Computers and the Mind

[00:00:03] SPEAKER 1:
Good afternoon. Welcome to this year’s Foerster Lecture. In nineteen twenty-eight, Miss Edith Zweybruck provided an endowment for the Agnes and Constantine Foerster Lectureship on the Berkeley campus.

The subject of the lectureship was to be the immortality of the soul or some kindred spiritual subject. The first Foerster Lecture was delivered in 1933, and by 1964, some twenty-five Foerster Lectures had been given, mainly by bishops, theologians, and moral philosophers. No lectures were given between nineteen sixty four and nineteen sixty nine, the five crisis years on the Berkeley campus, extending from the days of the Free Speech Movement to the climax of the People’s Park.

Upon the advent of the present post-revolutionary era, the Graduate Council decided to reactivate the Foerster Lectureship and appointed a new lectureship committee that was to take a so-called modern view of the words immortality and soul. What would be a modern view of the immortality of the soul? Insofar as there is a modern view at all, I think it can be traced back to Descartes and the seventeenth century.

Descartes, you may remember, laid the philosophical foundations for physiology by advancing the most fruitful notion that the bodies of humans and animals can be considered machines. But since moral principles obviously do not apply to machines, but do apply to humans, humans must be more than automata in human shape. The extra something that makes men more than automata, that’s the soul,

(clears throat)

an agency that is not itself part of the body. It is from their soul that men derive both freedom and responsibility for their actions. Though the subsequent rise of empiricist and, uh, particularly positivist philosophy relegated this Cartesian view to the realm of nonsensical metaphysical propositions, I think it can fairly be said that nothing so far has really replaced the Cartesian body-soul dualism when it comes to the problem of dealing between the intersection of morals and human biology.

In any case, uh, the question which Descartes defined so clearly, namely, Are men more than machines? has been debated ever since his time. What’s kept this debate alive is the simultaneous growth of the science of psychology on the one hand, and advances in machine technology on the other. Just as the analysis of human behavior came to show ever greater complexities in the workings of the mind, so also did we come into possession of ever more, uh, complicated and capable mechanical and electrical devices, which could perform ever more complicated tasks.

Today, this debate still continues. The rise of neo-Cartesian, neo-Kantian structuralism in the human sciences, said of more or less by Freud at the turn of the century and flourishing today in such disciplines as anthropology and linguistics, has revealed profound difficulties which face any attempt to find the causal connections that govern human behavior.

[00:03:27] MODERATOR:
At the same time, recent developments in the design of electronic computers have suggested, at least to some students of this problem, that there may not be any fundamental limitation to the degree to which the operation of such machines can be made to approximate the workings of the mind. The subject of the three Foerster lectures to be presented today, Computers and the Mind, is addressed precisely to this problem. We are most grateful that three outstanding authorities in this domain, Professor Seymour Papert of MIT, Professor Hilary Putnam of Harvard, and Professor Donald MacKay of the University of Keele, have accepted our invitation to present to us their views on this profound problem.

Two of these lectures will be given here this afternoon, and the third lecture by Professor MacKay will be given this evening here at eight PM. Well, now it’s my pleasant duty to introduce to you the first of this year’s Foerster lectures, Seymour Papert, Professor of Mathematics at MIT. Professor Papert was born in Pretoria, South Africa, and studied mathematics in the Witwatersrand University of Johannesburg, where he received his PhD in nineteen fifty-two.

He has been a member of the MIT faculty since nineteen sixty-three and became co-director with Marvin Minsky of its Artificial Intelligence Laboratory in nineteen sixty-seven. Professor Papert is eminently qualified to speak on our subject, computers and the mind, since he combines a high professional competence in the actual programming of computers for the performance of complex tasks, with a deep knowledge of cognitive psychology, particularly its ontogenetic aspects, which he acquired in the course of a long association with Piaget in his Center for Genetic Epistemology in Geneva. So it’s with great pleasure that I introduce to you now Seymour Papert.

(applause and cheering)

[00:05:27] SEYMOUR PAPERT:
Oh, It’s a complicated machine.

(laughter)

Well, I had a lot of trouble trying to decide what I should call this lecture, and then it appears that they don’t have to have names, but they are merely thoughts on the same subject by three people. However, I’d like to tell you about some of the names I, the two names that I couldn’t decide between, and they really, since these really are themes of what I’m going to say. One of the names is the Simplicity of Mind.

I think that’s quite a nice name, and I think it’s relevant to the question of, of artificial intelligence in various ways which will come out at a number of points in my– in the next three-quarters of an hour. I’d like to start off with a very simple one, though, and that is the following observation, that when you see somebody doing something that seems very difficult, There are two explanations. One is that he’s smarter than you thought he was, and the other is the task is simpler than you thought it was.

And it’s often very hard to know which it is. Indeed, I’m going to suggest that we are extremely bad at knowing which it is, and that we hardly ever do, except in those cases where we rarely find out that the task is simple. That in fact, we have very, very few examples where we can, in any competent, technical, rigorous way say, “This is a really difficult task of such and such a degree of, of difficulty.”

So it becomes at least an open question to be considered whether the problem of making machines be like people is going to resolve itself through discovering how to make machines much more complicated than we know how to make them now, or whether it’s going to resolve itself through discovering that people are really much simpler than we have l-been led to believe in the past. It must be remarked, of course, that as we’re hard– we’re bad at judging the difficulty of problems, we’re particularly dif-bad at finding out through introspection or any other direct means how complicated or even what our own mechanisms of behavior are. And I’m going to take a couple of examples later on to, to illustrate that.

I, I’ll mention just one right now, which is something that I’ve been studying a lot recently, and that’s how people ride bicycles. Now, almost everybody goes through the following sort of dialectic about riding bicycles, about your feeling of how complicated riding a bicycle is. First of all, you think if you haven’t thought about it at all, that there is nothing to it, it’s extremely simple.

You get on the bicycle and you ride. So that’s the extreme of simplicity. Then you might get involved in teaching somebody how to ride a bicycle, maybe a kid.

Better still, a grown-up, it’s harder. So then you start, you might try to formulate to him what he has to do in order to ride a bicycle, and after a while, you convince yourself that riding a bicycle must be extraordinarily difficult. Ah, well, A-a little investigation of the explanations that people give when asked to explain to somebody else how to ride a bicycle has revealed some interesting data.

Uh, for example, almost everybody, whether physicists or psychologists, has at least one false belief about how to ride a bicycle, about the control process in riding the bicycle. It follows that if they try to give explanations to pupils and express what they really think happens in riding a bicycle, they are almost sure to mislead. Hence, their verbal explanations are almost sure to run them into the hole of believing it must be an extraordinarily complicated and difficult thing because this damned pupil just can’t carry out or won’t carry out these instructions I’m giving him.

Now, the… uh… poignancy of the story comes out of the third phase of this dialectic process, having gone from a very simple view of the bicycle, you don’t do anything to this extremely complicated view because we can’t teach people how to do it. When you really come to understand it, you find the following dramatic fact, namely that, uh, bicycles are actually quite stable all by themselves, and if you, you push a bicycle down a hill, it goes quite a long way before it falls. If you put a beginner on it, it, it falls very quickly.

So the real task of the beginner is the simplest of all possible tasks, namely to do nothing. So this is why it appears that the the task of, that the way to learn a bicycle is not through verbal instruction or anything like that, but through, through learning by doing. Uh, well, but once we understand that what it is that you have to learn by doing, namely nothing, we’re inclined to swing over to the other side and say that, “Well, it’s extremely simple then to ride a bicycle.”

Uh, however, pursuing it a little further, we find that if we try to get ourselves to do nothing in various situations, like if there’s a lion over there, we find it’s extremely difficult to find– to make yourself do nothing. So we swing backwards towards a complexity again. And it’s that kind of play of is it complex, is it simple?

Depends how you look at it that I’m really going to talk about, and that I think is at the heart of the question of, uh, what k-kind of mechanism can behave like a person, what kind of a thing is the human mind, and so the title, The Simplicity of Mind, not as an assertion, but as a question. Uh, more than a question, actually. I’d like to convey to you a certain image of the development of a science which is generally called artificial intelligence.

And, uh, what I’d like to give you at the end of the, of, of the lecture is not so much a technical knowledge of this very difficult and complicated subject, but at least a view of what some people in it see as the main trends. What do we think it’s doing? Um, and how do we think it should be judged?

I think generally, judging from the, uh, public discussion on the subject, people show themselves to be extremely bad at talking about artificial intelligence. And you might say that what I’m talking to you about is at least how, in one person’s view, you– how to talk about talking about how– Sorry, I got tongue-tied there.

Uh, I’d like to talk about talking about artificial intelligence and, um, now, so that was another title talking about talking about artificial intelligence. I didn’t consider that one very seriously, but I considered one more very seriously, and that’s one that came to my mind once as a spontaneous remark when somebody said to me, “What do you think about artificial intelligence?” And instantly the thought came, I found myself saying, “Is there any other kind?”

And that leads to the question, the suggested title, The Artificiality of Intelligence.

(laughter)

Now, this really is a serious question. Is there any other kind? Look, is intelligence an artificial thing?

The point’s been nicely argued from an angle somewhat different from mine by Simon in that very brave, beautiful book called The Sciences of the Artificial. Uh, uh, I don’t agree with very much in it, but I think it’s one of the most clear and direct attempts to get out of the details of, of artificial intelligence and say what it’s about. And I think people in the area have been seriously remiss in not doing so, and Simon makes a very brave attempt there.

Us– Piaget, one might also say, has taught us a sense in which intelligence is artificial. That intelligence, in Piaget’s view, is something constructed by the child as he grows older, as he grows. Does he?

It’s not something innate, it’s not a mechanism that emerges from the bio-biological structure of the brain, it’s a construct. Uh, in a simple-minded way, one might say, let’s look at that howling infant just born into the world and look at the intellectual he might grow into twenty or thirty years later. And it’s clear that, uh, as he came howling into the world, he didn’t have very much

(coughs)

of that equipment that’s going to make him such a powerful thinker one day. Well, uh, the point of course is not it’s it that doesn’t prove anything, and the open question is the extent to which the, uh, that stuff that was there in the brain is, um, at, at birth is an essential part of the intellect that’s to grow from it. After all, if I say I made a roast beef last night, it would be perverse to argue from that, that roast beefs are artificial things, and the properties of roast beef don’t have anything to do with, with steers or biological meat just because I made it.

Uh, it– it– My contribution was relatively small. The contribution of the, of, of the biological system was enormous. Uh, if we accept that we make our minds, we don’t have– there’s a whole spectrum of the extent to which the prime matter out of which we made our minds, as it were, is the– is, is important.

But I think that this is an open question that, you know, spontaneous superstitious points of view on it have– don’t have very much value, and particularly since one can see some very powerful reasons why people are drawn into, into holding them. So the thesis about the artificial artificiality of mind is this: that it is an open question, and elucidating it is a lot of what the subject of artificial intelligence is about. Uh, let me take another example to point that, uh, rather curious issue, which I think, uh, people ha– needs a little reflection to, to see really what I’m trying to say there.

I’d like to contrast two kinds of theories of intelligence by considering the kind of remark that we all make about people, uh, let’s say two chess players. You know, these two chess players, A and B, they play chess equally well. But you say A does it in a plodding way because he’s read all those books, and he knows all the opening sequences by heart.

He’s just got a lot of knowledge. Whereas B has hardly read anything, has only played six games in his life, he does it by sheer native intelligence. So you might say this one does it by brain power, and this one does it by knowledge.

And I think that image of the opposition of brain power with knowledge is something that’s, uh, built into our culture to the extent that it’s very hard for us to, to even sta- to stand away from it long enough to, to ask whether it can, it– can be put into question at all. However, I think we’ve got to put it into question, and I think that the, uh, most important developments in artificial intelligence over the past ten years, five years particularly, have been, uh, steps towards making more precise that question that I just asked. The question of how much– how can we really define in a rigorous and responsible way that opposition between brainpower and, and, and, and knowledge as a source of, of, of intelligence?

And, uh, so there is developing two very distinct theories, I think, of, of, of the, the, the nature of intelligence. Let’s call them the brainpower theory, and let’s call them the, the, the epistemological theory. Uh, you might say, uh, surely it’s obvious in the case of my two chess players that the player A in fact does use knowledge and player B doesn’t use knowledge.

But a moment’s reflection should convince you that that is not at all obvious. All that’s obvious from my definition of the problem is that player B had less knowledge specifically acquired in order to play chess than player A. And this says nothing about how much general knowledge he had about strategies, about warfare, about psychology.

There’s– about how to organize the solution to a problem, et cetera, et cetera. He– It is an open question whether that which we call in him the native ability is itself the result of some other set of knowledge which is not so obviously and specifically related to chess.

So it is an open question how much knowledge he he contributes, and the technical problem for the study of artificial intelligence becomes the question, what kinds of knowledge are there and how can we classify them, and what kinds are relevant to different sorts of problems? So the study of artificial intelligence moves into the enterprise of creating what I think is fairly described as a new epistemology, a different kind of classification of knowledge and, uh, is giving rise not only to answers to questions about what kinds of knowledge they are, but in fact, to, uh, people recognizing important pieces of knowledge that had never been recognized before, mainly because they were too simple, but not always, and I’m going to give some examples of that. So that’s a preamble about what it’s all about.

Let me say then briefly, uh, asked to talk about computers and the mind, uh, I find I’m going to, uh, talk very little more about either computers or the mind. Uh, because I’m going to suggest that that’s not where the problem is. I’m suggesting a thesis that might at first sight seem quite paradoxical, that the, uh, problem about, that the problem about computers and the mind, or the problem that you think is, is about computers and the mind, really is about the nature of knowledge.

And so we’ll focus on the nature of knowledge as the, as the primary, uh, subject for, for discussion. Now, I said earlier that, uh, that’s a trend amongst people in the branch of, of inquiry called artificial intelligence. It’s not quite true, though, that it’s not a uniform trend.

And in fact, uh, I think that the subject is in a state of identity crisis, which will probably end in a kind of rupture of it being recognized, of people going off in different ways and, and s– people following absolutely different directions and not even trying to see them-themselves as belonging to one subject. Maybe so or maybe not. But, uh, I want to explain to you where I see the identity crisis.

And again, in the interest of trying to give you an image of the emergence of this science, a science trying to define itself, trying to set up its paradigms, and, uh, in doing so, going through many throes and crises, not unlike those that we saw physics in, in the time of Galileo or in the early quantum periods, and as constructive as those. Um, it’s only philosophers and journalists who seek to, uh, criticize and denigrate artificial intelligence by pitting us against one another and finding that whatever one of us says, somebody else will contradict, and, uh, I think that these contradictions are, in fact, enormously constructive and a rema– and a, a sign of great vitality. Well now, the identity crisis can be seen most clearly in– if you open different books.

I’ll take two extremes. A recent textbook by Nils Nilsson on problem-solving methods in artificial intelligence. You open it, and it looks like a book on logic.

It’s full of formulae, it’s full of, of algorithms for search, for proving things in the, in the quantifier calculus, and so on. Uh, so there’s an image of artificial intelligence as, as a hard mathematical subject, as a hard science. At the other extreme, uh, a lot of the activity at MIT, where there is no lack of, of mathematical activity, but the other extreme, we’ve recently had students writing theses on subjects such as how to make a computer program to understand children’s stories, where children’s stories involve incidents like Janet and Jack are sitting, playing together, And Janet has some, has just got a present, some crayons, And Jack comes in and says to Janet, um, “Those are terrible crayons.”

“Uh, I’ll give you my dog for them.” And it becomes obvious from the dialogue that, that in fact, Jack doesn’t think they’re terrible crayons and is trying to, to take in Janet and persuade Janet to, uh, to believe the crayons are terrible, so she’ll do this trade. Well, now, in the course of working on this problem, you find yourself getting into an activity that’s very far from the image of the hard-headed mathematician with his, with his rigorous proofs.

You’re much more like the image of the literary critic sitting examining a novel, or tear it– trying to see what the presuppositions and the, uh, worldviews of characters in this novel are. And indeed, you have to do that according to, well, certainly our epistemological view of artificial intelligence and, uh, probably on anybody’s a priori account, because how on earth could this program understand what’s going on in the story in any sense without knowing that there’s such a thing as trading, as lying, as et cetera, et cetera, as deliberately misleading people and, uh, so unless there is– unless these– the– this kind of knowledge is somehow, uh, in– somehow got into that machine’s repertoire of knowledge, it has no hope at all of doing anything with the little story of, of Jack and Janet. And so we have to sit and pore over that kind of soft knowledge that looked like it belonged to, to the literary criticis– literary critic, and this seems to be feels like a very far away extreme from, from proving a, a hardheaded mathematical theorem.

And well, I think the identity crisis is represented in the extremes by these two forms. Um, i-in fact, it’s going to resolve itself by finding more technical ways of talking about the kind of knowledge that the literary critic, critic teases out of the, of the story. Uh, however, uh, it’s an open question, and, uh, I see the possibility of the resolution.

In fact, I see almost the certainty of, of the, the main line of artificial intelligence getting to be the classification and study of that kind of knowledge which children have. That kind of knowledge which nobody’s ever written down in books because everybody knows it. There might even be an analogy with lex– with the development of lexicography.

There was– The making of big dictionaries is a relatively recent occupation. I believe that in medieval times, the only dictionaries that existed were foreign language dictionaries or Latin word lists, and the idea of writing down, of, of making a compendium of, of all that knowledge that goes into a dictionary was not something that had really been posed. Uh, even if I’m wrong about my history, which I don’t think, uh, the point in principle should be clear that it’s– there was a time it could happen that all that kind of knowledge was passed on, somehow permeated, diffused through the culture, and never had to be written down, and nobody ever tried to write it down or classify it.

There’s, in the same way, but that’s not to suppose that it was generated in a child by any innate mechanism or that it’s a reflection of, of some structure of the child’s brain, and, uh, the child doesn’t know what the word dog means except by knowing what the word dog means. It’s a hard piece of knowledge that has to come specifically from somewhere. Well, now, if one thinks of the kind of knowledge that one would call common sense, for example, I’ve got an area here, and I’ve got some objects on it, and I’d like to put a new object on it, and I can’t find a place for the new object.

Then we all know what to do. You could, you’ve got to make some space, and you can make the space by taking away some objects that are already there, or you could make the space by pushing them aside, compacting them in some way. In any case, we have a certain number of, of techniques for making space when we need it.

And this is a specific piece of knowledge which you don’t have to explicitly transmit to somebody because everybody has got it by the time that he could read it in a book. And nevertheless, in the same way as like the meaning of the word dog, there’s no reason to suppose that it emerges automatically, and one might see in the future the development of, uh, people who will put together, explore, classify these kinds of knowledge that, um, of which that’s a, a typical thing, knowledge of the very ordinary. Well, so that’s a rough picture of, uh, of the…

In a very abstract form of, of trends in artificial intelligence. I’d like to now turn to some more technical little examples to illustrate, uh, in, you know, on a more finer scale and more precisely what I might mean in, by the pro– by the, the problem of finding out, you know, by the suggestion that one would really find out that it’s easier than you thought it was rather than… Um, I’m sorry, I should– I don’t want to make a general statement.

I want to give you some very simple examples. I’ll take a first example, which is like riding the bicycle, and that’s catching a ball. Okay.

Uh, so let’s suppose we’re in a baseball field, and there’s somebody out there who throws a ball. It goes over like that, and somebody here is going to catch it. So, and now we say to ourselves, how can this guy catch that ball?

Um, well, you could set up the problem so that it seems extremely difficult. Certainly, if you set it up in the following way, you said, “Well, what must be happening here is that he must be setting up some differential equations of of motion, and he must be doing some time prediction analysis, and here are these equations, and somehow that brain must be solving these equations, and that’s fantastically complicated.” Uh, and you need all this immense number of ten to the twelve or whatever neurons to do that, and no mere serial digital computer could possibly do that in enough time to play a single baseball game.

Well, so you can. You can set up the problem, and if you’ve set it up that way, you’ve dug yourself into a trap of looking for impossible mechanisms in brains. However, uh, that’s not the way people catch the ball.

There isn’t any differential equation. There happens to be an extraordinarily simple algorithm that leads to the ball being caught. The extraordinary algorithm is as follows, that if you– I can, just to state it a little, so, uh, simply.

Let’s, let’s imagine a plane back there, and suppose the– that on the– you think of the ball projected on the plane. This is just a mathematical device for me to describe the algorithm to you. And let’s say that, that there, that distance is, is h at time t, that’s the height of the virtual ball.

Now, uh, the algorithm says as follows, that if the virtual– if the ball is going to hit you in the eye, the virtu-the virtual ball will increase at a constant rate, will be equal to k, all the time until it gets into your eye. Now, notice how counterintuitive that seems, and almost everybody tested on that says, “How could it be? Because how do you mean it goes up at a const–

It must first go up and come down. Well, you see, false. It doesn’t.

See, just imagine to see that what would happen if it went over your head that way. Uh, in fact, it would go up to infinity in your visual field. And of course, the projection, the ball is, is going up, up to proje– up to infinity there.

So if the thing is going up at an increasing rate, it’s gonna go back after your head, over your head. If it’s going to fall in front of you, indeed, it will go up and then come down. If it’s going to hit your eye, it will go up at a constant rate.

Now, I’m not going to prove that. Actually, any high school student could prove it, but I don’t want, I don’t want to take up the three minutes. It’s a nice exercise.

The point is that, uh, if you– once you’ve noticed that, you’ll find that there’s an extremely simple algorithm which says, uh, observe this height of this virtual ball, and if it seems to be increasing, observe the rate of change of this height of the virtual ball. If it seems to be increasing, then run forwards, run backwards, because the ball is going to come behind you. If it seems to be decreasing, run forward, and you will catch the ball by following that, that algorithm.

So there’s a case where, uh, it seemed to be much simpler than it was. Now, of course, there are lots of problems still. I haven’t told you how this thing is calculated.

I haven’t told you what neurological mechanisms there can be. However, what I have done is done something about delimiting the computational complexity of the task we have in mind. It looked like a very complex task.

When I made this mathematical discovery, this discovery about the world outside, the task would seem to be of high computational complexity turned into a task of, of low computational complexity. Now, that of course, doesn’t mean that the silly brain might not do it in an extremely inefficient way, But it means that we’ve changed the problem. The problem now isn’t how on earth can a mere brain do such a complicated thing, and the brain must have some extremely complicated properties to enable it to do it.

The problem is rather, well, now we’ve seen how simple it is, we can think of any number of mechanisms, and maybe we can get our neurologists to go and look for, for the actual one that, that operates. So there’s an image of what– of a kind of knowledge about the ball catching, about the behavior of balls, a purely mathematical knowledge that changes your picture of what the problem for intelligence is. And I want to suggest with obvious rhetorical exaggeration that that is a better model of the stuff out of which progress towards artificial intelligence is made than looking at neurons through microscopes, or dealing in philosophical paradoxes, or whatever, uh, you might be led to do either by poring over the computer or by, or over the mind.

It’s looking at what’s in the outside world. Uh, that at least might, and this view does give us the, the, uh, the kind of, of, of insight and possibility of, uh, believing that machines might one day do these things. Well, now this example of the ball is, um, it’s an example that didn’t come out of an artificial intelligence laboratory, and in many ways it’s not an ideal example.

And what I’d like to do is to give you just one other example of, uh, of what I think is a very nice piece of artificial intelligence fact, artificial intelligence knowledge. Uh, by the way, I think that I’m faced, you know, I’m faced with a problem that bothered me for a long time. What could I tell you in this to illustrate this?

And, uh, s- thinking about the, the problem, I recalled and reread that beautiful book by Hardy called A Mathematician’s Apology. And if you remember, which you ought to, or to read it if you haven’t, Hardy says– poses himself the problem of how to tell a non-mathematical audience, how to give them an example of what mathematical knowledge is. Yeah, how can I tell somebody who hasn’t done any mathematics, why math– what is important and what is real in mathematics?

Well, not by telling him about how mathematics was used to build a bridge. Uh, because that’s– it’s– it might or might not be the case that good mathematics was used to build the bridge. Hardy thinks that bad mathematics was in fact used to build the bridge.

But in any case, the fact that a bu-bridge was built is irrelevant to the, to, to the quality and nature of the mathematics. Uh, what he does in the end is he finds a very beautiful example, namely proves the irrationality of the square root of two, because this is a piece of mathematics that seems to him to be in some special way typical of real mathematics, and yet I can tell it to somebody in a, in a few lines. Uh, I’m not gonna take the few minutes either, and if you don’t know the example, then I’m not sure that you aren’t in such a different cultural world that I can’t even talk to you.

But,

(laughter)

uh, we’ll pass that by and and say that the example I’m going to give you now is, uh, I think of it as rather like a, a, in artificial intelligence, what the root two in, in is for, for Hardy. Namely, this is not the most important piece of mathematics, and nor is the fact I’m just about to give you the most important piece of artificial intelligence. Just is very pretty and in an, in a significantly elegant way.

Now, before I do it, though, I think it’s very– I’d like to take up Hardy’s point about the bridges. I said at the beginning, I wanted to talk about how to talk about artificial intelligence, and it’s getting towards the end, and I must say something about some ways not to talk about artificial intelligence.

And, um, one of them is not to focus on, um, this crass empiricism of seeing what the latest score in the chess games is or, or how big a tower the, the robots are building or whatever might be. Actually, I think at the beginning it was much easier to, to know how to judge of artificial intelligence. And there was a day not very long ago, twenty years approximately, when the first– a computer played chess for the first time, and everyone said, “Wow!” and thought it was a great thing.

In fact, it was a great thing. Then after a while, it came out that a ten-year-old child had beaten that computer program, and then Dreyfus and the New Yorker said, “You see, it’s–”

(laughter)

they’re not so smart after all.” And, well, so and then, um, somebody came along, Greenblatt came and wrote a chess program that could play in Chess Federation tournaments, and, uh, and now they say they stopped talking about ten-year-old children. They said, “Well, but it only beats

(coughs)

mediocre players.” And then they said, “Well, it only uses twenty, but it uses twenty-six thousand moves every time it has to choose one.” And then somebody wrote one that only looked at two thousand.

Well, and so it goes like that, you know, that, that, uh, tennis match of the, uh, s- male chauvinist pig and the… Uh, well, for a while, that was fun, and it was significant, but that’s no way we can judge things anymore. For one thing, there’s nothing intrinsic in it.

And so what? If you’re involved in chess programming, it’s nice to see that the rate of increase of points in the Fe- Chess Federation scores seems to be keeping up about a steady rate year by year, or it, it fluctuates a little, but it averages out. And, well, that’s nice to know, and you can predict from that when you think it’s going to be a- an A player or whatever.

But, uh, there’s– if you really look at the details of how it all happens, it’s all such a horrible mess inside those programs that there really is no significance to, to just where that, that, that, ra-that rating is. So we’ve got to look at a different kind of progress. Now, I’ve already mentioned one, which I think is the most important, and that is that artificial intelligence is beginning to define itself as a science of knowledge.

And, uh, now it’s time to give some examples of, of the kinds of knowledge, and I’m going to give, first of all, an example of specific knowledge on the lines of these. This comes from a very beautiful story, and I’ve used my favorite, a s– a terribly simplified example of– that everybody knows, I guess. Uh, I can’t draw it very well.

Uh, that’s meant to be a picture of a wooden block or a block, and behind it there’s a bar stretching out like that. And you’re meant to imagine a computer looking at this, and the computer is given the task of seeing whether there’s a cube there. Well, now, um, there once was a time, say, in 1950, when people began to talk about something they called the problem of pattern recognition.

And right through the ’60s or so, there was a– the, the, the problem of computer vision, if you like, was defined as pattern recognition, and pattern recognition was thought of as something like this, that I’m giving you something. Here it is.

Tell me what it is. Um, when robotics started and we tried to get computers to look at this picture and say there’s a cube, or better still, to say there’s a bar behind it, we realized, not immediately, that in fact the problem wasn’t inside the paradigm of pattern recognition, Because you weren’t given the thing. It wasn’t as if you were given that and you asked, “Is it a cube or a triangle or a pyramid?”

It’s that you were given this complex which might have had dozens or hundreds of things in it, and you had to pick this out as the, as the important piece. And it appeared that the most intellectually interesting and stimulating and fruitful, uh, source of, of thinking and, and advance there was that shift of the problem from recognizing to, to picking out. Now, I think it’s very interesting that the we might have seen that by reading back in psychologists, that the school of psychologists called Gestaltists had long ago said, “Well, of course, the problem in perception is the figure-ground problem.

And you can’t do the figure-ground problem by local whatever it was. Well, I noticed a number of points about that in a digression. First of all, that, uh, of– although psychologists had said what we found, it was embedded in a lot of other things we couldn’t use.

Um, notice that the psychologists who said it were the Gestalt psychologists who were considered and considered themselves the anti-behaviorists and the anti-mechanists. And I think that is a typical sideline about the history of artificial intelligence, namely that it’s quite wrong to think of artificial intelligence as lining up with certain schools of psychology that in the past called themselves mechanistic or behavioristic or, or whatever it is. It’s much more often the case that, that the methods that we find valuable were enunciated by the anti-mechanists and the transcendentalists and the gestaltists and so on, and for very good reason.

Namely that the mechanism that anybody had in mind before computers happened must have been so primitive that if he thought that with that mechanism it would be possible to achieve intelligence, he had to be wrong. So the mechanists of the past were systematically wrong, and the arguments brought against them by the various anti-mechanists and transcendentalists were very often right, but they were also very often wrong, so there’s no goldmine of good arguments by just going to read Kant. Ah, nevertheless, coming back to this, the shift– the first point I wanted to draw attention to is the shift in the problem from pattern recognition to a different paradigm of problem, namely the separation out of the object.

Uh, now how did we separate the object? And the way we separated the object was by hitting on a new piece– a new kind of knowledge. And this new kind of knowledge is about the local features of a scene like this that you can use to, to break it up.

And the knowledge has to do with noticing the kinds of configurations of where lines meet, for example. That’s a certain kind of vertex, a T-vertex. That’s a certain kind of vertex, a fork vertex.

And that observation that by once– if you recognized these vertex types, you could, by using these, formulate rules for separating out objects which could not be for– which had not been formulated before, and which nobody yet knows how to formulate without using such knowledge as, as in terms of vertex types or other new forms of knowledge. Um, you then see that, you know, this is the thing that’s like my root two, this example of a new kind of knowledge in terms of which, what seemed to be unrule-like and not formulatable in any simple rules now becomes formulatable in terms of some very simple rules. The simple rules really are very simple and such as where you see a T, that’s evidence for breaking it.

Uh, well, that was the first primitive step. Since then, the, the study of vertex types and edge types has become a whole new science and has led to some, some, some very beautiful theorems and, and, and theses and programs that work, and quite a number of people have contributed to it. Uh, so there’s an example of, of an entirely new piece of knowledge about new concepts, new statements, new theorems about an area of reality, namely how bodies configure.

It’s a new geometry which came out of the attempt to make a machine see, which is independent as pure knowledge from anything to do with machines or any properties of machines, whose effect is to change the apparent computational complexity of the problem of seeing. That what seemed very complex now becomes very simple. And that is the ex– and, and that’s, that’s the example I wanted to give.

And like the balls, typical of the sort of thing which I think is, um, is li– is, uh, on the main line of what’s happening in artificial intelligence as the source of knowledge. Uh, this might seem very strange, and I’d like to end up by making a few remarks about analogies with other sciences and philosophies and things that one might, uh, uh, that might bolster this paradigm of seeing artificial intelligence as quite a new science that doesn’t have much to do with the mind or computers directly, but changes our view of the relation between them. Uh, one, uh, I’d like to make an analogy f– with another, with some other sciences.

Aeronautics, flying is probably the, the, the one I can say most quickly ’cause I’ve said it often before. Uh, when people in the nineteenth century wanted to understand how birds flew, some of them did it by examining birds. And you might have imagined some of them pulling out all the feathers of a bird and finding, well, the bird doesn’t fly now, so the feathers are the organs of flight.

So we might study the feathers in great detail, like some people study neurons because they notice if you pull all the neurons out of a man’s head

(laughter)

Um, well, now that’s not the way it happened. Yeah, it didn’t happen that we understood how birds fly by watching birds fly. We understood how birds fly by doing something qualitatively different from that, namely by making a science of flight

And we made a science of flight partly by watching birds, of course, partly by building airplanes, partly by doing things in wind tunnels, partly by doing things with, with mathematics. A whole lot of activities went into this thing called making a science of flight. And when we’d made a science of flight, so we understood the principles of lift, for example, then we could go back and say, it’s now, how do the birds fly, and we could discover a lot about it, not that we yet know everything.

Uh, but and we could make airplanes. And I think that in artificial intelligence, the situation is very much the same, that it’s neither by studying the human brain to see how people think, you know, nor for that matter by studying the computer to see how the computer doesn’t think.

(laughter)

You know, that you’re going to, uh, make the progress, but that the analogy is that we are to make something which is to be a science of intelligence. And I think what the science of intelligence is about is how simple are the things that seem to be complex. Uh, so, uh, that’s that pa– the paradigm.

It’s hard for us sometimes to believe in it. We are inclined. There are many reasons we that, that many dazzling features of this problem which tend to capture the gaze, and you can’t get away from them, like the snake looking at the bird or the bird looking at the snake, whichever.

The bird looking at the snake. Some of these are paradoxes, some of them are fallacies. I think it’s important if one wants to think about the problem, at least to ration the amount of time one will allow oneself to think about them.

An example of a paradox, well, this is a, an oversimplified one that I learned, of one that I learned from Professor Mackay, who will be talking this evening, is, for example, saying, well, if you could make a s– a computer, if you could make a mechanistic model of mind, then you’d be able to predict its behavior. So you’d be able to predict people’s behavior. But we can’t predict people’s behavior.

At least in some conditions, circumstances, we can’t predict people’s behavior. For example, if I wanted to predict that the next word you say is no, and my calculations of, of your state, etcetera, had led me to, to know that what you’re gonna do is to say the opposite of what I’ll predict, then I’m really in a bad way, and nothing I can do will make me come right, because whatever I predict to you, you you– it’s in your power to, to disconfirm my prediction.

Well, now that I think it’s a paradox. It’s a paradox like the paradox of the barber who had to shave all the people ex– who didn’t shave himself– themselves, and so didn’t know how to– whether to shave himself or not. Uh, the point about the para– the point about the paradox of the barber is that it says nothing about the limitations of barbers.

It, it only points to some conceptual difficulties that we have in the way our logic works. And many paradoxes of mind, such as that you can’t predict people, are of that nature. They say something about the– our difficulties in conceptualizing everything about mind.

In any case, they only point to the extraordinary. And that’s really the theme on which I’d like to end, that, uh, what we are concerned with in artificial intelligence is not mainly, or even at all, what happens in these extraordinary situations where you say to somebody, “I’m going to predict your behavior, and let’s see what you do.” Uh, we’re much more concerned with the ordinary, everyday situation where we do predict people’s behavior, and we wouldn’t be able to survive in a society if we didn’t predict pretty well.

Related to this is what I like to call the superhuman human fallacy. And this one’s a fallacy. Uh, and you can posit like this, the superhuman humans.

Uh, many times when people discuss what machines can’t do, they ask the machine to do things that people, that most people certainly can’t do. Uh, example, to given no knowledge except these four axioms for group theory, to prove a theorem.

(laughter)

Well, the machine screws up. Of course, it screws up. So would a person.

(background chatter)

Uh, right along the line, There’s the, the tendency to, uh, well, the extreme case is the argument from the Turing, that Turing’s, you s– you might say, Turing has shown that for any machine, there’s a problem which that machine can’t solve. Ergo, people aren’t machines. And the conclusion would be true if it weren’t unfortunately the case that for any person, any person I’ve known, there’s a problem he can’t solve.

(laughter)

And, uh, so talking not about superhuman humans, but about ordinary humans, I think these paradoxes and these Turing-like arguments are rather irrelevant. Even the question of whether they can play chess becomes irrelevant because already the machines can play chess better than most ordinary people. And whether they, and what it’s going to take to make them play chess at the peak of human performance is another problem which doesn’t seem to us to be in the forefront of attention.

Our– We’re not simply shirking the issue in that, we are following what’s turned out to be the strategy in all sciences. And all sciences have advanced by solving a piece of the problem at a time. There’s a difficulty, though, with intelligence.

Namely, if you take a small piece of intelligence, you have a special name for that. You call it stupidity and you say it’s not intelligence. And so any of the steps towards artificial intelligence has to be classified as stupidity, and so as non-intelligent.

And so on that way of thinking, you can never get there until you’ve got there. That fallacy is known as modus postponens, which… Um, so, uh, that’s what–

Oh, I’d like to say one other ordinary thing about an ordinary use of intelligence. A theory of intelligence would be– should be of use in more than one way. And now, I know two ways in which you’d expect a theory of intelligence to be useful.

One is in making machines. Of course, there are others in the pure intellectual satisfaction, understanding ourselves. So there are many.

But amongst all of those, there ought to be– it ought to tell us how to teach children to be more intelligent and ought to have implications for a theory of education. And following that thought at the MIT Artificial Intelligence Lab, the, um, there’s been a growing project that’s now about a quarter of the activity of the lab that consists, in fact, of trying to develop an alternate to the whole concept of, of school as it exists for children. Now, uh, why us?

For a number of reasons. First of all, uh, we put our– we’re in a different situation from the psychologists who seem hidebound by the fact that people are limited in their, in their intelligence as shown by IQ scores, and we know that if you apply those IQ scores to our PDP-10 computers, the computers flunked completely. and so if IQ scores are going to be accepted, we might as well give up, so.

But, um, there’s a more important point, and that’s related to my last point about the ordinariness, the study of int- of intelligence in ordinary situations like catching balls and seeing blocks. And that’s this, that if I would like a child to have a theory of intelligence. By that I mean, your child’s business is to learn, for example, or to understand.

That’s what he goes to school for, they say. Well, if that’s his job there, his professional job, then he ought to have a lot of specialized knowledge about it. So he ought to be an expert on learning and understanding.

So you ought to have some kind of theory or model of the cogni– of the intellectual processes of understanding and and learning. And so the heart of our replacement for the school idea, or one of its, its hearts, is to give the kids that kind of expertise about understanding, learning, intelligence in general that they ought to have as practitioners of those, of those arts. Well, now, what model can you give a child of intelligence?

And let’s consider the models that exist. The models of himself that are offered to him by our culture. Well, I’ve seen in books where it’s said that a man is like a hydraulic system.

(background chatter)

I’ve seen in books that it says a man is like a worm or a rat in a maze, or like a collection of SR reactions. Uh, I think if you go and look at all the models that exist in those textbooks of the human being, he’s like a telephone exchange.

(background chatter)

What others? You, you must know them. They’ll just come to mind if you think for a while.

Compare those models with the model I’m going to tell you. Namely, you are like a complex system of computer programs. Now, of course you’re not a complex system of computer programs.

Of course that has limitations. But I think that compared with all those other models, it’s vastly richer and has the merit of being actually useful, that thinking of yourself in those terms leads you when you get your sums wrong, instead of saying, “I am dumb,” or, “I don’t have a mind for mathematics,” to saying, “Well, let me debug myself.”

(laughter)

Uh, I think it’s a poor model, but I think it’s the best we have. And I think that when psychologists say, “You guys are making a model of the mind, but you’ll have to work for another lot of years before we’ll even take it seriously because you haven’t “developed it yet,” I say, “uh, that’s only because they’re stuck with theirs and they’ve grown up with them.” If you look objectively at the array of models that exist, weak as they all are, this one is way out ahead of the others.

Well, that’s what I think about computers and models.

(applause and cheering)