in a human university and take classes in the same way as humans, and get its degree, then I’ll [say] we’ve created [an]… artificial general intelligence.
I do not disagree that such a robot, when walking towards the classroom, if it is being obstructed by a fellow human student, could attempt to kill this human, in order to get to the classroom.
Killing a fellow human, from the perspective of the human creators of the robot, is clearly a mistake. From a human perspective, it means that the robot failed.
You believe that the robot was just following its programming/construction. Indeed, the robot is its programming. I agree with this. I agree that the human creators were mistaken about what dynamic state sequence the robot will exhibit by computing the code.
What the “dumb superintelligence” argument tries to highlight is that if humans are incapable of predicting such behavior, then they will also be mistaken about predicting behavior that is harmful to the robots power. For example, while fighting with the human in order to kill it, for a split-second it mistakes its own arm with that of the human and breaks it.
You might now argue that such a robot isn’t much of a risk. It is pretty stupid to mistake its own arm with that of the enemy it tries to kill. True. But the point is that there is no relevant difference between failing to predict behavior that will harm the robot itself, and behavior that will harm a human. Except that you might believe the former is much easier than the latter. I dispute this.
For the robot to master a complex environment, like a university full of humans, without harming itself, or decreasing the chance of achieving its goals, is already very difficult. Not stabbing or strangling other human students is not more difficult than not jumping from the 4th floor, instead of taking the stairs. This is the “dumb superintelligence” argument.
What the “dumb superintelligence” argument tries to highlight is that if humans are incapable of predicting such behavior, then they will also be mistaken about predicting behavior that is harmful to the robots power.
To some extent. Perhaps it would be helpful to distinguish four different kinds of defeater:
early intelligence defeater: We try to build a seed AI, but our self-rewriting AI quickly hits a wall or explodes. This is most likely if we start with a subhuman intelligence and have serious resource constraints (so we can’t, e.g., just run an evolutionary algorithm over millions of copies of the AGI until we happen upon a variant that works).
late intelligence defeater: The seed AI works just fine, but at some late stage, when it’s already at or near superintelligence, it suddenly explodes. Apparently it went down a blind alley at some point early on that led it to plateau or self-destruct later on, and neither it nor humanity is smart enough yet to figure out where exactly the problem arose. So the FOOM fizzles.
early Friendliness defeater: From the outset, the seed AI’s behavior already significantly diverges from Friendliness.
late Friendliness defeater: The seed AI starts off as a reasonable approximation of Friendliness, but as it approaches superintelligence its values diverge from anything we’d consider Friendly, either because it wasn’t previously smart enough to figure out how to self-modify while keeping its values stable, or because it was never perfectly Friendly and the new circumstances its power puts it in now make the imperfections much more glaring.
In general, late defeaters are much harder for humans to understand than early defeaters, because an AI undergoing FOOM is too fast and complex to be readily understood. Your three main arguments, if I’m understanding them, have been:
(a) Early intelligence defeaters are so numerous that there’s no point thinking much about other kinds of defeaters yet.
(b) Friendliness defeaters imply a level of incompetence on the programmers’ part that strongly suggest intelligence defeaters will arise in the same situation.
(c) If an initially somewhat-smart AI is smart enough to foresee and avoid late intelligence defeaters, then an initially somewhat-nice AI should be smart enough to foresee and avoid late Friendliness defeaters.
I reject (a), because I haven’t seen any specific reason a self-improving AGI will be particularly difficult to make FOOM—‘it would require lots of complicated things to happen’ is very nearly a fully general argument against any novel technology, so I can’t get very far on that point alone. I accept (b), at least for a lot of early defeaters. But my concern is that while non-Friendliness predicts non-intelligence (and non-intelligence predicts non-Friendliness), intelligence also predicts non-Friendliness.
But our interesting disagreement seems to be over (c). Interesting because it illuminates general differences between the basic idea of a domain-general optimization process (intelligence) and the (not-so-)basic idea of Everything Humans Like. One important difference is that if an AGI optimizes for anything, it will have strong reason to steer clear of possible late intelligence defeaters. Late Friendliness defeaters, on the other hand, won’t scare optimization-process-optimizers in general.
It’s easy to see in advance that most beings that lack obvious early Friendliness defeaters will nonetheless have late Friendliness defeaters. In contrast, it’s much less clear that most beings lacking early intelligence defeaters will have late intelligence defeaters. That’s extremely speculative at this point—we simply don’t know what sorts of intelligence-destroying attractors might exist out there, or what sorts of paradoxes and complications are difficult v. trivial to overcome.
there is no relevant difference between failing to predict behavior that will harm the robot itself, and behavior that will harm a human. Except that you might believe the former is much easier than the latter. I dispute this.
But, once again, it doesn’t take any stupidity on the AI’s part to disvalue physically injuring a human, even if it does take stupidity to not understand that one is physically injuring a human. It only takes a different value system. Valuing one’s own survival is not orthogonal to valuing becoming more intelligent; but valuing human survival is orthogonal to valuing becoming more intelligent. (Indeed, to the extent they aren’t orthogonal it’s because valuing becoming more intelligent tends to imply disvaluing human survival, because humans are hard to control and made of atoms that can be used for other ends, including increased computing power.) This is the whole point of the article we’re commenting on.
Your three main arguments, if I’m understanding them, have been:
Here is part of my stance towards AI risks:
1. I assign a negligible probability to the possibility of a sudden transition from well-behaved narrow AIs to general AIs (see below).
2. An AI will not be pulled at random from mind design space. An AI will be the result of a research and development process. A new generation of AIs will need to be better than other products at “Understand What Humans Mean” and “Do What Humans Mean”, in order to survive the research phase and subsequent market pressure.
3. Commercial, research or military products are created with efficiency in mind. An AI that was prone to take unbounded actions given any terminal goal would either be fixed or abandoned during the early stages of research. If early stages showed that inputs such as the natural language query would yield results such as then the AI would never reach a stage in which it was sufficiently clever and trained to understand what results would satisfy its creators in order to deceive them.
4. I assign a negligible probability to the possibility of a consequentialist AI / expected utility maximizer / approximation to AIXI.
Given that the kind of AIs from point 4 are possible:
5. Omohundro’s AI drives are what make the kind of AIs mentioned in point 1 dangerous. Making an AI that does not exhibit these drives in an unbounded manner is probably a prerequisite to get an AI to work at all (there are not enough resources to think about being obstructed by simulator gods etc.), or should otherwise be easy compared to the general difficulties involved in making an AI work using limited resources.
6. An AI from point 4 will only ever do what it has been explicitly programmed to do. Such an AI is not going to protect its utility-function, acquire resources or preemptively eliminate obstacles in an unbounded fashion. Because it is not intrinsically rational to do so. What specifically constitutes rational, economic behavior is inseparable with an agent’s terminal goal. That any terminal goal can be realized in an infinite number of ways implies an infinite number of instrumental goals to choose from.
7. Unintended consequences are by definition not intended. They are not intelligently designed but detrimental side effects, failures. Whereas intended consequences, such as acting intelligently, are intelligently designed. If software was not constantly improved to be better at doing what humans intend it to do we would never be able to reach a level of sophistication where a software could work well enough to outsmart us. To do so it would have to work as intended along a huge number of dimensions. For an AI to constitute a risk as a result of unintended consequences those unintended consequences would have to have no, or little, negative influence on the huge number of intended consequences that are necessary for it to be able to overpower humanity.
I haven’t seen any specific reason a self-improving AGI will be particularly difficult to make FOOM...
I am not yet at a point of my education where I can say with confidence that this is the wrong way to think, but I do believe it is.
If someone walked up to you and told you about a risk only he can solve, and that you should therefore give this person money, would you give him money because you do not see any specific reason for why he could be wrong? Personally I would perceive the burden of proof to be on him to show me that the risk is real.
Despite this, I have specific reasons to personally believe that the kind of AI you have in mind is impossible. I have thought about such concepts as consequentialism / expected utility maximization, and do not see that they could be made to work, other than under very limited circumstances. And I also asked other people, outside of LessWrong, who are more educated and smarter than me, and they also told me that these kind of AIs are not feasible, they are uncomputable.
But our interesting disagreement seems to be over (c).
I am not sure I understand what you mean by c. I don’t think I agree with it.
One important difference is that if an AGI optimizes for anything,
I don’t know what this means.
Valuing one’s own survival is not orthogonal to valuing becoming more intelligent; but valuing human survival is orthogonal to valuing becoming more intelligent.
That this black box you call “intelligence” might be useful to achieve a lot of goals is not an argument in support of humans wanting to and succeeding at the implementation of “value to maximize intelligence” in conjunction with “by all means”.
Most definitions of intelligence that I am aware of are in terms of the ability to achieve goals. Saying that a system values to become more intelligent then just means that a system values to increase its ability to achieve its goals. In this context, what you suggest is that humans will want to, and will succeed to, implement an AI that in order to beat humans at Tic-tac-toe is first going to take over the universe and make itself capable of building such things as Dyson spheres.
What I am saying is that it is much easier to create a Tic-tac-toe playing AI, or an AI that can earn a university degree, than the former in conjunction with being able to take over the universe and build Dyson spheres.
The argument that valuing not to kill humans is orthogonal to taking over the universe and building Dyson spheres is completely irrelevant.
An AI will not be pulled at random from mind design space.
I don’t think anyone’s ever disputed this. (However, that’s not very useful if the deterministic process resulting in the SI is too complex for humans to distinguish it in advance from the outcome of a random walk.)
An AI will be the result of a research and development process. A new generation of AIs will need to be better than other products at “Understand What Humans Mean” and “Do What Humans Mean”, in order to survive the research phase and subsequent market pressure.
Agreed. But by default, a machine that is better than other rival machines at satisfying our short-term desires will not satisfy our long-term desires. The concern isn’t that we’ll suddenly start building AIs with the express purpose of hitting humans in the face with mallets. The concern is that we’ll code for short-term rather than long-term goals, due to a mixture of disinterest in Friendliness and incompetence at Friendliness. But if intelligence explosion occurs, ‘the long run’ will arrive very suddenly, and very soon. So we need to adjust our research priorities to more seriously assess and modulate the long-term consequences of our technology.
An AI that was prone to take unbounded actions given any terminal goal would either be fixed or abandoned during the early stages of research.
That may be a reason to think that recursively self-improving AGI won’t occur. But it’s not a reason to expect such AGI, if it occurs, to be Friendly.
If early stages showed that inputs such as the natural language query would yield results such as
The seed is not the superintelligence. We shouldn’t expect the seed to automatically know whether the superintelligence will be Friendly, any more than we should expect humans to automatically know whether the superintelligence will be Friendly.
Making an AI that does not exhibit these drives in an unbounded manner is probably a prerequisite to get an AI to work at all (there are not enough resources to think about being obstructed by simulator gods etc.)
I’m not following. Why does an AGI have to have a halting condition (specifically, one that actually occurs at some point) in order to be able to productively rewrite its own source code?
An AI from point 4 will only ever do what it has been explicitly programmed to do.
You don’t seem to be internalizing my arguments. This is just the restatement of a claim I pointed out was not just wrong but dishonestly statedhere.
That any terminal goal can be realized in an infinite number of ways implies an infinite number of instrumental goals to choose from.
Sure, but the list of instrumental goals overlap more than the list of terminal goals, because energy from one project can be converted to energy for a different project. This is an empirical discovery about our world; we could have found ourselves in the sort of universe where instrumental goals don’t converge that much, e.g., because once energy’s been locked down into organisms or computer chips you just Can’t convert it into useful work for anything else. In a world where we couldn’t interfere with the AI’s alien goals, nor could our component parts and resources be harvested to build very different structures, nor could we be modified to work for the AI, the UFAI would just ignore us and zip off into space to try and find more useful objects. We don’t live in that world because complicated things can be broken down into simpler things at a net gain in our world, and humans value a specific set of complicated things.
‘These two sets are both infinite’ does not imply ‘we can’t reason about these two things’ relative size, or how often the same elements recur in their elements’.
I am not yet at a point of my education where I can say with confidence that this is the wrong way to think, but I do believe it is.
If someone walked up to you and told you about a risk only he can solve, and that you should therefore give this person money, would you give him money because you do not see any specific reason for why he could be wrong? Personally I would perceive the burden of proof to be on him to show me that the risk is real.
You’ve spent an awful lot of time writing about the varied ways in which you’ve not yet been convinced by claims you haven’t put much time into actively investigating. Maybe some of that time could be better spent researching these topics you keep writing about? I’m not saying to stop talking about this, but there’s plenty of material on a lot of these issues to be found. Have you read Intelligence Explosion Microeconomics?
succeeding at the implementation of “value to maximize intelligence” in conjunction with “by all means”.
As a rule, adding halting conditions adds complexity to an algorithm, rather than removing complexity.
Saying that a system values to become more intelligent then just means that a system values to increase its ability to achieve its goals.
No, this is a serious misunderstanding. Yudkowsky’s definition of ‘intelligence’ is about the ability to achieve goals in general, not about the ability to achieve the system’s goals. That’s why you can’t increase a system’s intelligence by lowering its standards, i.e., making its preferences easier to satisfy.
what you suggest is that humans will want to, and will succeed to, implement an AI that in order to beat humans at Tic-tac-toe is first going to take over the universe and make itself capable of building such things as Dyson spheres.
Straw-man; no one has claimed that humans are likely to want to create an UFAI. What we’ve suggested is that humans are likely to want to create an algorithm, X, that will turn out to be a UFAI. (In other words, the fallacy you’re committing is confusing intension with extension.)
That aside: Are you saying Dyson spheres wouldn’t be useful for beating more humans at more tic-tac-toe games? Seems like a pretty good way to win at tic-tac-toe to me.
Yudkowsky’s definition of ‘intelligence’ is about the ability to achieve goals in general, not about the ability to achieve the system’s goals. That’s why you can’t increase a system’s intelligence by lowering its standards, i.e., making its preferences easier to satisfy.
Actually I do define intelligence as ability to hit a narrow outcome target relative to your own goals, but if your goals are very relaxed then the volume of outcome space with equal or greater utility will be very large. However one would expect that many of the processes involved in hitting a narrow target in outcome space (such that few other outcomes are rated equal or greater in the agent’s preference ordering), such as building a good epistemic model or running on a fast computer, would generalize across many utility functions; this is why we can speak of properties apt to intelligence apart from particular utility functions.
Actually I do define intelligence as ability to hit a narrow outcome target relative to your own goals
Hmm. But this just sounds like optimization power to me. You’ve defined intelligence in the past as “efficient cross-domain optimization”. The “cross-domain” part I’ve taken to mean that you’re able to hit narrow targets in general, not just ones you happen to like. So you can become more intelligent by being better at hitting targets you hate, or by being better at hitting targets you like.
The former are harder to test, but something you’d hate doing now could become instrumentally useful to know how to do later. And your intelligence level doesn’t change when the circumstance shifts which part of your skillset is instrumentally useful. For that matter, I’m missing why it’s useful to think that your intelligence level could drastically shift if your abilities remained constant but your terminal values were shifted. (E.g., if you became pickier.)
No, “cross-domain” means that I can optimize across instrumental domains. Like, I can figure out how to go through water, air, or space if that’s the fastest way to my destination, I am not limited to land like a ground sloth.
Measured intelligence shouldn’t shift if you become pickier—if you could previously hit a point such that only 1/1000th of the space was more preferred than it, we’d still expect you to hit around that narrow a volume of the space given your intelligence even if you claimed afterward that a point like that only corresponded to 0.25 utility on your 0-1 scale instead of 0.75 utility due to being pickier ([expected] utilities sloping more sharply downward with increasing distance from the optimum).
But by default, a machine that is better than other rival machines at satisfying our short-term desires will not satisfy our long-term desires.
You might be not aware of this but I wrote a sequence of short blog posts where I tried to think of concrete scenarios that could lead to human extinction. Each of which raised many questions.
What might seem to appear completely obvious to you for reasons that I do not understand, e.g. that an AI can take over the world, appears to me largely like magic (I am not trying to be rude, by magic I only mean that I don’t understand the details). At the very least there are a lot of open questions. Even given that for the sake of the above posts I accepted that the AI is superhuman and can do such things as deceive humans by its superior knowledge of human psychology. Which seems to be non-trivial assumption, to say the least.
That may be a reason to think that recursively self-improving AGI won’t occur. But it’s not a reason to expect such AGI, if it occurs, to be Friendly.
Over and over I told you that given all your assumptions, I agree that AGI is an existential risk.
The seed is not the superintelligence. We shouldn’t expect the seed to automatically know whether the superintelligence will be Friendly, any more than we should expect humans to automatically know whether the superintelligence will be Friendly.
You did not reply to my argument. My argument was that if the seed is unfriendly then it will not be smart enough to hide its unfriendliness. My argument did not pertain the possibility of a friendly seed turning unfriendly.
Why does an AGI have to have a halting condition (specifically, one that actually occurs at some point) in order to be able to productively rewrite its own source code?
What I have been arguing is that an AI should not be expected, by default, to want to eliminate all possible obstructions. There are many graduations here. That, by some economic or otherwise theoretic argument, it might be instrumentally rational for some ideal AI to take over the world, does not mean that humans would create such an AI, or that an AI could not be limited to care about fires in its server farm rather than that Russia might nuke the U.S. and thereby destroy its servers.
You don’t seem to be internalizing my arguments.
Did you mean to reply to another point? I don’t see how the reply you linked to is relevant to what I wrote.
Sure, but the list of instrumental goals overlap more than the list of terminal goals, because energy from one project can be converted to energy for a different project.
My argument is that an AI does not need to consider all possible threats and care to acquire all possible resources. Based on its design it could just want to optimize using its initial resources while only considering mundane threats. I just don’t see real-world AIs to conclude that they need to take over the world. I don’t think an AI is likely going to be designed that way. I also don’t think such an AI could work, because such inferences would require enormous amounts of resources.
You’ve spent an awful lot of time writing about the varied ways in which you’ve not yet been convinced by claims you haven’t put much time into actively investigating. Maybe some of that time could be better spent researching these topics you keep writing about?
I have done what is possible given my current level of education and what I perceive to be useful. I have e.g. asked experts about their opinion.
A few general remarks about the kind of papers such as the one that you linked to.
How much should I update towards MIRI’s position if I (1) understood the arguments in the paper (2) found the arguments convincing?
My answer is the following. If the paper was about the abc conjecture, the P versus NP problem, climate change, or even such mundane topics as psychology, I would either not be able to understand the paper, would be unable to verify the claims, or would have very little confidence in my judgement.
So what about ‘Intelligence Explosion Microeconomics’? That I can read most of it is only due to the fact that it is very informally written. The topic itself is more difficult and complex than all of the above mentioned problems together. Yet the arguments in support of it, to exaggerate a little bit, contain less rigor than the abstract of one of Shinichi Mochizuki’s papers on the abc conjecture.
Which means that my answer is that I should update very little towards MIRI’s position and that any confidence I gain about MIRI’s position is probably highly unreliable.
Thanks. My feeling is that to gain any confidence into what all this technically means, and to answer all the questions this raises, I’d probably need about 20 years of study.
No, this is a serious misunderstanding. Yudkowsky’s definition of ‘intelligence’ is
Here is part of a post exemplifying how I understand the relation between goals and intelligence:
If a goal has very few constraints then the set that satisfies all constraints is very large. A vague and ambiguous goal allows for too much freedom in the sense that a wide range of world states would have the same expected value and therefore imply a very large solution space, since a wide range of AI’s will be able to achieve those world states and thereby satisfy the condition of being improved versions of their predecessor.
This means that in order to get an AI to become superhuman at all, and very quickly in particular, you will need to encode a very specific goal against which mistakes, optimization power and achievement can be judged.
It is really hard to communicate how I perceive this and other discussions about MIRI’s position without offending people, or killing the discussion.
I am saying this in full honesty. The position you appear to support seems so utterly “complex” (far-fetched) that the current arguments are unconvincing.
Here is my perception of the scenario that you try to sell me (exaggerated to make a point). I have a million questions about it that I can’t answer and which your answers either sidestep or explain away by using “magic”.
At this point I probably made 90% of the people reading this comment incredible angry. My perception is that you cannot communicate this perception on LessWrong without getting into serious trouble. That’s also what I meant when I told you that I cannot be completely honest if you want to discuss this on LessWrong.
I can also assure you that many people who are much smarter and higher status than me think so as well. Many people communicated the absurdity of all this to me but told me that they would not repeat this in public.
My argument was that if the seed is unfriendly then it will not be smart enough to hide its unfriendliness.
Pretending to be friendly when you’re actually not is something that doesn’t even require human level intelligence. You could even do it accidentally.
In general, the appearance of Friendliness at low levels of ability to influence the world doesn’t guarantee actual Friendliness at high levels of ability to influence the world. (If it did, elected politicians would be much higher quality.)
But our interesting disagreement seems to be over (c). Interesting because it illuminates general differences between the basic idea of a domain-general optimization process (intelligence) and the (not-so-)basic idea of Everything Humans Like. One important difference is that if an AGI optimizes for anything, it will have strong reason to steer clear of possible late intelligence defeaters. Late Friendliness defeaters, on the other hand, won’t scare optimization-process-optimizers in general.
But it will scare friendly ones, which will want to keep their values stable.
But, once again, it doesn’t take any stupidity on the AI’s part to disvalue physically injuring a human,
But it will scare friendly ones, which will want to keep their values stable.
Yes. If an AI is Friendly at one stage, then it is Friendly at every subsequent stage. This doesn’t help make almost-Friendly AIs become genuinely Friendly, though.
It takes stupidity to misinterpret friendlienss.
Yes, but that’s stupidity on the part of the human programmer, and/or on the part of the seed AI if we ask it for advice. The superintelligence didn’t write its own utility function; the superintelligence may well understand Friendliness perfectly, but that doesn’t matter if it hasn’t been programmed to rewrite its source code to reflect its best understanding of ‘Friendliness’. The seed is not the superintelligence. See: http://lesswrong.com/lw/igf/the_genie_knows_but_doesnt_care/
Yes, but that’s stupidity on the part of the human programmer, and/or on the part of the seed AI if we ask it for advice.
That depends on the architecture. In a Loosemore architecture, the AI interprets high-level directives itself, so if it gets them wrong, that’s it’s mistake.
Say we find an algorithm for producing progressively more accurate beliefs about itself and the world. This algorithm may be long and complicated—perhaps augmented by rules-of-thumb whenever the evidence available to it says these rules make better predictions. (E.g, “nine times out of ten the Enterprise is not destroyed.”) Combine this with an arbitrary goal and we have the making of a seed AI.
Seems like this could straightforwardly improve its ability to predict humans without changing its goal, which may be ‘maximize pleasure’ or ‘maximize X’. Why would it need to change its goal?
We seem to agree that for an AI to talk itself out of a confinement (like in the AI box experiment), the AI would have to understand what humans mean and want.
As far as I understand your position, you believe that it is difficult to make an AI care to do what humans want, apart from situations where it is temporarily instrumentally useful to do what humans want.
Do you agree that for such an AI to do what humans want, in order to deceive them, humans would have to succeed at either encoding the capability to understand what humans want, or succeed at encoding the capability to make itself capable of understanding what humans want?
My question, do you believe there to be a conceptual difference between encoding capabilities, what an AI can do, and goals, what an AI will do? As far as I understand, capabilities and goals are both encodings of how humans want an AI to behave.
In other words, humans intend an AI to be intelligent and use its intelligence in a certain way. And in order to be an existential risk, humans need to succeed making and AI behave intelligently but fail at making it use its intelligence in a way that does not kill everyone.
Do you agree?
Your summaries of my views here are correct, given that we’re talking about a superintelligence.
Well, there’s obviously a difference; ‘what an AI can do’ and ‘what an AI will do’ mean two different things. I agree with you that this difference isn’t a particularly profound one, and the argument shouldn’t rest on it.
What the argument rests on is, I believe, that it’s easier to put a system into a positive feedback loop that helps it better model its environment and/or itself, than it is to put a system into a positive feedback loop that helps it better pursue a specific set of highly complex goals we have in mind (but don’t know how to fully formalize).
If the AI incorrectly models some feature of itself or its environment, reality will bite back. But if it doesn’t value our well-being, how do we make reality bite back and change the AI’s course? How do we give our morality teeth?
Whatever goals it initially tries to pursue, it will fail in those goals more often the less accurate its models are of its circumstances; so if we have successfully programmed it to do increasingly well at any difficult goal at all (even if it’s not the goal we intended it to be good at), then it doesn’t take a large leap of the imagination to see how it could receive feedback from its environment about how well it’s doing at modeling states of affairs. ‘Modeling states of affairs well’ is not a highly specific goal, it’s instrumental to nearly all goals, and it’s easy to measure how well you’re doing at it if you’re entangled with anything about your environment at all, e.g., your proximity to a reward button.
(And when a system gets very good at modeling itself, its environment, and the interactions between the two, such that it can predict what changes its behaviors are likely to effect and choose its behaviors accordingly, then we call its behavior ‘intelligent’.)
This stands in stark contrast to the difficulty of setting up a positive feedback loop that will allow an AGI to approximate our True Values with increasing fidelity. We understand how accurately modeling something works; we understand the basic principles of intelligence. We don’t understand the basic principles of moral value, and we don’t even have a firm grasp about how to go about finding out the answer to moral questions. Presumably our values are encoded in some way in our brains, such that there is some possible feedback loop we could use to guide an AGI gradually toward Friendliness. But how do we figure out in advance what that feedback loop needs to look like, without asking the superintelligence? (We can’t ask the superintelligence what algorithm to use to make it start becoming Friendly, because to the extent it isn’t already Friendly it isn’t a trustworthy source of information. This is in addition to the seed/intelligence distinction I noted above.)
If we slightly screw up the AGI’s utility function, it will still need to to succeed at modeling things accurately in order to do anything complicated at all. But it will not need to succeed at optimally caring about what humans care about in order to do anything complicated at all.
This can be understood as both a capability and as a goal. What humans mean an AI to do is to undergo recursive self-improvement. What humans mean an AI to be capable of is to undergo recursive self-improvement.
I am only trying to clarify the situation here. Please correct me if you think that above is wrong.
I do not disagree with the orthogonality thesis insofar as an AI can have goals that interfere with human values in a catastrophic way, possibly leading to human extinction.
I believe here is where we start to disagree. I do not understand how the “improvement” part of recursive self-improvement can be independent of properties such as the coherence and specificity of the goal the AI is supposed to achieve.
Either you have a perfectly specified goal, such as “maximizing paperclips”, where it is clear what “maximization” means, and what the properties of “paperclips” are, or there is some amount of uncertainty about what it means to achieve the goal of “maximizing paperclips”.
Consider the programmers forgot to encode what shape the paperclips are supposed to have. How do you suppose would that influence the behavior of the AI. Would it just choose some shape at random, or would it conclude that shape is not part of its goal? If the former, where would the decision to randomly choose a shape come from? If the latter, what would it mean to maximize shapeless objects?
I am just trying to understand what kind of AI you have in mind.
This is a clearer point of disagreement.
An AI needs to be able to draw clear lines where exploration ends and exploitation starts. For example, an AI that thinks about every decision for a year would never get anything done.
An AI also needs to discount low probability possibilities, as to not be vulnerable to internal or external Pascal’s mugging scenarios.
These are problems that humans need to solve and encode in order for an AI to be a danger.
But these problems are in essence confinements, or bounds on how an AI is going to behave.
How likely is an AI then going to take over the world, or look for dangerous aliens, in order to make sure that neither aliens nor humans obstruct it from achieving its goal?
Similarly, how likely is such an AI to convert all resources into computronium in order to be better able to model states of affairs well?
I understand this. And given your assumptions about how an AI will affect the whole world in a powerful way, it makes sense to make sure that it does so in a way that preserves human values.
I have previously compared this to uncontrollable self-replicating nanobots. Given that you cannot confine the speed or scope of their self-replication, only the nature of the transformation that they cause, you will have to make sure that they transform the world into a paradise rather than grey goo.
“uncertainty” is in your human understanding of the program, not in the actual program. A program doesn’t go “I don’t know what I’m supposed to do next”, it follows instructions step-by-step.
It would mean exactly what it’s programmed to mean, without any uncertainty in it at all.
Yes. To divide it more finely, it could be a terminal goal, or an instrumental goal; it could be a goal of the AI, or a goal of the human; it could be a goal the human would reflectively endorse, or a goal the human would reflectively reject but is inadvertently promoting anyway.
I agree that, at a given time, the AI must have a determinate goal. (Though the encoding of that goal may be extremely complicated and unintentional. And it may need to be time-indexed.) I’m not dogmatically set on the idea that a self-improving AGI is easy to program; at this point it wouldn’t shock me if it took over 100 years to finish making the thing. What you’re alluding to are the variety of ways we could fail to construct a self-improving AGI at all. Obviously there are plenty of ways to fail to make an AGI that can improve its own ability to track things about its environment in a domain-general way, without bursting into flames at any point. If there weren’t plenty of ways to fail, we’d have already succeeded.
Our main difference in focus is that I’m worried about what happens if we do succeed in building a self-improving AGI that doesn’t randomly melt down. Conditioned on our succeeding in the next few centuries in making a machine that actually optimizes for anything at all, and that optimizes for its own ability to generally represent its environment in a way that helps it in whatever else it’s optimizing for, we should currently expect humans to go extinct as a result. Even if the odds of our succeeding in the next few centuries were small, it would be worth thinking about how to make that extinction event less likely. (Though they aren’t small.)
I gather that you think that making an artificial process behave in any particular way at all (i.e., optimizing for something), while recursively doing surgery on its own source code in the radical way MIRI is interested in, is very tough. My concern is that, no matter how true that is, it doesn’t entail that if we succeed at that tough task, we’ll therefore have made much progress on other important tough tasks, like Friendliness. It does give us more time to work on Friendliness, but if we convince ourselves that intelligence explosion is a completely pie-in-the-sky possibility, then we won’t use that time effectively.
I also gather that you have a hard time imagining our screwing up on a goal architecture without simply breaking the AGI. Perhaps by ‘screwing up’ you’re imagining failing to close a set of parentheses. But I think you should be at least as worried about philosophical, as opposed to technical, errors. A huge worry isn’t just that we’ll fail to make the AI we intended; it’s that our intentions while we’re coding the thing will fail to align with the long-term interests of ourselves, much less of the human race.
We agree that it’s possible to ‘bind’ a superintelligence. (By this you don’t mean boxing it; you just mean programming it to behave in some ways as opposed to others.) But if the bindings fall short of Friendliness, while enabling superintelligence to arise at all, then a serious risk remains. Is your thought that Friendliness is probably an easier ‘binding’ to figure out how to code than are, say, resisting Pascal’s mugging, or having consistent arithmetical reasoning?
I am trying to understand if the kind of AI, that is underlying the scenario that you have in mind, is a possible and likely outcome of human AI research.
As far as I am aware, as a layman, goals and capabilities are intrinsically tied together. How could a chess computer be capable of winning against humans at chess without the terminal goal of achieving a checkmate?
Coherent and specific goals are necessary to (1) decide which actions are instrumental useful (2) judge the success of self-improvement. If the given goal is logically incoherent, or too vague for the AI to be able to tell apart success from failure, would it work at all?
If I understand your position correctly, you would expect a chess playing general AI, one that does not know about checkmate, instead of “winning at chess”, to improve against such goals as “modeling states of affairs well” or “make sure nothing intervenes chess playing”. You believe that these goals do not have to be programmed by humans, because they are emergent goals, an instrumental consequence of being general intelligent.
These universal instrumental goals, these “AI drives”, seem to be a major reason for why you believe it to be important to make the AI care about behaving correctly. You believe that these AI drives are a given, and the only way to prevent an AI from being an existential risk is to channel these drives, is to focus this power on protecting and amplifying human values.
My perception is that these drives that you imagine are not special and will be as difficult to get “right” than any other goal. I think that the idea that humans not only want to make an AI exhibit such drives, but also succeed at making such drives emerge, is a very unlikely outcome.
As far as I am aware, here is what you believe an AI to want:
It will want to self-improve
It will want to be rational
It will try to preserve their utility functions
It will try to prevent counterfeit utility
It will be self-protective It will want to acquire resources and use them efficiently
What AIs that humans would ever want to create would require all of these drives, and how easy will it be for humans to make an AI exhibit these drives compared to making an AI that can do what humans want without these drives?
Take mathematics. What are the difficulties associated with making an AI better than humans at mathematics, and will an AI need these drives in order to do so?
Humans did not evolve to play chess or do mathematics. Yet it is considerably more difficult to design a chess AI than an AI that is capable of discovering interesting and useful mathematics.
I believe that the difficulty is due to the fact that it is much easier to formalize what it means to play chess than doing mathematics. The difference between chess and mathematics is that chess has a specific terminal goal in the form of a clear definition of what constitutes winning. Although mathematics has unambiguous rules, there is no specific terminal goal and no clear definition of what constitutes winning.
The progress of the capability of artificial intelligence is not only related to whether humans have evolved for a certain skill or to how much computational resources it requires but also to how difficult it is to formalize the skill, its rules and what it means to succeed.
In the light of this, how difficult would it be to program the drives that you imagine, versus just making an AI win against humans at a given activity without exhibiting these drives?
All these drives are very vague ideas, not like “winning at chess”, but more like “being better at mathematics than Terence Tao”.
The point I am trying to make is that these drives constitute additional complexity, rather than being simple ideas that you can just assume, and from which you can reason about the behavior of an AI.
It is this context that the “dumb superintelligence” argument tries to highlight. It is likely incredibly hard to make these drives emerge in a seed AI. They implicitly presuppose that humans succeed at encoding intricate ideas about what “winning” means in all those cases required to overpower humans, but not in the case of e.g. winning at chess or doing mathematics. I like to analogize such a scenario to the creation of a generally intelligent autonomous car that works perfectly well at not destroying itself in a crash but which somehow manages to maximize the number of people to run over.
I agree that if you believe that it is much easier to create a seed AI to exhibit the drives that you imagine, than it is to make a seed AI use its initial resources to figure out how to solve a specific problem, then we agree about AI risks.
Humans are capable of winning at chess without the terminal goal of doing so. Nor were humans designed by evolution specifically for chess. Why should we expect a general superintelligence to have intelligence that generalizes less easily than a human’s does?
You keep coming back to this ‘logically incoherent goals’ and ‘vague goals’ idea. Honestly, I don’t have the slightest idea what you mean by those things. A goal that can’t motivate one to do anything ain’t a goal; it’s decor, it’s noise. ‘Goals’ are just the outcomes systems tend to produce, especially systems too complex to be easily modeled as, say, physical or chemical processes. Certainly it’s possible for goals to be incredibly complicated, or to vary over time. But there’s no such thing as a ‘logically incoherent outcome’. So what’s relevant to our purposes is whether failing to make a powerful optimization process human-friendly will also consistently stop the process from optimizing for anything whatsoever.
Conditioned on a self-modifying AGI (say, an AGI that can quine its source code, edit it, then run the edited program and repeat the process) achieving domain-general situation-manipulating abilities (i.e., intelligence), analogous to humans’ but to a far greater degree, which of the AI drives do you think are likely to be present, and which absent? ‘It wants to self-improve’ is taken as a given, because that’s the hypothetical we’re trying to assess. Now, should we expect such a machine to be indifferent to its own survival and to the use of environmental resources?
Sometimes a more complex phenomenon is the implication of a simpler hypothesis. A much narrower set of goals will have intelligence-but-not-resource-acquisition as instrumental than will have both as instrumental, because it’s unlikely to hit upon a goal that requires large reasoning abilities but does not call for many material resources.
You haven’t given arguments suggesting that here. At most, you’ve given arguments against expecting a seed AI to be easy to invent. Be careful to note, to yourself and others, when you switch between the claims ‘a superintelligence is too hard to make’ and ‘if we made a superintelligence it would probably be safe’.
Well, I’m not sure what XXD means by them, but…
G1 (“Everything is painted red”) seems like a perfectly coherent goal. A system optimizing G1 paints things red, hires people to paint things red, makes money to hire people to paint things red, invents superior paint-distribution technologies to deposit a layer of red paint over things, etc.
G2 (“Everything is painted blue”) similarly seems like a coherent goal.
G3 (G1 AND G2) seems like an incoherent goal. A system with that goal… well, I’m not really sure what it does.
A system’s goals have to be some event that can be brought about. In our world, ‘2+2=4’ and ‘2+2=5’ are not goals; ‘everything is painted red and not-red’ may not be a goal for similar reasons. When we’re talking about an artificial intelligence’s preferences, we’re talking about the things it tends to optimize for, not the things it ‘has in mind’ or the things it believes are its preferences.
This is part of what makes the terminology misleading, and is also why we don’t ask ‘can a superintelligence be irrational?‘. Irrationality is dissonance between my experienced-‘goals’ (and/or, perhaps, reflective-second-order-‘goals’) and my what-events-I-produce-‘goals’; but we don’t care about the superintelligence’s phenomenology. We only care about what events it tends to produce.
Tabooing ‘goal’ and just talking about the events a process-that-models-its-environment-and-directs-the-future tends to produce would, I think, undermine a lot of XiXiDu’s intuitions about goals being complex explicit objects you have to painstakingly code in. The only thing that makes it more useful to model a superintelligence as having ‘goals’ than modeling a blue-minimizing robot as having ‘goals’ is that the superintelligence responds to environmental variation in a vastly more complicated way. (Because, in order to be even a mediocre programmer, its model-of-the-world-that-determines-action has to be more complicated than a simple camcorder feed.)
Oh.
Well, in that case, all right. If there exists some X a system S is in fact optimizing for, and what we mean by “S’s goals” is X, regardless of what target S “has in mind”, then sure, I agree that systems never have vague or logically incoherent goals.
Well, wait. Where did “models its environment” come from?
If we’re talking about the things S optimizes its environment for, not the things S “has in mind”, then it would seem that whether S models its environment or not is entirely irrelevant to the conversation.
In fact, given how you’ve defined “goal” here, I’m not sure why we’re talking about intelligence at all. If that is what we mean by “goal” then intelligence has nothing to do with goals, or optimizing for goals. Volcanoes have goals, in that sense. Protons have goals.
I suspect I’m still misunderstanding you.
From Eliezer’s Belief in Intelligence:
“Since I am so uncertain of Kasparov’s moves, what is the empirical content of my belief that ‘Kasparov is a highly intelligent chess player’? What real-world experience does my belief tell me to anticipate? [...]
“The empirical content of my belief is the testable, falsifiable prediction that the final chess position will occupy the class of chess positions that are wins for Kasparov, rather than drawn games or wins for Mr. G. [...] The degree to which I think Kasparov is a ‘better player’ is reflected in the amount of probability mass I concentrate into the ‘Kasparov wins’ class of outcomes, versus the ‘drawn game’ and ‘Mr. G wins’ class of outcomes.”
From Measuring Optimization Power:
“When I think you’re a powerful intelligence, and I think I know something about your preferences, then I’ll predict that you’ll steer reality into regions that are higher in your preference ordering. [...]
“Ah, but how do you know a mind’s preference ordering? Suppose you flip a coin 30 times and it comes up with some random-looking string—how do you know this wasn’t because a mind wanted it to produce that string?
“This, in turn, is reminiscent of the Minimum Message Length formulation of Occam’s Razor: if you send me a message telling me what a mind wants and how powerful it is, then this should enable you to compress your description of future events and observations, so that the total message is shorter. Otherwise there is no predictive benefit to viewing a system as an optimization process. This criterion tells us when to take the intentional stance.
“(3) Actually, you need to fit another criterion to take the intentional stance—there can’t be a better description that averts the need to talk about optimization. This is an epistemic criterion more than a physical one—a sufficiently powerful mind might have no need to take the intentional stance toward a human, because it could just model the regularity of our brains like moving parts in a machine.
“(4) If you have a coin that always comes up heads, there’s no need to say “The coin always wants to come up heads” because you can just say “the coin always comes up heads”. Optimization will beat alternative mechanical explanations when our ability to perturb a system defeats our ability to predict its interim steps in detail, but not our ability to predict a narrow final outcome. (Again, note that this is an epistemic criterion.)
“(5) Suppose you believe a mind exists, but you don’t know its preferences? Then you use some of your evidence to infer the mind’s preference ordering, and then use the inferred preferences to infer the mind’s power, then use those two beliefs to testably predict future outcomes. The total gain in predictive accuracy should exceed the complexity-cost of supposing that ‘there’s a mind of unknown preferences around’, the initial hypothesis.”
Notice that throughout this discussion, what matters is the mind’s effect on its environment, not any internal experience of the mind. Unconscious preferences are just as relevant to this method as are conscious preferences, and both are examples of the intentional stance. Note also that you can’t really measure the rationality of a system you’re modeling in this way; any evidence you raise for ‘irrationality’ could just as easily be used as evidence that the system has more complicated preferences than you initially thought, or that they’re encoded in a more distributed way than you had previously hypothesized.
My take-away from this is that there are two ways we generally think about minds on LessWrong: Rational Choice Theory, on which all minds are equally rational and strange or irregular behaviors are seen as evidence of strange preferences; and what we might call the Ideal Self Theory, on which minds’ revealed preferences can differ from their ‘true self’ preferences, resulting in irrationality. One way of unpacking my idealized values is that they’re the rational-choice-theory preferences I would exhibit if my conscious desires exhibited perfect control over my consciously controllable behavior, and those desires were the desires my ideal self would reflectively prefer, where my ideal self is the best trade-off between preserving my current psychology and enhancing that psychology’s understanding of itself and its environment.
We care about ideal selves when we think about humans, because we value our conscious, ‘felt’ desires (especially when they are stable under reflection) more than our unconscious dispositions. So we want to bring our actual behavior (and thus our rational-choice-theory preferences, the ‘preferences’ we talk about when we speak of an AI) more in line with our phenomenological longings and their idealized enhancements. But since we don’t care about making non-person AIs more self-actualized, but just care about how they tend to guide their environment, we generally just assume that they’re rational. Thus if an AI behaves in a crazy way (e.g., alternating between destroying and creating paperclips depending on what day of the week it is), it’s not because it’s a sane rational ghost trapped by crazy constraints. It’s because the AI has crazy core preferences.
Yes, in principle. But in practice, a system that doesn’t have internal states that track the world around it in a reliable and useable way won’t be able to optimize very well for anything particularly unlikely across a diverse set of environments. In other words, it won’t be very intelligent. To clarify, this is an empirical claim I’m making about what it takes to be particularly intelligent in our universe; it’s not part of the definition for ‘intelligent’.
Yes, that seems plausible.
I would say rather that modeling one’s environment is an effective tool for consistently optimizing for some specific unlikely thing X across a range of environments, so optimizers that do so will be more successful at optimizing for X, all else being equal, but it more or less amounts to the same thing.
But… so what?
I mean, it also seems plausible that optimizers that explicitly represent X as a goal will be more successful at consistently optimizing for X, all else being equal… but that doesn’t stop you from asserting that explicit representation of X is irrelevant to whether a system has X as its goal.
So why isn’t modeling the environment equally irrelevant? Both features, on your account, are optional enhancements an optimizer might or might not display.
It keeps seeming like all the stuff you quote and say before your last two paragraphs ought to provide an answer that question, but after reading it several times I can’t see what answer it might be providing. Perhaps your argument is just going over my head, in which case I apologize for wasting your time by getting into a conversation I’m not equipped for..
Maybe it will help to keep in mind that this is one small branch of my conversation with Alexander Kruel. Alexander’s two main objections to funding Friendly Artificial Intelligence research are that (1) advanced intelligence is very complicated and difficult to make, and (2) getting a thing to pursue a determinate goal at all is extraordinarily difficult. So a superintelligence will never be invented, or at least not for the foreseeable future; so we shouldn’t think about SI-related existential risks. (This is my steel-manning of his view. The way he actually argues seems to instead be predicated on inventing SI being tied to perfecting Friendliness Theory, but I haven’t heard a consistent argument for why that should be so.)
Both of these views, I believe, are predicated on a misunderstanding of how simple and disjunctive ‘intelligence’ and ‘goal’ are, for present purposes. So I’ve mainly been working on tabooing and demystifying those concepts. Intelligence is simply a disposition to efficiently convert a wide variety of circumstances into some set of specific complex events. Goals are simply the circumstances that occur more often when a given intelligence is around. These are both very general and disjunctive ideas, in stark contrast to Friendliness; so it will be difficult to argue that a superintelligence simply can’t be made, and difficult too to argue that optimizing for intelligence requires one to have a good grasp on Friendliness Theory.
Because I’m trying to taboo the idea of superintelligence, and explain what it is about seed AI that will allow it to start recursively improving its own intelligence, I’ve been talking a lot about the important role modeling plays in high-level intelligent processes. Recognizing what a simple idea modeling is, and how far it gets one toward superintelligence once one has domain-general modeling proficiency, helps a great deal with greasing the intuition pump ‘Explosive AGI is a simple, disjunctive event, a low-hanging fruit, relative to Friendliness.’ Demystifying unpacking makes things seem less improbable and convoluted.
I think this is a map/territory confusion. I’m not denying that superintelligences will have a map of their own preferences; at a bare minimum, they need to know what they want in order to prevent themselves from accidentally changing their own preferences. But this map won’t be the AI’s preferences—those may be a very complicated causal process bound up with, say, certain environmental factors surrounding the AI, or oscillating with time, or who-knows-what.
There may not be a sharp line between the ‘preference’ part of the AI and the ‘non-preference’ part. Since any superintelligence will be exemplary at reasoning with uncertainty and fuzzy categories, I don’t think that will be a serious obstacle.
Does that help explain why I’m coming from? If not, maybe I’m missing the thread unifying your comments.
I suppose it helps, if only in that it establishes that much of what you’re saying to me is actually being addressed indirectly to somebody else, so it ought not surprise me that I can’t quite connect much of it to anything I’ve said. Thanks for clarifying your intent.
For my own part, I’m certainly not functioning here as Alex’s proxy; while I don’t consider explosive intelligence growth as much of a foregone conclusion as many folks here do, I also don’t consider Alex’s passionate rejection of the possibility justified, and have had extended discussions on related subjects with him myself in past years. So most of what you write in response to Alex’s positions is largely talking right past me.
(Which is not to say that you ought not be doing it. If this is in effect a private argument between you and Alex that I’ve stuck my nose into, let me know and I’ll apologize and leave y’all to it in peace.)
Anyway, I certainly agree that a system might have a representation of its goals that is distinct from the mechanisms that cause it to pursue those goals. I have one of those, myself. (Indeed, several.) But if a system is capable of affecting its pursuit of its goals (for example, if it is capable of correcting the effects of a state-change that would, uncorrected, have led to value drift), it is not merely interacting with maps. It is also interacting with the territory… that is, it is modifying the mechanisms that cause it to pursue those goals… in order to bring that territory into line with its pre-existing map.
And in order to do that, it must have such a mechanism, and that mechanism must be consistently isomorphic to its representations of its goals.
Yes?
Right. I’m not saying that there aren’t things about the AI that make it behave the way it does; what the AI optimizes for is a deterministic result of its properties plus environment. I’m just saying that something about the environment might be necessary for it to have the sorts of preferences we can most usefully model it as having; and/or there may be multiple equally good candidates for the parts of the AI that are its values, or their encoding. If we reify preferences in an uncautious way, we’ll start thinking of the AI’s ‘desires’ too much as its first-person-experienced urges, as opposed to just thinking of them as the effect the local system we’re talking about tends to have on the global system.
Hm.
So, all right. Cconsider two systems, S1 and S2, both of which happen to be constructed in such a way that right now, they are maximizing the number of things in their environment that appear blue to human observers, by going around painting everything blue.
Suppose we add to the global system a button that alters all human brains so that everything appears blue to us, and we find that S1 presses the button and stops painting, and S2 ignores the button and goes on painting.
Suppose that similarly, across a wide range of global system changes, we find that S1 consistently chooses the action that maximizes the number of things in its environment that appear blue to human observers, while S2 consistently goes on painting.
I agree with you that if I reify S2′s preferences in an uncautious way, I might start thinkng of S2 as “wanting to paint things blue” or “wanting everything to be blue” or “enjoying painting things blue” or as having various other similar internal states that might simply not exist, and that I do better to say it has a particular effect on the global system. S2 simply paints things blue; whether it has the goal of painting things blue or not, I have no idea.
I am far less comfortable saying that S1 has no goals, precisely because of how flexibly and consistently it is revising its actions so as to consistently create a state-change across wide ranges of environments. To use Dennett’s terminology, I am more willing to adopt an intentional stance with respect to S1 than S2.
If I’ve understood your position correctly, you’re saying that I’m unjustified in making that distinction… that to the extent that we can say that S1 and S2 have “goals,” the word “goals” simply refer to the state changes they create in the world. Initially they both have the goal of painting things blue, but S1′s goals keep changing: first it paints things blue, then it presses a button, then it does other things. And, sure, I can make up some story like “S1 maximizes the number of things in its environment that appear blue to human observers, while S2 just paints stuff blue” and that story might even have predictive power, but I ought not fall into the trap of reifying some actual thing that corresponds to those notional “goals”.
Am I in the right ballpark?
I think you’re switching back and forth between a Rational Choice Theory ‘preference’ and an Ideal Self Theory ‘preference’. To disambiguate, I’ll call the former R-preferences and the latter I-preferences. My R-preferences—the preferences you’d infer I had from my behaviors if you treated me as a rational agent—are extremely convoluted, indeed they need to be strongly time-indexed to maintain consistency. My I-preferences are the things I experience a desire for, whether or not that desire impacts my behavior. (Or they’re the things I would, with sufficient reflective insight and understanding into my situation, experience a desire for.)
We have no direct evidence from your story addressing whether S1 or S2 have I-preferences at all. Are they sentient? Do they create models of their own cognitive states? Perhaps we have a little more evidence that S1 has I-preferences than that S2 does, but only by assuming that a system whose goals require more intelligence or theory-of-mind will have a phenomenology more similar to a human’s. I wouldn’t be surprised if that assumption turns out to break down in some important ways, as we explore more of mind-space.
But my main point was that it doesn’t much matter what S1 or S2′s I-preferences are, if all we’re concerned about is what effect they’ll have on their environment. Then we should think about their R-preferences, and bracket exactly what psychological mechanism is resulting in their behavior, and how that psychological mechanism relates to itself.
I’ve said that R-preferences are theoretical constructs that happen to be useful a lot of the time for modeling complex behavior; I’m not sure whether I-preferences are closer to nature’s joints.
S1′s instrumental goals may keep changing, because its circumstances are changing. But I don’t think its terminal goals are changing. The only reason to model it as having two completely incommensurate goal sets at different times would be if there were no simple terminal goal that could explain the change in instrumental behavior.
I don’t think I’m switching back and forth between I-preferences and R-preferences.
I don’t think I’m talking about I-preferences at all, nor that I ever have been.
I completely agree with you that they don’t matter for our purposes here, so if I am talking about them, I am very very confused. (Which is certainly possible.)
But I don’t think that R-preferences (preferences, goals, etc.) can sensibly be equated with the actual effects a local system has on a global system. If they could, we could talk equally sensibly about earthquakes having R-preferences (preferences, goals, etc.), and I don’t think it’s sensible to talk that way.
R-preferences (preferences, goals, etc.) are, rather, internal states of a system S.
If S is a competent optimizer (or “rational agent,” if you prefer) with R-preferences (preferences, goals, etc.) P, the existence of P will cause S to behave in ways that cause isomorphic effects (E) on a global system, so we can use observations of E as evidence of P (positing that S is a competent optimizer) or as evidence that S is a competent optimizer (positing the existence of P) or a little of both.
But however we slice it, P is not the same thing as E, E is merely evidence of P’s existence. We can infer P’s existence in other ways as well, even if we never observe E… indeed, even if E never gets produced. And the presence or absence of a given P in S is something we can be mistaken about; there’s a fact of the matter.
I think you disagree with the above paragraph, because you describe R-preferences (preferences, goals, etc.) as theoretical constructs rather than parts of the system, which suggests that there is no fact of the matter… a different theoretical approach might never include P, and it would not be mistaken, it would just be a different theoretical approach.
I also think that because way back at the beginning of this exchange when I suggested “paint everything red AND paint everything blue” was an example of an incoherent goal (R-preference, preference, P), your reply was that it wasn’t a goal at all, since that state can’t actually exist in the world. Which suggests that you don’t see goals as internal states of optimizers and that you do equate P with E.
This is what I’ve been disputing from the beginning.
But to be honest, I’m not sure whether you disagree or not, as I’m not sure we have yet succeeded in actually engaging with one another’s ideas in this exchange.
You can treat earthquakes and thunderstorms and even individual particles as having ‘preferences’. It’s just not very useful to do so, because we can give an equally simple explanation for what effects things like earthquakes tend to have that is more transparent about the physical mechanism at work. The intentional strategy is a heuristic for black-boxing physical processes that are too complicated to usefully describe in their physical dynamics, but that can be discussed in terms of the complicated outcomes they tend to promote.
(I’d frame it: We’re exploiting the fact that humans are intuitively dualistic by taking the non-physical modeling device of humans (theory of mind, etc.) and appropriating this mental language and concept-web for all sorts of systems whose nuts and bolts we want to bracket. Slightly regimented mental concepts and terms are useful, not because they apply to all the systems we’re talking about in the same way they were originally applied to humans, but because they’re vague in ways that map onto the things we’re uncertain about or indifferent to.)
‘X wants to do Y’ means that the specific features of X tend to result in Y when its causal influence is relatively large and direct. But, for clarity’s sake, we adopt the convention of only dropping into want-speak when a system is too complicated for us to easily grasp in mechanistic terms why it’s having these complex effects, yet when we can predict that, whatever the mechanism happens to be, it is the sort of mechanism that has those particular complex effects.
Thus we speak of evolution as an optimization process, as though it had a ‘preference ordering’ in the intuitively human (i.e., I-preference) sense, even though in the phenomenological sense it’s just as mindless as an earthquake. We do this because black-boxing the physical mechanisms and just focusing on the likely outcomes is often predictively useful here, and because the outcomes are complicated and specific. This is useful for AIs because we care about the AI’s consequences and not its subjectivity (hence we focused on R-preference), and because AIs are optimization processes of even greater complex specificity in mechanism and outcome than evolution (hence we adopted the intentional stance of ‘preference’-talk in the first place).
I agree this is often the case, because when we define ‘what is this system capable of?’ we often hold the system fixed while examining possible worlds where the environment varies in all kinds of ways. But if the possible worlds we care about all have a certain environmental feature in common—say, because we know in reality that the environmental condition obtains, and we’re trying to figure out all the ways the AI might in fact behave given different values for the variables we don’t know about with confidence—then we may, in effect, include something about the environment ‘in the AI’ for the purposes of assessing its optimization power and/or preference ordering.
For instance, we might model the AI as having the preference ‘surround the Sun with a dyson sphere’ rather than ‘conditioned on there being a Sun, surround it with a dyson sphere’; if we do the former, then the fact that that is the system’s preference depends in part on the actual existence of the Sun. Does that mean the Sun is a part of the AI’s preference encoding? Is the Sun a component of the AI? I don’t think these questions are important or interesting, so I don’t want us to be too committed to reifying AI preferences. They’re just a useful shorthand for the expected outcomes of the AI’s distinguishing features having a more large and direct causal impact on things.
Yes, agreed, for some fuzzy notion of “easily grasp” and “too complicated.” That is, there’s a sense in which thunderstorms are too complicated for me to describe in mechanistic terms why they’re having the effects they have… I certainly can’t predict those effects. But there’s also a sense in which I can describe (and even predict) the effects of a thunderstorm that feels simple, whereas I can’t do the same thing for a human being without invoking “want-speak”/intentional stance.
I’m not sure any of this is [i]justified[/i], but I agree that it is what we do… this is how we speak, and we draw these distinctions. So far, so good.
I’m not really sure what you mean by “in the AI” here, but I guess I agree that the boundary between an agent and its environment is always a fuzzy one. So, OK, I suppose we can include things about the environment “in the AI” if we choose. (I can similarly choose to include things about the environment “in myself.”) So far, so good.
Here is where you lose me again… once again you talk as though there’s simply no fact of the matter as to which preference the AI has, merely our choice as to how we model it.
But it seems to me that there are observations I can make which would provide evidence one way or the other. For example, if it has the preference ‘surround the Sun with a dyson sphere,’ then in an environment lacking the Sun I would expect it to first seek to create the Sun… how else can it implement its preferences? Whereas if it has the preference ‘conditioned on there being a Sun, surround it with a dyson sphere’; in an environment lacking the Sun I would not expect it to create the Sun.
So does the AI seek create the Sun in such an environment, or not? Surely that doesn’t depend on how I choose to model it. The AI’s preference is whatever it is, and controls its behavior. Of course, as you say, if the real world always includes a sun, then I might not be able to tell which preference the AI has. (Then again I might… the test I describe above isn’t the only test I can perform, just the first one I thought of, and other tests might not depend on the Sun’s absence.)
But whether I can tell or not doesn’t affect whether the AI has the preference or not.
Again, no. Regardless of how we model it, the system’s preference is what it is, and we can study the system (e.g., see whether it creates the Sun) to develop more accurate models of its preferences.
I agree. But I do think the question of what the AI (or, more generally, an optimizing agent) will do in various situations is interesting, and it seems to be that you’re consistently eliding over that question in ways I find puzzling.
This sounds like a potentially confusing level of simplification; a goal should be regarded as at least a way of comparing possible events.
Its behavior is what makes its goal important. But in a system designed to follow an explicitly specified goal, it does make sense to talk of its goal apart from its behavior. Even though its behavior will reflect its goal, the goal itself will reflect itself better.
If the goal is implemented as a part of the system, other parts of the system can store some information about the goal, certain summaries or inferences based on it. This information can be thought of as beliefs about the goal. And if the goal is not “logically transparent”, that is its specification is such that making concrete conclusions about what it states in particular cases is computationally expensive, then the system never knows what its goal says explicitly, it only ever has beliefs about particular aspects of the goal.
Perhaps, but I suspect that for most possible AIs there won’t always be a fact of the matter about where its preference is encoded. The blue-minimizing robot is a good example. If we treat it as a perfectly rational agent, then we might say that it has temporally stable preferences that are very complicated and conditional; or we might say that its preferences change at various times, and are partly encoded, for instance, in the properties of the color-inverting lens on its camera. An AGI’s response to environmental fluctuation will probably be vastly more complicated than a blue-minimizer’s, but the same sorts of problems arise in modeling it.
I think it’s more useful to think of rational-choice-theory-style preferences as useful theoretical constructs—like a system’s center of gravity, or its coherently extrapolated volition—than as real objects in the machine’s hardware or software. This sidesteps the problem of haggling over which exact preferences a system has, how those preferences are distributed over the environment, how to decide between causally redundant encodings which is ‘really’ the preference encoding, etc. See my response to Dave.
“Goal” is a natural idea for describing AIs with limited resources: these AIs won’t be able to make optimal decisions, and their decisions can’t be easily summarized in terms of some goal, but unlike the blue-minimizing robot they have a fixed preference ordering that doesn’t gradually drift away from what it was originally, and eventually they tend to get better at following it.
For example, if a goal is encrypted, and it takes a huge amount of computation to decrypt it, system’s behavior prior to that point won’t depend on the goal, but it’s going to work on decrypting it and eventually will follow it. This encrypted goal is probably more predictive of long-term consequences than anything else in the details of the original design, but it also doesn’t predict its behavior during the first stage (and if there is only a small probability that all resources in the universe will allow decrypting the goal, it’s probable that system’s behavior will never depend on the goal). Similarly, even if there is no explicit goal, as in the case of humans, it might be possible to work with an idealized goal that, like the encrypted goal, can’t be easily evaluated, and so won’t influence behavior for a long time.
My point is that there are natural examples where goals and the character of behavior don’t resemble each other, so that each can’t be easily inferred from the other, while both can be observed as aspects of the system. It’s useful to distinguish these ideas.
I agree preferences aren’t reducible to actual behavior. But I think they are reducible to dispositions to behave, i.e., behavior across counterfactual worlds. If a system prefers a specific event Z, that means that, across counterfactual environments you could have put it in, the future would on average have had more Z the more its specific distinguishing features had a large and direct causal impact on the world.
The examples I used seem to apply to “dispositions” to behave, in the same way (I wasn’t making this distinction). There are settings where the goal can’t be clearly inferred from behavior, or collection of hypothetical behaviors in response to various environments, at least if we keep environments relatively close to what might naturally occur, even as in those settings the goal can be observed “directly” (defined as an idealization based in AI’s design).
An AI with encypted goal (i.e. the AI itself doesn’t know the goal in explicit form, but the goal can be abstractly defined as the result of decryption) won’t behave in accordance with it in any environment that doesn’t magically let it decrypt its goal quickly, there is no tendency to push the events towards what the encrypted goal specifies, until the goal is decrypted (which might be never with high probability).
I don’t think a sufficiently well-encrypted ‘preference’ should be counted as a preference for present purposes. In principle, you can treat any physical chunk of matter as an ‘encrypted preference’, because if the AI just were a key of exactly the right shape, then it could physically interact with the lock in question to acquire a new optimization target. But if neither the AI nor anything very similar to the AI in nearby possible worlds actually acts as a key of the requisite sort, then we should treat the parts of the world that a distant AI could interact with to acquire a preference as, in our world, mere window dressing.
Perhaps if we actually built a bunch of AIs, and one of them was just like the others except where others of its kind had a preference module, it had a copy of The Wind in the Willows, we would speak of this new AI as having an ‘encrypted preference’ consisting of a book, with no easy way to treat that book as a decision criterion like its brother- and sister-AIs do for their homologous components. But I don’t see any reason right now to make our real-world usage of the word ‘preference’ correspond to that possible world’s usage. It’s too many levels of abstraction away from what we should be worried about, which are the actual real-world effects different AI architectures would have.
Here is what I mean:
Evolution was able to come up with cats. Cats are immensely complex objects. Evolution did not intend to create cats. Now consider you wanted to create an expected utility maximizer to accomplish something similar, except that it would be goal-directed, think ahead, and jump fitness gaps. Further suppose that you wanted your AI to create qucks, instead of cats. How would it do this?
Given that your AI is not supposed to search design space at random, but rather look for something particular, you would have to define what exactly qucks are. The problem is that defining what a quck is, is the hardest part. And since nobody has any idea what a quck is, nobody can design a quck creator.
The point is that thinking about the optimization of optimization is misleading, as most of the difficulty is with defining what to optimize, rather than figuring out how to optimize it. In other words, the efficiency of e.g. the scientific method depends critically on being able to formulate a specific hypothesis.
Trying to create an optimization optimizer would be akin to creating an autonomous car to find the shortest route between Gotham City and Atlantis. The problem is not how to get your AI to calculate a route, or optimize how to calculate such a route, but rather that the problem is not well-defined. You have no idea what it means to travel between two fictional cities. Which in turn means that you have no idea what optimization even means in this context, let alone meta-level optimization.
The problem is, you don’t have to program the bit that says “now make yourself more intelligent.” You only have to program the bit that says “here’s how to make a new copy of yourself, and here’s how to prove it shares your goals without running out of math.”
And the bit that says “Try things until something works, then figure out why it worked.” AKA modeling.
The AI isn’t actually an intelligence optimizer. But it notes that when it takes certain actions, it is better able to model the world, which in turn allows it to make more paperclips (or whatever). So it’ll take those actions more often.
Biological evolution is not the full picture here. Humans were programmed to be capable of winning at chess, and to care to do so, by cultural evolution, education, and environmental feedback in the form of incentives given by other people challenging them to play.
I don’t know how this works. But I do not dispute the danger of neuromorphic AIs, as you know from a comment elsewhere.
Do you suggest that from the expected behavior of neuromorphic AIs it is possible to draw conclusions about the behavior of what you call a ‘seed AI’? Would such a seed AI, as would be the case with neuromorphic AIs, be constantly programmed by environmental feedback?
What I mean is that if you program a perfect scientist but give this perfect scientist a hypothesis that does not make any predictions, then it will not be able to unfold its power.
I believe that I already wrote that I do not dispute that the idea you seem to have in mind is a risk by definition. If such an AI is likely, then we are likely going extinct if we fail at making it care about human values.
I feel uncomfortable to say this, but I do not see that the burden of proof is on me to show that it takes deliberate and intentional effort to make an AI exhibit those drives, as long that is not part of your very definition. I find the current argument in favor of AI drives to be thoroughly unconvincing.
The former has always been one of the arguments in favor of the latter in the posts I wrote on my blog.
(Note: I’m also a layman, so my non-expert opinions necessarily come with a large salt side-dish)
My guess here is that most of the “AI Drives” to self-improve, be rational, retaining it’s goal structure, etc. are considered necessary for a functional learning/self-improving algorithm. If the program cannot recognize and make rules for new patterns observed in data, make sound inferences based on known information or keep after it’s objective it will not be much of an AGI at all; it will not even be able to function as well as a modern targeted advertising program.
The rest, such as self-preservation, are justified as being logical requirements of the task. Rather than having self-preservation as a terminal value, the paperclip maximizer will value it’s own existence as an optimal means of proliferating paperclips. It makes intuitive sense that those sorts of ‘drives’ would emerge from most-any goal, but then again my intuition is not necessarily very useful for these sorts of questions.
This point might also be a source of confusion;
As Dr Valiant (great name or the greatest name?) classifies things in Probably Approximately Correct, Winning Chess would be a ‘theoryful’ task while Discovering (Interesting) Mathematical Proofs would be a ‘theoryless’ one. In essence, the theoryful has simple and well established rules for the process which could be programmed optimally in advance with little-to-no modification needed afterwards while the theoryless is complex and messy enough that an imperfect (Probably Approximately Correct) learning process would have to be employed to suss out all the rules.
Now obviously the program will benefit from labeling in it’s training data for what is and is not an “interesting” mathematical proof, otherwise it can just screw around with computationally-cheap arithmetic proofs (1 + 1 = 2, 1.1 + 1 = 2.1, 1.2 + 1 = 2.2, etc.) until the heat death of the universe. Less obviously, as the hidden tank example shows, insufficient labeling or bad labels will lead to other unintended results.
So applying that back to Friendliness; despite attempts to construct a Fun Theory, human value is currently (and may well forever remain) theoryless. A learning process whose goal is to maximize human value is going to have to be both well constructed and have very good labels initially to not be Unfriendly. Of course, it could very well correct itself later on, that is in fact at the core of a PAC algorithm, but then we get into questions of FOOM-ing and labels of human value in the environment which I am not equipped to deal with.
To explain what I have in mind, consider Ben Goertzel’s example of how to test for general intelligence:
I do not disagree that such a robot, when walking towards the classroom, if it is being obstructed by a fellow human student, could attempt to kill this human, in order to get to the classroom.
Killing a fellow human, from the perspective of the human creators of the robot, is clearly a mistake. From a human perspective, it means that the robot failed.
You believe that the robot was just following its programming/construction. Indeed, the robot is its programming. I agree with this. I agree that the human creators were mistaken about what dynamic state sequence the robot will exhibit by computing the code.
What the “dumb superintelligence” argument tries to highlight is that if humans are incapable of predicting such behavior, then they will also be mistaken about predicting behavior that is harmful to the robots power. For example, while fighting with the human in order to kill it, for a split-second it mistakes its own arm with that of the human and breaks it.
You might now argue that such a robot isn’t much of a risk. It is pretty stupid to mistake its own arm with that of the enemy it tries to kill. True. But the point is that there is no relevant difference between failing to predict behavior that will harm the robot itself, and behavior that will harm a human. Except that you might believe the former is much easier than the latter. I dispute this.
For the robot to master a complex environment, like a university full of humans, without harming itself, or decreasing the chance of achieving its goals, is already very difficult. Not stabbing or strangling other human students is not more difficult than not jumping from the 4th floor, instead of taking the stairs. This is the “dumb superintelligence” argument.
To some extent. Perhaps it would be helpful to distinguish four different kinds of defeater:
early intelligence defeater: We try to build a seed AI, but our self-rewriting AI quickly hits a wall or explodes. This is most likely if we start with a subhuman intelligence and have serious resource constraints (so we can’t, e.g., just run an evolutionary algorithm over millions of copies of the AGI until we happen upon a variant that works).
late intelligence defeater: The seed AI works just fine, but at some late stage, when it’s already at or near superintelligence, it suddenly explodes. Apparently it went down a blind alley at some point early on that led it to plateau or self-destruct later on, and neither it nor humanity is smart enough yet to figure out where exactly the problem arose. So the FOOM fizzles.
early Friendliness defeater: From the outset, the seed AI’s behavior already significantly diverges from Friendliness.
late Friendliness defeater: The seed AI starts off as a reasonable approximation of Friendliness, but as it approaches superintelligence its values diverge from anything we’d consider Friendly, either because it wasn’t previously smart enough to figure out how to self-modify while keeping its values stable, or because it was never perfectly Friendly and the new circumstances its power puts it in now make the imperfections much more glaring.
In general, late defeaters are much harder for humans to understand than early defeaters, because an AI undergoing FOOM is too fast and complex to be readily understood. Your three main arguments, if I’m understanding them, have been:
(a) Early intelligence defeaters are so numerous that there’s no point thinking much about other kinds of defeaters yet.
(b) Friendliness defeaters imply a level of incompetence on the programmers’ part that strongly suggest intelligence defeaters will arise in the same situation.
(c) If an initially somewhat-smart AI is smart enough to foresee and avoid late intelligence defeaters, then an initially somewhat-nice AI should be smart enough to foresee and avoid late Friendliness defeaters.
I reject (a), because I haven’t seen any specific reason a self-improving AGI will be particularly difficult to make FOOM—‘it would require lots of complicated things to happen’ is very nearly a fully general argument against any novel technology, so I can’t get very far on that point alone. I accept (b), at least for a lot of early defeaters. But my concern is that while non-Friendliness predicts non-intelligence (and non-intelligence predicts non-Friendliness), intelligence also predicts non-Friendliness.
But our interesting disagreement seems to be over (c). Interesting because it illuminates general differences between the basic idea of a domain-general optimization process (intelligence) and the (not-so-)basic idea of Everything Humans Like. One important difference is that if an AGI optimizes for anything, it will have strong reason to steer clear of possible late intelligence defeaters. Late Friendliness defeaters, on the other hand, won’t scare optimization-process-optimizers in general.
It’s easy to see in advance that most beings that lack obvious early Friendliness defeaters will nonetheless have late Friendliness defeaters. In contrast, it’s much less clear that most beings lacking early intelligence defeaters will have late intelligence defeaters. That’s extremely speculative at this point—we simply don’t know what sorts of intelligence-destroying attractors might exist out there, or what sorts of paradoxes and complications are difficult v. trivial to overcome.
But, once again, it doesn’t take any stupidity on the AI’s part to disvalue physically injuring a human, even if it does take stupidity to not understand that one is physically injuring a human. It only takes a different value system. Valuing one’s own survival is not orthogonal to valuing becoming more intelligent; but valuing human survival is orthogonal to valuing becoming more intelligent. (Indeed, to the extent they aren’t orthogonal it’s because valuing becoming more intelligent tends to imply disvaluing human survival, because humans are hard to control and made of atoms that can be used for other ends, including increased computing power.) This is the whole point of the article we’re commenting on.
Here is part of my stance towards AI risks:
1. I assign a negligible probability to the possibility of a sudden transition from well-behaved narrow AIs to general AIs (see below).
2. An AI will not be pulled at random from mind design space. An AI will be the result of a research and development process. A new generation of AIs will need to be better than other products at “Understand What Humans Mean” and “Do What Humans Mean”, in order to survive the research phase and subsequent market pressure.
3. Commercial, research or military products are created with efficiency in mind. An AI that was prone to take unbounded actions given any terminal goal would either be fixed or abandoned during the early stages of research. If early stages showed that inputs such as the natural language query would yield results such as then the AI would never reach a stage in which it was sufficiently clever and trained to understand what results would satisfy its creators in order to deceive them.
4. I assign a negligible probability to the possibility of a consequentialist AI / expected utility maximizer / approximation to AIXI.
Given that the kind of AIs from point 4 are possible:
5. Omohundro’s AI drives are what make the kind of AIs mentioned in point 1 dangerous. Making an AI that does not exhibit these drives in an unbounded manner is probably a prerequisite to get an AI to work at all (there are not enough resources to think about being obstructed by simulator gods etc.), or should otherwise be easy compared to the general difficulties involved in making an AI work using limited resources.
6. An AI from point 4 will only ever do what it has been explicitly programmed to do. Such an AI is not going to protect its utility-function, acquire resources or preemptively eliminate obstacles in an unbounded fashion. Because it is not intrinsically rational to do so. What specifically constitutes rational, economic behavior is inseparable with an agent’s terminal goal. That any terminal goal can be realized in an infinite number of ways implies an infinite number of instrumental goals to choose from.
7. Unintended consequences are by definition not intended. They are not intelligently designed but detrimental side effects, failures. Whereas intended consequences, such as acting intelligently, are intelligently designed. If software was not constantly improved to be better at doing what humans intend it to do we would never be able to reach a level of sophistication where a software could work well enough to outsmart us. To do so it would have to work as intended along a huge number of dimensions. For an AI to constitute a risk as a result of unintended consequences those unintended consequences would have to have no, or little, negative influence on the huge number of intended consequences that are necessary for it to be able to overpower humanity.
I am not yet at a point of my education where I can say with confidence that this is the wrong way to think, but I do believe it is.
If someone walked up to you and told you about a risk only he can solve, and that you should therefore give this person money, would you give him money because you do not see any specific reason for why he could be wrong? Personally I would perceive the burden of proof to be on him to show me that the risk is real.
Despite this, I have specific reasons to personally believe that the kind of AI you have in mind is impossible. I have thought about such concepts as consequentialism / expected utility maximization, and do not see that they could be made to work, other than under very limited circumstances. And I also asked other people, outside of LessWrong, who are more educated and smarter than me, and they also told me that these kind of AIs are not feasible, they are uncomputable.
I am not sure I understand what you mean by c. I don’t think I agree with it.
I don’t know what this means.
That this black box you call “intelligence” might be useful to achieve a lot of goals is not an argument in support of humans wanting to and succeeding at the implementation of “value to maximize intelligence” in conjunction with “by all means”.
Most definitions of intelligence that I am aware of are in terms of the ability to achieve goals. Saying that a system values to become more intelligent then just means that a system values to increase its ability to achieve its goals. In this context, what you suggest is that humans will want to, and will succeed to, implement an AI that in order to beat humans at Tic-tac-toe is first going to take over the universe and make itself capable of building such things as Dyson spheres.
What I am saying is that it is much easier to create a Tic-tac-toe playing AI, or an AI that can earn a university degree, than the former in conjunction with being able to take over the universe and build Dyson spheres.
The argument that valuing not to kill humans is orthogonal to taking over the universe and building Dyson spheres is completely irrelevant.
I don’t think anyone’s ever disputed this. (However, that’s not very useful if the deterministic process resulting in the SI is too complex for humans to distinguish it in advance from the outcome of a random walk.)
Agreed. But by default, a machine that is better than other rival machines at satisfying our short-term desires will not satisfy our long-term desires. The concern isn’t that we’ll suddenly start building AIs with the express purpose of hitting humans in the face with mallets. The concern is that we’ll code for short-term rather than long-term goals, due to a mixture of disinterest in Friendliness and incompetence at Friendliness. But if intelligence explosion occurs, ‘the long run’ will arrive very suddenly, and very soon. So we need to adjust our research priorities to more seriously assess and modulate the long-term consequences of our technology.
That may be a reason to think that recursively self-improving AGI won’t occur. But it’s not a reason to expect such AGI, if it occurs, to be Friendly.
The seed is not the superintelligence. We shouldn’t expect the seed to automatically know whether the superintelligence will be Friendly, any more than we should expect humans to automatically know whether the superintelligence will be Friendly.
I’m not following. Why does an AGI have to have a halting condition (specifically, one that actually occurs at some point) in order to be able to productively rewrite its own source code?
You don’t seem to be internalizing my arguments. This is just the restatement of a claim I pointed out was not just wrong but dishonestly stated here.
Sure, but the list of instrumental goals overlap more than the list of terminal goals, because energy from one project can be converted to energy for a different project. This is an empirical discovery about our world; we could have found ourselves in the sort of universe where instrumental goals don’t converge that much, e.g., because once energy’s been locked down into organisms or computer chips you just Can’t convert it into useful work for anything else. In a world where we couldn’t interfere with the AI’s alien goals, nor could our component parts and resources be harvested to build very different structures, nor could we be modified to work for the AI, the UFAI would just ignore us and zip off into space to try and find more useful objects. We don’t live in that world because complicated things can be broken down into simpler things at a net gain in our world, and humans value a specific set of complicated things.
‘These two sets are both infinite’ does not imply ‘we can’t reason about these two things’ relative size, or how often the same elements recur in their elements’.
You’ve spent an awful lot of time writing about the varied ways in which you’ve not yet been convinced by claims you haven’t put much time into actively investigating. Maybe some of that time could be better spent researching these topics you keep writing about? I’m not saying to stop talking about this, but there’s plenty of material on a lot of these issues to be found. Have you read Intelligence Explosion Microeconomics?
http://wiki.lesswrong.com/wiki/Optimization_process
As a rule, adding halting conditions adds complexity to an algorithm, rather than removing complexity.
No, this is a serious misunderstanding. Yudkowsky’s definition of ‘intelligence’ is about the ability to achieve goals in general, not about the ability to achieve the system’s goals. That’s why you can’t increase a system’s intelligence by lowering its standards, i.e., making its preferences easier to satisfy.
Straw-man; no one has claimed that humans are likely to want to create an UFAI. What we’ve suggested is that humans are likely to want to create an algorithm, X, that will turn out to be a UFAI. (In other words, the fallacy you’re committing is confusing intension with extension.)
That aside: Are you saying Dyson spheres wouldn’t be useful for beating more humans at more tic-tac-toe games? Seems like a pretty good way to win at tic-tac-toe to me.
Actually I do define intelligence as ability to hit a narrow outcome target relative to your own goals, but if your goals are very relaxed then the volume of outcome space with equal or greater utility will be very large. However one would expect that many of the processes involved in hitting a narrow target in outcome space (such that few other outcomes are rated equal or greater in the agent’s preference ordering), such as building a good epistemic model or running on a fast computer, would generalize across many utility functions; this is why we can speak of properties apt to intelligence apart from particular utility functions.
Hmm. But this just sounds like optimization power to me. You’ve defined intelligence in the past as “efficient cross-domain optimization”. The “cross-domain” part I’ve taken to mean that you’re able to hit narrow targets in general, not just ones you happen to like. So you can become more intelligent by being better at hitting targets you hate, or by being better at hitting targets you like.
The former are harder to test, but something you’d hate doing now could become instrumentally useful to know how to do later. And your intelligence level doesn’t change when the circumstance shifts which part of your skillset is instrumentally useful. For that matter, I’m missing why it’s useful to think that your intelligence level could drastically shift if your abilities remained constant but your terminal values were shifted. (E.g., if you became pickier.)
No, “cross-domain” means that I can optimize across instrumental domains. Like, I can figure out how to go through water, air, or space if that’s the fastest way to my destination, I am not limited to land like a ground sloth.
Measured intelligence shouldn’t shift if you become pickier—if you could previously hit a point such that only 1/1000th of the space was more preferred than it, we’d still expect you to hit around that narrow a volume of the space given your intelligence even if you claimed afterward that a point like that only corresponded to 0.25 utility on your 0-1 scale instead of 0.75 utility due to being pickier ([expected] utilities sloping more sharply downward with increasing distance from the optimum).
You might be not aware of this but I wrote a sequence of short blog posts where I tried to think of concrete scenarios that could lead to human extinction. Each of which raised many questions.
The introductory post is ‘AI vs. humanity and the lack of concrete scenarios’.
1. Questions regarding the nanotechnology-AI-risk conjunction
2. AI risk scenario: Deceptive long-term replacement of the human workforce
3. AI risk scenario: Social engineering
4. AI risk scenario: Elite Cabal
5. AI risk scenario: Insect-sized drones
6. AI risks scenario: Biological warfare
What might seem to appear completely obvious to you for reasons that I do not understand, e.g. that an AI can take over the world, appears to me largely like magic (I am not trying to be rude, by magic I only mean that I don’t understand the details). At the very least there are a lot of open questions. Even given that for the sake of the above posts I accepted that the AI is superhuman and can do such things as deceive humans by its superior knowledge of human psychology. Which seems to be non-trivial assumption, to say the least.
Over and over I told you that given all your assumptions, I agree that AGI is an existential risk.
You did not reply to my argument. My argument was that if the seed is unfriendly then it will not be smart enough to hide its unfriendliness. My argument did not pertain the possibility of a friendly seed turning unfriendly.
What I have been arguing is that an AI should not be expected, by default, to want to eliminate all possible obstructions. There are many graduations here. That, by some economic or otherwise theoretic argument, it might be instrumentally rational for some ideal AI to take over the world, does not mean that humans would create such an AI, or that an AI could not be limited to care about fires in its server farm rather than that Russia might nuke the U.S. and thereby destroy its servers.
Did you mean to reply to another point? I don’t see how the reply you linked to is relevant to what I wrote.
My argument is that an AI does not need to consider all possible threats and care to acquire all possible resources. Based on its design it could just want to optimize using its initial resources while only considering mundane threats. I just don’t see real-world AIs to conclude that they need to take over the world. I don’t think an AI is likely going to be designed that way. I also don’t think such an AI could work, because such inferences would require enormous amounts of resources.
I have done what is possible given my current level of education and what I perceive to be useful. I have e.g. asked experts about their opinion.
A few general remarks about the kind of papers such as the one that you linked to.
How much should I update towards MIRI’s position if I (1) understood the arguments in the paper (2) found the arguments convincing?
My answer is the following. If the paper was about the abc conjecture, the P versus NP problem, climate change, or even such mundane topics as psychology, I would either not be able to understand the paper, would be unable to verify the claims, or would have very little confidence in my judgement.
So what about ‘Intelligence Explosion Microeconomics’? That I can read most of it is only due to the fact that it is very informally written. The topic itself is more difficult and complex than all of the above mentioned problems together. Yet the arguments in support of it, to exaggerate a little bit, contain less rigor than the abstract of one of Shinichi Mochizuki’s papers on the abc conjecture.
Which means that my answer is that I should update very little towards MIRI’s position and that any confidence I gain about MIRI’s position is probably highly unreliable.
Thanks. My feeling is that to gain any confidence into what all this technically means, and to answer all the questions this raises, I’d probably need about 20 years of study.
Here is part of a post exemplifying how I understand the relation between goals and intelligence:
If a goal has very few constraints then the set that satisfies all constraints is very large. A vague and ambiguous goal allows for too much freedom in the sense that a wide range of world states would have the same expected value and therefore imply a very large solution space, since a wide range of AI’s will be able to achieve those world states and thereby satisfy the condition of being improved versions of their predecessor.
This means that in order to get an AI to become superhuman at all, and very quickly in particular, you will need to encode a very specific goal against which mistakes, optimization power and achievement can be judged.
It is really hard to communicate how I perceive this and other discussions about MIRI’s position without offending people, or killing the discussion.
I am saying this in full honesty. The position you appear to support seems so utterly “complex” (far-fetched) that the current arguments are unconvincing.
Here is my perception of the scenario that you try to sell me (exaggerated to make a point). I have a million questions about it that I can’t answer and which your answers either sidestep or explain away by using “magic”.
At this point I probably made 90% of the people reading this comment incredible angry. My perception is that you cannot communicate this perception on LessWrong without getting into serious trouble. That’s also what I meant when I told you that I cannot be completely honest if you want to discuss this on LessWrong.
I can also assure you that many people who are much smarter and higher status than me think so as well. Many people communicated the absurdity of all this to me but told me that they would not repeat this in public.
Pretending to be friendly when you’re actually not is something that doesn’t even require human level intelligence. You could even do it accidentally.
In general, the appearance of Friendliness at low levels of ability to influence the world doesn’t guarantee actual Friendliness at high levels of ability to influence the world. (If it did, elected politicians would be much higher quality.)
But it will scare friendly ones, which will want to keep their values stable.
It takes stupidity to misinterpret friendlienss.
Yes. If an AI is Friendly at one stage, then it is Friendly at every subsequent stage. This doesn’t help make almost-Friendly AIs become genuinely Friendly, though.
Yes, but that’s stupidity on the part of the human programmer, and/or on the part of the seed AI if we ask it for advice. The superintelligence didn’t write its own utility function; the superintelligence may well understand Friendliness perfectly, but that doesn’t matter if it hasn’t been programmed to rewrite its source code to reflect its best understanding of ‘Friendliness’. The seed is not the superintelligence. See: http://lesswrong.com/lw/igf/the_genie_knows_but_doesnt_care/
That depends on the architecture. In a Loosemore architecture, the AI interprets high-level directives itself, so if it gets them wrong, that’s it’s mistake.
… and whose fault is that?
http://lesswrong.com/lw/rf/ghosts_in_the_machine/
Say we find an algorithm for producing progressively more accurate beliefs about itself and the world. This algorithm may be long and complicated—perhaps augmented by rules-of-thumb whenever the evidence available to it says these rules make better predictions. (E.g, “nine times out of ten the Enterprise is not destroyed.”) Combine this with an arbitrary goal and we have the making of a seed AI.
Seems like this could straightforwardly improve its ability to predict humans without changing its goal, which may be ‘maximize pleasure’ or ‘maximize X’. Why would it need to change its goal?
If you deny the possibility of the above algorithm, then before giving any habitual response please remember what humanity knows about clinical vs. actuarial judgment. What lesson do you take from this?