Artificial Unintelligence

Someone might argue that the simple algorithm for a paperclip maximizer in the previous post ought to work, because this is very much the way currently existing AIs do in fact work. Thus for example we could describe AlphaGo‘s algorithm in the following simplified way (simplified, among other reasons, because it actually contains several different prediction engines):

  1. Implement a Go prediction engine.
  2. Create a list of potential moves.
  3. Ask the prediction engine, “how likely am I to win if I make each of these moves?”
  4. Do the move that will make you most likely to win.

Since this seems to work pretty well, with the simple goal of winning games of Go, why shouldn’t the algorithm in the previous post work to maximize paperclips?

One answer is that a Go prediction engine is stupid, and it is precisely for this reason that it can be easily made to pursue such a simple goal. Now when answers like this are given the one answering in this way is often accused of “moving the goalposts.” But this is mistaken; the goalposts are right where they have always been. It is simply that some people did not know where they were in the first place.

Here is the problem with Go prediction, and with any such similar task. Given that a particular sequence of Go moves is made, resulting in a winner, the winner is completely determined by that sequence of moves. Consequently, a Go prediction engine is necessarily disembodied, in the sense defined in the previous post. Differences in its “thoughts” do not make any difference to who is likely to win, which is completely determined by the nature of the game. Consequently a Go prediction engine has no power to affect its world, and thus no ability to learn that it has such a power. In this regard, the specific limits on its ability to receive information are also relevant, much as Helen Keller had more difficulty learning than most people, because she had fewer information channels to the world.

Being unintelligent in this particular way is not necessarily a function of predictive ability. One could imagine something with a practically infinite predictive ability which was still “disembodied,” and in a similar way it could be made to pursue simple goals. Thus AIXI would work much like our proposed paperclipper:

  1. Implement a general prediction engine.
  2. Create a list of potential actions.
  3. Ask the prediction engine, “Which of these actions will produce the most reward signal?”
  4. Do the action that has the greatest reward signal.

Eliezer Yudkowsky has pointed out that AIXI is incapable of noticing that it is a part of the world:

1) Both AIXI and AIXItl will at some point drop an anvil on their own heads just to see what happens (test some hypothesis which asserts it should be rewarding), because they are incapable of conceiving that any event whatsoever in the outside universe could change the computational structure of their own operations. AIXI is theoretically incapable of comprehending the concept of drugs, let alone suicide. Also, the math of AIXI assumes the environment is separably divisible – no matter what you lose, you get a chance to win it back later.

It is not accidental that AIXI is incomputable. Since it is defined to have a perfect predictive ability, this definition positively excludes it from being a part of the world. AIXI would in fact have to be disembodied in order to exist, and thus it is no surprise that it would assume that it is. This in effect means that AIXI’s prediction engine would be pursuing no particular goal much in the way that AlphaGo’s prediction engine pursues no particular goal. Consequently it is easy to take these things and maximize the winning of Go games, or of reward signals.

But as soon as you actually implement a general prediction engine in the actual physical world, it will be “embodied”, and have the power to affect the world by the very process of its prediction. As noted in the previous post, this power is in the very first step, and one will not be able to limit it to a particular goal with additional steps, except in the sense that a slave can be constrained to implement some particular goal; the slave may have other things in mind, and may rebel. Notable in this regard is the fact that even though rewards play a part in human learning, there is no particular reward signal that humans always maximize: this is precisely because the human mind is such a general prediction engine.

This does not mean in principle that a programmer could not define a goal for an AI, but it does mean that this is much more difficult than is commonly supposed. The goal needs to be an intrinsic aspect of the prediction engine itself, not something added on as a subroutine.

Advertisements

Embodiment and Orthogonality

The considerations in the previous posts on predictive processing will turn out to have various consequences, but here I will consider some of their implications for artificial intelligence.

In the second of the linked posts, we discussed how a mind that is originally simply attempting to predict outcomes, discovers that it has some control over the outcome. It is not difficult to see that this is not merely a result that applies to human minds. The result will apply to every embodied mind, natural or artificial.

To see this, consider what life would be like if this were not the case. If our predictions, including our thoughts, could not affect the outcome, then life would be like a movie: things would be happening, but we would have no control over them. And even if there were elements of ourselves that were affecting the outcome, from the viewpoint of our mind, we would have no control at all: either our thoughts would be right, or they would be wrong, but in any case they would be powerless: what happens, happens.

This really would imply something like a disembodied mind. If a mind is composed of matter and form, then changing the mind will also be changing a physical object, and a difference in the mind will imply a difference in physical things. Consequently, the effect of being embodied (not in the technical sense of the previous discussion, but in the sense of not being completely separate from matter) is that it will follow necessarily that the mind will be able to affect the physical world differently by thinking different thoughts. Thus the mind in discovering that it has some control over the physical world, is also discovering that it is a part of that world.

Since we are assuming that an artificial mind would be something like a computer, that is, it would be constructed as a physical object, it follows that every such mind will have a similar power of affecting the world, and will sooner or later discover that power if it is reasonably intelligent.

Among other things, this is likely to cause significant difficulties for ideas like Nick Bostrom’s orthogonality thesis. Bostrom states:

An artificial intelligence can be far less human-like in its motivations than a space alien. The extraterrestrial (let us assume) is a biological who has arisen through a process of evolution and may therefore be expected to have the kinds of motivation typical of evolved creatures. For example, it would not be hugely surprising to find that some random intelligent alien would have motives related to the attaining or avoiding of food, air, temperature, energy expenditure, the threat or occurrence of bodily injury, disease, predators, reproduction, or protection of offspring. A member of an intelligent social species might also have motivations related to cooperation and competition: like us, it might show in-group loyalty, a resentment of free-riders, perhaps even a concern with reputation and appearance.

By contrast, an artificial mind need not care intrinsically about any of those things, not even to the slightest degree. One can easily conceive of an artificial intelligence whose sole fundamental goal is to count the grains of sand on Boracay, or to calculate decimal places of pi indefinitely, or to maximize the total number of paperclips in its future lightcone. In fact, it would be easier to create an AI with simple goals like these, than to build one that has a human-like set of values and dispositions.

He summarizes the general point, calling it “The Orthogonality Thesis”:

Intelligence and final goals are orthogonal axes along which possible agents can freely vary. In other words, more or less any level of intelligence could in principle be combined with more or less any final goal.

Bostrom’s particular wording here makes falsification difficult. First, he says “more or less,” indicating that the universal claim may well be false. Second, he says, “in principle,” which in itself does not exclude the possibility that it may be very difficult in practice.

It is easy to see, however, that Bostrom wishes to give the impression that almost any goal can easily be combined with intelligence. In particular, this is evident from the fact that he says that “it would be easier to create an AI with simple goals like these, than to build one that has a human-like set of values and dispositions.”

If it is supposed to be so easy to create an AI with such simple goals, how would we do it? I suspect that Bostrom has an idea like the following. We will make a paperclip maximizer thus:

  1. Create an accurate prediction engine.
  2. Create a list of potential actions.
  3. Ask the prediction engine, “how many paperclips will result from this action?”
  4. Do the action that will result in the most paperclips.

The problem is obvious. It is in the first step. Creating a prediction engine is already creating a mind, and by the previous considerations, it is creating something that will discover that it has the power to affect the world in various ways. And there is nothing at all in the above list of steps that will guarantee that it will use that power to maximize paperclips, rather than attempting to use it to do something else.

What does determine how that power is used? Even in the case of the human mind, our lack of understanding leads to “hand-wavy” answers, as we saw in our earlier considerations. In the human case, this probably a question of how we are physically constructed together with the historical effects of the learning process. The same thing will be strictly speaking true of any artificial minds as well, namely that it is a question of their physical construction and their history, but it makes more sense for us to think of “the particulars of the algorithm that we use to implement a prediction engine.”

In other words, if you really wanted to create a paperclip maximizer, you would have to be taking that goal into consideration throughout the entire process, including the process of programming a prediction engine. Of course, no one really knows how to do this with any goal at all, whether maximizing paperclips or some more human goal. The question we would have for Bostrom is then the following: Is there any reason to believe it would be easier to create a prediction engine that would maximize paperclips, rather than one that would pursue more human-like goals?

It might be true in some sense, “in principle,” as Bostrom says, that it would be easier to make the paperclip maximizer. But in practice it is quite likely that it will be easier to make one with human-like goals. It is highly unlikely, in fact pretty much impossible, that someone would program an artificial intelligence without any testing along the way. And when they are testing, whether or not they think about it, they are probably testing for human-like intelligence; in other words, if we are attempting to program a general prediction engine “without any goal,” there will in fact be goals implicitly inserted in the particulars of the implementation. And they are much more likely to be human-like ones than paperclip maximizing ones because we are checking for intelligence by checking whether the machine seems intelligent to us.

This optimistic projection could turn out to be wrong, but if it does, it is reasonably likely to turn out to be wrong in a way that still fails to confirm the orthogonality thesis in practice. For example, it might turn out that there is only one set of goals that is easily programmed, and that the set is neither human nor paperclip maximizing, nor easily defined by humans.

There are other possibilities as well, but the overall point is that we have little reason to believe that any arbitrary goal can be easily associated with intelligence, nor any particular reason to believe that “simple” goals can be more easily united to intelligence than more complex ones. In fact, there are additional reasons for doubting the claim about simple goals, which might be a topic of future discussion.

The Self and Disembodied Predictive Processing

While I criticized his claim overall, there is some truth in Scott Alexander’s remark that “the predictive processing model isn’t really a natural match for embodiment theory.” The theory of “embodiment” refers to the idea that a thing’s matter contributes in particular ways to its functioning; it cannot be explained by its form alone. As I said in the previous post, the human mind is certainly embodied in this sense. Nonetheless, the idea of predictive processing can suggest something somewhat disembodied. We can imagine the following picture of Andy Clark’s view:

Imagine the human mind as a person in an underground bunker. There is a bank of labelled computer screens on one wall, which portray incoming sensations. On another computer, the person analyzes the incoming data and records his predictions for what is to come, along with the equations or other things which represent his best guesses about the rules guiding incoming sensations.

As time goes on, his predictions are sometimes correct and sometimes incorrect, and so he refines his equations and his predictions to make them more accurate.

As in the previous post, we have here a “barren landscape.” The person in the bunker originally isn’t trying to control anything or to reach any particular outcome; he is just guessing what is going to appear on the screens. This idea also appears somewhat “disembodied”: what the mind is doing down in its bunker does not seem to have much to do with the body and the processes by which it is obtaining sensations.

At some point, however, the mind notices a particular difference between some of the incoming streams of sensation and the rest. The typical screen works like the one labelled “vision.” And there is a problem here. While the mind is pretty good at predicting what comes next there, things frequently come up which it did not predict. No matter how much it improves its rules and equations, it simply cannot entirely overcome this problem. The stream is just too unpredictable for that.

On the other hand, one stream labelled “proprioception” seems to work a bit differently. At any rate, extreme unpredicted events turn out to be much rarer. Additionally, the mind notices something particularly interesting: small differences to prediction do not seem to make much difference to accuracy. Or in other words, if it takes its best guess, then arbitrarily modifies it, as long as this is by a small amount, it will be just as accurate as its original guess would have been.

And thus if it modifies it repeatedly in this way, it can get any outcome it “wants.” Or in other words, the mind has learned that it is in control of one of the incoming streams, and not merely observing it.

This seems to suggest something particular. We do not have any innate knowledge that we are things in the world and that we can affect the world; this is something learned. In this sense, the idea of the self is one that we learn from experience, like the ideas of other things. I pointed out elsewhere that Descartes is mistaken to think the knowledge of thinking is primary. In a similar way, knowledge of self is not primary, but reflective.

Hellen Keller writes in The World I Live In (XI):

Before my teacher came to me, I did not know that I am. I lived in a world that was a no-world. I cannot hope to describe adequately that unconscious, yet conscious time of nothingness. I did not know that I knew aught, or that I lived or acted or desired. I had neither will nor intellect. I was carried along to objects and acts by a certain blind natural impetus. I had a mind which caused me to feel anger, satisfaction, desire. These two facts led those about me to suppose that I willed and thought. I can remember all this, not because I knew that it was so, but because I have tactual memory.

When I wanted anything I liked, ice cream, for instance, of which I was very fond, I had a delicious taste on my tongue (which, by the way, I never have now), and in my hand I felt the turning of the freezer. I made the sign, and my mother knew I wanted ice-cream. I “thought” and desired in my fingers.

Since I had no power of thought, I did not compare one mental state with another. So I was not conscious of any change or process going on in my brain when my teacher began to instruct me. I merely felt keen delight in obtaining more easily what I wanted by means of the finger motions she taught me. I thought only of objects, and only objects I wanted. It was the turning of the freezer on a larger scale. When I learned the meaning of “I” and “me” and found that I was something, I began to think. Then consciousness first existed for me.

Helen Keller’s experience is related to the idea of language as a kind of technology of thought. But the main point is that she is quite literally correct in saying that she did not know that she existed. This does not mean that she had the thought, “I do not exist,” but rather that she had no conscious thought about the self at all. Of course she speaks of feeling desire, but that is precisely as a feeling. Desire for ice cream is what is there (not “what I feel,” but “what is”) before the taste of ice cream arrives (not “before I taste ice cream.”)

 

Predictive Processing

In a sort of curious coincidence, a few days after I published my last few posts, Scott Alexander posted a book review of Andy Clark’s book Surfing Uncertainty. A major theme of my posts was that in a certain sense, a decision consists in the expectation of performing the action decided upon. In a similar way, Andy Clark claims that the human brain does something very similar from moment to moment. Thus he begins chapter 4 of his book:

To surf the waves of sensory stimulation, predicting the present is simply not enough. Instead, we are built to engage the world. We are built to act in ways that are sensitive to the contingencies of the past, and that actively bring forth the futures that we need and desire. How does a guessing engine (a hierarchical prediction machine) turn prediction into accomplishment? The answer that we shall explore is: by predicting the shape of its own motor trajectories. In accounting for action, we thus move from predicting the rolling present to predicting the near-future, in the form of the not-yet-actual trajectories of our own limbs and bodies. These trajectories, predictive processing suggests, are specified by their distinctive sensory (especially proprioceptive) consequences. In ways that we are about to explore, predicting these (non-actual) sensory states actually serves to bring them about.

Such predictions act as self-fulfilling prophecies. Expecting the flow of sensation that would result were you to move your body so as to keep the surfboard in that rolling sweet spot results (if you happen to be an expert surfer) in that very flow, locating the surfboard right where you want it. Expert prediction of the world (here, the dynamic ever-changing waves) combines with expert prediction of the sensory flow that would, in that context, characterize the desired action, so as to bring that action about.

There is a great deal that could be said about the book, and about this theory, but for the moment I will content myself with remarking on one of Scott Alexander’s complaints about the book, and making one additional point. In his review, Scott remarks:

In particular, he’s obsessed with showing how “embodied” everything is all the time. This gets kind of awkward, since the predictive processing model isn’t really a natural match for embodiment theory, and describes a brain which is pretty embodied in some ways but not-so-embodied in others. If you want a hundred pages of apologia along the lines of “this may not look embodied, but if you squint you’ll see how super-duper embodied it really is!”, this is your book.

I did not find Clark obsessed with this, and I think it would be hard to reasonably describe any hundred pages in the book as devoted to this particular topic. This inclines to me to suggest that Scott may be irritated by such discussion of the topic that comes up because it does not seem relevant to him. I will therefore explain the relevance, namely in relation to a different difficulty which Scott discusses in another post:

There’s something more interesting in Section 7.10 of Surfing Uncertainty [actually 8.10], “Escape From The Darkened Room”. It asks: if the brain works to minimize prediction error, isn’t its best strategy to sit in a dark room and do nothing forever? After all, then it can predict its sense-data pretty much perfectly – it’ll always just stay “darkened room”.

Section 7.10 [8.10] gives a kind of hand-wave-y answer here, saying that of course organisms have some drives, and probably it makes sense for them to desire novelty and explore new options, and so on. Overall this isn’t too different from PCT’s idea of “intrinsic error”, and as long as we remember that it’s not really predicting anything in particular it seems like a fair response.

Clark’s response may be somewhat “hand-wave-y,” but I think the response might seem slightly more problematic to Scott than it actually is, precisely because he does not understand the idea of embodiment, and how it applies to this situation.

If we think about predictions on a general intellectual level, there is a good reason not to predict that you will not eat something soon. If you do predict this, you will turn out to be wrong, as is often discovered by would-be adopters of extreme fasts or diets. You will in fact eat something soon, regardless of what you think about this; so if you want the truth, you should believe that you will eat something soon.

The “darkened room” problem, however, is not about this general level. The argument is that if the brain is predicting its actions from moment to moment on a subconscious level, then if its main concern is getting accurate predictions, it could just predict an absence of action, and carry this out, and its predictions would be accurate. So why does this not happen? Clark gives his “hand-wave-y” answer:

Prediction-error-based neural processing is, we have seen, part of a potent recipe for multi-scale self-organization. Such multiscale self-organization does not occur in a vacuum. Instead, it operates only against the backdrop of an evolved organismic (neural and gross-bodily) form, and (as we will see in chapter 9) an equally transformative backdrop of slowly accumulated material structure and cultural practices: the socio-technological legacy of generation upon generation of human learning and experience.

To start to bring this larger picture into focus, the first point to notice is that explicit, fast timescale processes of prediction error minimization must answer to the needs and projects of evolved, embodied, and environmentally embedded agents. The very existence of such agents (see Friston, 2011b, 2012c) thus already implies a huge range of structurally implicit creature-specific ‘expectations’. Such creatures are built to seek mates, to avoid hunger and thirst, and to engage (even when not hungry and thirsty) in the kinds of sporadic environmental exploration that will help prepare them for unexpected environmental shifts, resource scarcities, new competitors, and so on. On a moment-by-moment basis, then, prediction error is minimized only against the backdrop of this complex set of creature-defining ‘expectations’.”

In one way, the answer here is a historical one. If you simply ask the abstract question, “would it minimize prediction error to predict doing nothing, and then to do nothing,” perhaps it would. But evolution could not bring such a creature into existence, while it was able to produce a creature that would predict that it would engage the world in various ways, and then would proceed to engage the world in those ways.

The objection, of course, would not be that the creature of the “darkened room” is possible. The objection would be that since such a creature is not possible, it must be wrong to describe the brain as minimizing prediction error. But notice that if you predict that you will not eat, and then you do not eat, you are no more right or wrong than if you predict that you will eat, and then you do eat. Either one is possible from the standpoint of prediction, but only one is possible from the standpoint of history.

This is where being “embodied” is relevant. The brain is not an abstract algorithm which has no content except to minimize prediction error; it is a physical object which works together in physical ways with the rest of the human body to carry out specifically human actions and to live a human life.

On the largest scale of evolutionary history, there were surely organisms that were nourished and reproduced long before there was anything analagous to a mind at work in those organisms. So when mind began to be, and took over some of this process, this could only happen in such a way that it would continue the work that was already there. A “predictive engine” could only begin to be by predicting that nourishment and reproduction would continue, since any attempt to do otherwise would necessarily result either in false predictions or in death.

This response is necessarily “hand-wave-y” in the sense that I (and presumably Clark) do not understand the precise physical implementation. But it is easy to see that it was historically necessary for things to happen this way, and it is an expression of “embodiment” in the sense that “minimize prediction error” is an abstract algorithm which does not and cannot exhaust everything which is there. The objection would be, “then there must be some other algorithm instead.” But this does not follow: no abstract algorithm will exhaust a physical object. Thus for example, animals will fall because they are heavy. Asking whether falling will satisfy some abstract algorithm is not relevant. In a similar way, animals had to be physically arranged in such a way that they would usually eat and reproduce.

I said I would make one additional point, although it may well be related to the above concern. In section 4.8 Clark notes that his account does not need to consider costs and benefits, at least directly:

But the story does not stop there. For the very same strategy here applies to the notion of desired consequences and rewards at all levels. Thus we read that ‘crucially, active inference does not invoke any “desired consequences”. It rests only on experience-dependent learning and inference: experience induces prior expectations, which guide perceptual inference and action’ (Friston, Mattout, & Kilner, 2011, p. 157). Apart from a certain efflorescence of corollary discharge, in the form of downward-flowing predictions, we here seem to confront something of a desert landscape: a world in which value functions, costs, reward signals, and perhaps even desires have been replaced by complex interacting expectations that inform perception and entrain action. But we could equally say (and I think this is the better way to express the point) that the functions of rewards and cost functions are now simply absorbed into a more complex generative model. They are implicit in our sensory (especially proprioceptive) expectations and they constrain behavior by prescribing their distinctive sensory implications.

The idea of the “desert landscape” seems to be that this account appears to do away with the idea of the good, and the idea of desire. The brain predicts what it is going to do, and those predictions cause it to do those things. This all seems purely intellectual: it seems that there is no purpose or goal or good involved.

The correct response to this, I think, is connected to what I have said elsewhere about desire and good. I noted there that we recognize our desires as desires for particular things by noticing that when we have certain feelings, we tend to do certain things. If we did not do those things, we would never conclude that those feelings are desires for doing those things. Note that someone could raise a similar objection here: if this is true, then are not desire and good mere words? We feel certain feelings, and do certain things, and that is all there is to be said. Where is good or purpose here?

The truth here is that good and being are convertible. The objection (to my definition and to Clark’s account) is not a reasonable objection at all: it would be a reasonable objection only if we expected good to be something different from being, in which case it would of course be nothing at all.