1. Introduction

The following content is provided under a Creative Commons license Your support will help MIT OpenCourseWare continue to offer high quality educational resources for free To make a donation or to view additional materials from hundreds of MIT courses, visit MIT OpenCourseWare at ocw.mit.edu PROFESSOR: So what I’m going to do in this course is discuss mostly ideas that are already in the book called “The Emotion Machine.” I’m sorry I used that title, and the older book called it “The Society of Mind.” The books are not quite the same They overlap a bit in material, but they’re sort of complementary I like the old one better, because the chapters are all one page long And they’re moderately independent So if you don’t like one, you can skip it The new book is much denser, and it has a smaller number of long chapters And I think it’s– over the years, I got lots of reactions from young people in high school, for example Almost all of whom liked “The Society of Mind,” and found it easy to read, and seem to understand it There are lots of criticisms by older people who maybe some of them found it harder to put so many fragments together Who knows? But most of this class, most of the things I’d like to say are in those books So it’s really like a big seminar, and my hope is that everyone who comes to this class would have a couple of questions that they’d like to discuss And if I can’t answer them, maybe some others of you can So I like to think of this as a super seminar, and normally, I don’t prepare lectures And I just start off asking if there are any questions And if there are not, I get really pissed off But anyway, I’m going to start with a series of slides So why do we need machines? And partly, there are a lot of problems Unlike most species or kinds of animals, humans have only been around a few million years And they’re very clever compared to other animals, but it’s not clear how long they will last And when we go, we might take all the others with us So there are a whole set of serious problems that are arising, because there are so many humans And here’s just a little list of things There’s a better list in a book by the Astronomer Royal Martin Rees of England Anybody know the title? AUDIENCE: “Our Final Hour.” PROFESSOR: Yes, “Our Final Hour.” It’s a slightly scary title And when I was a teenager, World War II came to an end with the dropping of two atomic– oh, this is getting terrible –two atomic bombs on Japan And I didn’t believe the first one was real, because it was in Hiroshima So I assumed that the US had somehow

made a big underwater tanker with 20,000 tons of TNT, and some few grams of radium or something, and blown it up in the harbor And first, it flew an airplane over, dropping some little thing And this was to fool the Japanese into thinking that we have an atomic bomb But when they did it, again, over Nagasaki, that wasn’t feasible And when I was in grade school, Sometimes, if I said something very bright, I would hear a teacher saying, maybe he’s another J Robert Oppenheimer Because that was the name of a scientist who had been head of the Manhattan Project, and he was, I think, three or four years earlier in grade school than I was And I thought it was very strange for a person to have a first name just being a letter rather than a name Many years later when I was at Princeton in graduate school, I met the Robert Oppenheimer, and that was a great pleasure And in fact, he took me to lunch with a couple of other people I admired, namely, Girdle and Einstein, which was very exciting Except I couldn’t understand Einstein, because I wasn’t used to people with a strong German accent But I understood Girdle just fine And after that lunch was over, I went and spent about a year learning about Turing machines, and trying to prove theorems about them, and so forth So anyway, in the course of these talks, we’ll run across a few of these people And here’s a big list of the people that I’m mostly indebted to for the ideas in the society of mind and the emotion machine The ones in blue are people I’ve actually met It would be nice to have met Aristotle, because no one really knows much about him But you really should read, just skim through some of that, and you’ll find that this is a really smart guy We don’t know if he wrote this stuff, or if it were compiled by his students, like a lot of Feynman’s writing is and Von Neumann’s writing is edited from notes by their students Anyway, the astonishing thing about Aristotle is that he seems to be slightly more imaginative than most cognitive scientists you’ll run into in the present day It would have been nice to know Spinoza, and Kant, and the others also Freud wrote 30 or 40 books Did he fall off this list? There he is I just made this list the other day, and I was looking up these people to find their birthdays and stuff Yes? AUDIENCE: Why are there no Eastern philosophers in history? PROFESSOR: Because they’re religious as far as I can see AUDIENCE: Religious? PROFESSOR: Well, who would you– would you say, Buddha? AUDIENCE: No, I mean, just Eastern thinkers thought that PROFESSOR: Name one Maybe I’ve never heard of them AUDIENCE: Confucius PROFESSOR: Who? AUDIENCE: Confucius PROFESSOR: Confucius? AUDIENCE: Or [INAUDIBLE] great thinkers from China

PROFESSOR: Well, I only know of them through aphorisms, single proverbs, but I don’t know that Confucius had a theory of thinking You think he did? AUDIENCE: There are a lot differences of thoughts, and I think they probably do have [INAUDIBLE] PROFESSOR: Well, I’ve looked at Buddhist theories, and they’re– I don’t think they would get a C plus And one problem is that there are cultures– there’s something about Greek culture Because it had science It had experiments Somebody has a theory, and they say, like Epimenides Lucretius Somewhere in the society mind, I think I quoted Lucretius about translucent objects And he says, they have the particular appearance, because the rays of light bounce many times before they get to the surface So you can’t tell where they started And I don’t find in eastern philosophy theories that say, here’s what, I think, and here’s a reason why I’ve looked at Buddhist stuff, and it’s strange lists of psychological principles Every one of which looks pretty wrong, and they make nice two dimensional diagrams But no evidence for any of them, so I don’t know whether to take it seriously AUDIENCE: I think knowledge is from observation I think you’re right that in some them probably didn’t really test it, because a lot of the ideology cannot be tested On the other hand, there are scientists– PROFESSOR: But what can’t be tested? AUDIENCE: I mean, some of the ideologies probably PROFESSOR: If they can’t be tested, why should one look at it twice? AUDIENCE: Test it in terms of something logically I don’t know Like culture, can you test culture? PROFESSOR: OK, I think this is a serious argument It seems to me that science began a little bit in China, a little bit in India In the Arabic world, they got up to the middle of high school algebra, but then– AUDIENCE: That’s the foundation PROFESSOR: What? AUDIENCE: That’s the foundation PROFESSOR: Well, but it wasn’t as good as Archimedes, who got to the beginning of calculus So if you look at most cultures, they never got to the critical point of getting theories, doing experiments, discussing them, and then throwing them out And so if you look at Buddhist philosophy, it’s 2,500 years old If you look at Greek physics, yes, Archimedes almost got calculus, and he got lots of nice principles And Buddha mentions, at some point, if you want to weigh an elephant, put him in a boat And then take the elephant out and put rocks in, until the boat sinks to the same level So there, you see a good idea But if you look at the history of the culture, if people still say, this thousand year old stuff is good, then you should say, no, it’s not AUDIENCE: By the way, same story about the elephants There’s like a story in Chinese history that has the same PROFESSOR: Sure AUDIENCE: I mean, maybe there’s no one person that [INAUDIBLE] PROFESSOR: No, but the question is, why did it stop? Why did it stop? Ancient wisdom is generally not very good, and we shouldn’t respect it for too long And that’s– AUDIENCE: [INAUDIBLE] past where everybody’s standing on the giant’s shoulders, right? PROFESSOR: No, we got rid of alchemy We got rid of– what do you call it? What’s caloric? You jump off their shoulder You don’t stay on them, so it’s good to know history But if the history doesn’t get anywhere,

then you don’t want to admire it too much Because you have to ask, why did it stop? What went wrong? And usually, it went wrong, because barbarians came in and– well, you know what happened to Archimedes Some Roman killed him Anyways– AUDIENCE: [INAUDIBLE] I’m sorry PROFESSOR: No, it’s a good question Why didn’t science happen a million years ago? Because humans are five million years old, so what took it so long? AUDIENCE: [INAUDIBLE] PROFESSOR: No, it’s more– AUDIENCE: [INAUDIBLE] PROFESSOR: Sure, OK Do you have a theory of why science didn’t develop for so long? In most cultures, it might be religion, which is a sort of science that doesn’t use evidence, and in fact, kills people who try to get it So there are systematic reasons why most cultures failed, and maybe somebody has written it Is there a book on why science disappeared, except once? It’s rather remarkable Isn’t it? After all, the idea, if somebody says something, and somebody else says, OK, let’s do an experiment to see if that’s right, you don’t have to be very bright So how come it didn’t happen all the time everywhere? Here he is AUDIENCE: I don’t know the answer to that, but I know Paul Davies has sort of an anecdote about that Or he’s exactly speculating, even in Europe when it did happen was a fluke And he gives the example of suppose an asteroid or a comet crashed in Paris in– I forget what year he gives –1150, or 1200, or something Then what? Whether it’s science [INAUDIBLE] as a thought problem PROFESSOR: History is full of flukes I’m trying to remember who wrote that nice book about the plague, some woman And she mentions that this was spread by rats and fleas or something And 30% or 40% of the population of many countries in Europe died, and the next generation had a lot of furniture The standard of living went way up, so anyway, here’s a list of disasters Oh, come on And Martin Rees is the royal astronomer and has that book about the last hour or whatever I’m making another longer list But he has lots of obvious disasters, like some high school student looks up the genetic sequence for smallpox virus has been published, and now, you can write a list of nucleotides and send it somewhere And they’ll make it for about $0.50 or $1 per nucleotide So for a couple of hundred dollars, you can make a virus or a few hundred So one possibility is that some high school student makes some smallpox only gets it wrong, and it kills everyone So there are lots of disasters like that, and no one knows what to do about that Because the DNA synthesis machinery is becoming less and less expensive, and probably the average rich private high school could afford one So there are lots of other things that could happen But one particular one is this graph, which I just made up An interesting fact is that since 1950 when the first antibiotics started to appear, as I mentioned, I was a kid in the 1940s And penicillin had just hit the stands,

and there wasn’t much of it And there was a researcher who lived a few blocks from us whose dog had cancer So its father– I don’t know what you call the owner of a dog –sneaked some penicillin out of the lab and gave it to the dog, who died anyway But he said, well, nobody’s tried penicillin on cancer yet Maybe it will work And a lot of people were mad at him, because he probably cost some human its life But he said, he might have saved a billion humans their lives, so ethics Ethicists are people who give reasons not to do things, and I’m not saying they’re wrong But it’s a funny job Anyway, since that sort of thing happened and medicine began to advance, people have been living one year longer every 12 So it’s 60 years since 1950, so that’s five of those six So they’re living six or seven years longer now than they were when I was born And somebody mentioned that curve stopped the last few years for other reasons, but anyway, if you extrapolated, you’ll find that the lifespan is going to keep increasing How much we don’t know– another problem is that you might discover enough about genetics to get rid of most of the serious diseases Maybe just 20 or 30 genes are responsible for most deaths right now And if you could fix those, which we can’t do yet, there’s no way to change a gene in a person Because invading all the cells is a pretty massive intervention, but we’ll get around that And then it might be that people suddenly start living 200 or 300 years Now at some point, the population has to slow down So you can only reach equilibrium with one child per family and probably less than that So all the work has to be done by 200 or 300-year-olds, and let’s hope they’re good and healthy So anyway, I think it’s very important that we get smart robots, because we’re going to have to stem the population And I hope people will live longer and blah, blah, blah So these robots have to be smart enough to replace most people, and how do you make something smart? Well, artificial intelligence is the field whose goal with has been to make machines that do things that we regard as smart, or intelligent, or whatever you want to call it And the idea of seriously making machines smart has roots that go back to a few pioneers, like Leibniz, who wrote about automata and that sort of thing But the idea of a general purpose computer didn’t appear till the 1930s and ’40s in some sense The first form of the general purpose computer appears really in the 1920s and ’30 with the work of a mathematician, Emil Post at NYU, who I happen to never meet But we had some friends in common, and he had the idea of production rules And basically, rule based systems and prove various theorems about them Then Kurt Girdle showed that, if you had something, like a computer or a procedure that had the right kinds of rules, it could compute all sorts of things But there were some things it couldn’t compute,

unsolvable problems, and that became an exciting branch of mathematics And the star thinker in that field was Alan Turing, who invented a very simple kind of universal general purpose computer Instead of a random access memory, it just had a tape, which it could write on, and read, and change symbols And it would go back and forth And if it’s in state x at, say, symbol y, it will print symbol z over the x and move to the left or right and just a bunch of rules like that, where it was enough to make a universal computer So from about 1936, it was sort of clear to a large mathematical community that these were great things And a couple of general purpose light computers, very simple ones, were built in the 1930s and more in the 1940s And in the 1950s, big companies started to make big computers, which were rooms full of equipment But as you know, most programs could only do some particular thing, and none of them were very smart Whereas a human can handle lots of kinds of situations And if you have one that you’ve never seen before, there’s a good chance you’ll think of a new way to deal with that and so forth So how do you make a machine that doesn’t get stuck almost all the time? And I like to use the word resourcefulness Although, I left an R out of that one Is there a shorter word? So here’s a good example My favorite example of a situation where a person is born, more or less, with a dozen different ways of dealing with something And the problem that I imagine that you’re dealing with is this My favorite example is I’m thirsty, so I see that glass of water And I do that and get it Actually, I am On the other hand, if I were here, I would never in a whole lifetime do this You never walk out a window by mistake We’re incredibly reliable, so how do I know how far it is? And that slide shows you 12 different ways that your vision system– that’s only your vision system –has to measure distances So gradients, if things are sort of blurry, then they must be pretty far away That’s sort of on a foggy day outside Here’s a situation If you assume those are both chairs of the same size, and you know that this chair is about twice as far away as that, although, you don’t– well, and how far away they are pretty much by the absolute size If you have two eyes that work well, then, if something is less than 30 feet away, you can make a pretty good estimate of its distance by focusing both eyes on some feature And your brain can tell how far apart your eyes are looking, so there’s 12 different things It’s more than you need Lots of people are missing half of those Lots of people have very poor vision in one eye Some people cannot fuse stereo images,

even though both eyes have 20/20 vision And in some cases, nobody knows why they can’t do that I think I once took a test for being a pilot, and they wanted to be sure you could do stereo vision, which seem very strange Because if you’re an airplane, and you’re less than 30 or 40 feet away from something, it’s too– you could use stereo But it’s too late Anyway, that’s interesting See if you can think of an example, where a person has even more 12 of these But it’s pretty amazing, isn’t it? It’s more redundancy This is too hard to read, but somehow, I found it in Aristotle essay the idea that you should represent things in multiple ways You might describe a house One person might describe a house as a shelter against destruction by wind, rain, and heat Another might describe it as a construction of stones, bricks, and timbers But a third possible description would say, it was in that form in that material with that purpose So you see there’s two different descriptions One is the functional description It’s a shelter The second one is a structural description, how it’s made And Aristotle says, which is the better description? And he dismisses the material one or the functional one is not rather the person who combines both in a single statement And then I found a paragraph by Fineman, who says, every theoretical physicist who is any good knows six or seven different ways to represent exactly the same physics And you know that they’re all equivalent, but you keep them all in your head hoping that they will give you different ideas for guessing I should put more dots Anyway, that whole argument is to say that the interesting thing about people is that they have so many ways to do things, and perceive things, and think of things And in some cases, we even know that there are different parts of the brain that are involved in one aspect or another of constructing those different representations or descriptions If you look at one of my favorite books, it weighs about 20 pounds It’s the book on the nervous system by Kandel and Schwartz And the index to that book is quite a lot of pages long, and it mentions 400 different structures in the brain So the brain is not like the– well, I shouldn’t make fun of the liver Because for all I know, the liver has 400 different many processes for doing things But the brain has distinguishable areas that seem to perform several hundred different functions And with a microscope, at first, they all look pretty much the same But if you look closely, you see slightly different patterns of how the most layers of the cortex of the brain, most parts of it have six layers, and each has a population of different kinds of cells There are a lot of cross connections up and down and sideways to other They’re arranged in columns of between 400 and the 1,000 cells, and you have a couple of million of those And there are lots of differences between the columns in different areas, and we know some of the functions

In most cases, we don’t know much about how any of them actually work with the main exception of vision, where the functions of the cells in the visual cortex are fairly well understood at low levels So we know how that part of the brain finds the edges and boundaries of different areas, and textures, and regions of the visual field But we do not know even a little bit about how the brain recognizes something as a chair, and an overhead projector, and a CRT screen, and that sort of thing The kind of question that I got interested in was, how can you have a system, which has a very large number of different kinds of computers? Each of which by itself might be relatively simple or might not, I suppose And how could you put them together into a larger system, which could do things, like learn language, and prove theorems, and convince people to do things that they would never have dreamed of doing five minutes earlier, and stuff like that? Now the first sort of things I was interested in was, in fact, how to simulate simple kinds of nerve cells Because in the 1950s, there was about almost 100 years, really more like 50 years of science discovering things about neurons and nerve cells, the axons, and dendrites that they use to communicate with other neurons So if you go back to 1890, you find a few anatomists discovering some of the functions or connections of neurons in the brain And you find a few experimental physicists There was no oscilloscope yet, but there were very high gain galvanometers, which could detect pulses going along a nerve fiber And by 1900, it was pretty clear that part of the activity in a nerve cell was chemical and part was electrical And by 1920 or ’30 with the cathode ray tube appearing mostly because of television, but it became possible to do a lot of neurophysiology by sticking needles in brains The vacuum tube appears around 1900, and you can make amplifiers that can see millivolts and then microvolts So in the beginning of the 20th century, there was lots of progress By 1950, we knew a lot about the nervous system, but we still don’t know much about how you learn something in the brain It’s quite clear that the things called synapses are involved The connections between two neurons become better at conducting nerve impulses under some conditions, but no one knows how higher level knowledge is represented in the brain yet And the Society of Mind book had a lot of theories about that And in particular, there was a theory called k line’s, knowledge lines, or something that came partly from me and partly from a couple of other researchers named David Waltz and Jordan Pollock That’s a sort of nice theory of how neural networks might remember higher level concepts And for some reason, although that kind of work is from around 1980, which is 30 years ago, it has not hit the neuroscience community

So if you look at the emotion machine book or the society minded in Amazon, you might run across a review by a neurologist named Richard Restak, who says that Minsky makes up a lot of concepts, like K-lines, and micronemes, and stuff like that, that nobody’s ever heard of And there’s no evidence for them, and he ignores the possibility that it isn’t the nerve cells in the brain that are important But the supporting tissues called glia, which hold the neurons up and feed them And he goes on for a couple of insane paragraphs It’s very interesting, because it doesn’t occur to him that you can’t look for something, until you have the idea of it So here is this 30-year-old idea of K-lines, and go and ask your favorite neurologist, neuroscientist what it is And he said, oh, I think that’s some AI thing, but where’s the evidence for it? What do you suppose is my reaction to that? Who is supposed to get the evidence? So it seems to me that there’s a strange field in neuroscience, which is that it doesn’t want new ideas, unless you’ve proved them So I try to have conversations with them, but get somewhat tired of it Anyway, in this course, I’m taking the opposite approach, which is that we don’t want a theory of thinking We want a lot of them, because probably, psychology is not like physics What’s the most wonderful thing about physics? The most wonderful thing is that they have unified theories There wasn’t much of a unified theory, until Newton, and he got these three wonderful laws One was the gravitational idea that bodies attract each other with a force that’s the inverse square of the distance between them Another is that kinetic energy is conserved I forget with the third one is Oh, equal reaction is equal and opposite If two things collide, they transfer equal amount of momentum to both There was a little problem up to Newton’s time Galileo got some of those ideas, and my impression from reading him is that he has a dim idea that there are two things around There’s kinetic energy, which is MV– oops, momentum is MV And there’s kinetic energy, which is MV squared, and he doesn’t have the clear idea that there are two different things here And you can’t blame him I would think– you wouldn’t think that two quantities would combine in two different ways to make two important different concepts Well, that got clear to Newton somehow, and Galileo is a bit muddled Although, he gets almost all the consequences of those things right, but he doesn’t get the orbits and things to come out Anyway, what’s happened in artificial intelligence, like most fields, is that people said, well, let’s try to understand thinking and psychology And let’s use physics as our model, so what we want is to get a very small number of universal laws And a lot of psychologists struggled around to do that, and then they gradually separated So that there were some psychologists, like Bill Estes, who worked out some very nice mathematical rules for reinforcement based learning, got a simple rule If you designed an experiment right, it predicted pretty well how many trials

it would take a rat, or a pigeon, or a dog, or whatever to learn a certain thing from trial and error And Este’s got a set of four or five rules, which looked like Newton’s laws And if you designed your experiment very carefully and shielded the animal from noise and everything else, which is what a physicist would do for a physics experiment, the reinforcement theories got some pretty good models of how to make a machine learn But they weren’t good enough So here’s a whole list of things that happened in the early years of cognitive psychology when people were trying to make theories of thinking, and they were imitating the physicists By physics envy to borrow a term of Freud, the idea is, can you find a few simple rules that will apply to very broad classes of psychological phenomena? And this led to various kinds of projects Lots of neural network, and reinforcement, and statistical based methods led to learning machines that were pretty good at learning in some kinds of situations And they’re becoming very popular, but I don’t like them Because, if you have a lot of variables, like 50 or 100, then to use a probabilistic analysis, you have to think of all combinations of those variables Because if two of them are combined in something, like a exclusive or a manner, you know, I just put the light pen in a pocket It’s either in the left pocket or a right pocket It can’t be both That’s an x or That will cause a lot of trouble to a learning machine And if there are a hundreds variables, there’s no way you could decide which of the two to the 100th Boolean combinations of those variables you should think about So lots of statistical learning systems are good for lots of applications But they just won’t cut it to solve hard problems, where the hypothesis is a little bit complicated and has seven or eight variables with complicated interactions Most statistical learning people assume that, if you get a lot of partial ones, then you can look for combinations of ones that have high correlations with the result. Then you can start combining them, and things get better and better However, mathematically, if an effect you’re looking for depends on the exclusive or of several variables, there’s no way to approach that by successive approximations If any one of the variables is missing, there won’t be any correlation of the phenomenon with the others Anyway, that’s a long story, but I think it’s worth complaining about Because almost all young people who start working on artificial intelligence look around and say, what’s popular? Statistical learning, so I’ll do that That’s exactly the way to kill yourself scientifically You don’t want to get the most popular thing You want to say, what are my really good at that’s different? And what are the chances that would provide another thing? End of long speech Another problem in the last 30 years– and as you’ll see during my lectures, I think a lot of wonderful things happened between 1950 when the idea of AI first got articulated in the 1950s And then the 20 years after that from 1960 to 1980, a lot of early experiments– and I’ll show you some of them

–looked very promising In fact, they may be– here we go 1961, Jim Slagle was a young graduate student here at MIT He was blind He had gotten some retinal degeneration thing in his first or second year of high school He was told that he would lose all his vision, and there was no treatment or hope So he learned Braille while he could still see And when he got to MIT, he was completely blind, but there was a nice big parking lot in technology square And he would ride a bicycle And people, like Sussman, and Winston, and whoever was around, would yell at him, telling him where the next obstacle would be Jim got better and better at that, and nothing would stop him And he decided he would write a program that– oh, I wrote a program that would take any formula and find its derivative It was really easy, because there were just about five rules Like if there’s a product UV, then you compute U times the derivative of V plus V times U of DV plus DVU So I wrote a 20 line list program that did all the algebraic expressions, and what it would do is put Ds in the right place And then it would go back through the expression again Wherever it saw a D, it would do the derivative of the thing after that and nothing to it So Slagle said, well, I’ll do integrals And we all said, well, that’s very hard Nobody knows how to do it And in fact, in Providence at the home of the American Mathematical Society, there is a big library called the Bateman Manuscript Project, which has been collecting all known integrals for 100 years And when anybody finds a new integral that they can integrate in closed form, they send the formulas to the Bateman Manuscript Project, and some hackers there develop ways to index it So if you had an integral, and you didn’t know how to integrate it, you could look it up And that was pretty big I should say that Slagle succeeded in writing a program that managed to do all of the kinds of integrals that one usually found on the first year calculus course at MIT and got an A in those He couldn’t do word problems And the uncanny thing is that, if it was a problem that usually took a MIT student five or 10 minutes, Slagle’s program would take five or 10 minutes It’s running on an IBM 701 with a 20 millisecond cycle time It’s incredibly slow You can type almost that fast and 16 K of words of memory So there’s no significance whatever to this accident of time It would now take a microsecond or so It would be 1,000 million times faster than a student Quite remarkable I don’t have a slide Joel Moses, then Slagle went and graduated Joel Moses was another student, who is– is he provost now or what?

He got tired of it A terrific student, and he set up a project called maxima for project Max Symbolic Algebra and got several people all over the country working on integration And at some point, a couple of them, Bobby Caviness and– forget the other one– found a procedure that could, in fact, integrate everything– every algebraic expression that has a– can be integrated in closed form I forget the couple of constraints on it And that became a widely used system It ultimately got replaced by Steven Wolfram’s Mathematica But Maxima was sort of the world-class symbolic mathematician for quite a few years And Moses mentioned to me he had read Slagle’s program thesis And it took him a couple of weeks to understand the two pages of– or three pages of– Lisp that Slagle had written Because being blind, Slagle had tried to get the thing into as compact a form as possible But that’s symbolic It’s too easy It was in a more ambitious one, which was, three years later, Dan Bobrow, who is now a vice president doing something at Xerox– and it solved problems like this The gas consumption of my car is 15 miles per gallon The distance between Boston and New York is 250 miles What is the number of gallons used on a trip between Boston and New York? And it chomps away and solves that It has about 100 rules It doesn’t really know what any of those words mean But it thinks that the word “is” is equals The distance between– doesn’t care what Boston and New York is It has a format thing which says the distance between two things And it never bothers to– you see, because the phrase “Boston and New York” occurs twice in the example, it just replaces that by some symbol It was fairly remarkable And generally, if you had an algebra problem, and you told it to Bobrow, Bobrow could type something in, and it would solve it If you typed it in, it probably wouldn’t But it was– it had more than half a chance, or less– about half a chance So it was pretty good And if you look at an out-of-print book I wrote called– I compiled, called– Semantic Information Processing, most of Bobrow’s program is in that So that’s 1964 I’ll skip Winograd, which is, perhaps, the most interesting program This was a program where you could talk to a robot that– I don’t have a good picture in this slide But they’re a bunch of blocks of different colors They’re all cubes in the– or rectangular blocks And you can say, which is the largest block on top of the big blue block? And it would answer you And you could say, put the large red block on top of the small green block, and it would do that And Winograd’s program was, of course, a symbolic one We actually built a robot And I guess we built it second Our friends at Stanford built a robot And they imported Winograd’s program And they had the robot actually performing these operations that you told it to do by typing And it was pretty exciting My favorite program in that period

was this one, because it’s so psychological This is called a geometrical analogy test And it’s on some IQ tests A is to B as C is to which of the following five? And Evans wrote a set of rules which were pretty good at this Did as well as 16-year-olds And it picks this one And if you ask it why, it says something like, I don’t have a reason It moves the largest object down or something like that, makes up different reasons So you see, in some sense, we’re going backwards in age Because we’re going from calculus, to algebra, to simple analogies Oh, there it is That’s one where the largest object moves down I don’t know why I have two of them These are for another lecture OK So that was a period in which we picked problems that people considered hard, because they were mathematical But when you think about it more, you see, well, those math things are just procedures And it’s once you know what Laplace, and Gauss, and those mathematicians– Newton and people– did, you can write down systematic procedures for integrating, or for solving simultaneous algebraic constraint equations, or things like that And so there’s very little to it So in some sense, if you look at the– what you’re doing in math in high school, in education, you’re going from hard to easy It’s just that people aren’t– most people aren’t very good at obeying really simple rules, because it’s so hideously boring or something So we gradually started to ask, well, why can’t we make machines understand everyday things, and the things that everyone regards as common sense, and people can do so you don’t need machines to do them? And one of my favorite examples is, why can you pull something with a string but not push? And there’s been a lot of publicity recently about that interesting program written at a group at IBM called Watson, which is good at finding facts about sportspeople, and celebrities, and politics, and so forth But there’s no way it could understand why you could push– pull something with a string but not push And I don’t know of any program that has that concept or way of dealing with it So that’s what I got interested in And starting around the– maybe the middle 1970s or late 1970s, several of us started to stop doing the easy stuff and try to make theories of how you would do the kinds of things that people are uniquely good at I don’t know if animals– well, I don’t know I’m sure a monkey wouldn’t try to push anything with a string Maybe it does it very quickly, and you don’t notice And one aspect of commonsense thinking is going right back to that idea of vision having a dozen different systems, is that, whatever

a person normally is doing, they are probably representing it in several different ways And here’s an actual scene of two kids named Julie and Henry who are playing with blocks It’s pretty hard to see those blocks And you can think that Julie is thinking seven thoughts I’d like to see a longer list Maybe a good essay would be to take a few examples and say, what are the most common micro-worlds? See physical, social, emotional, mental, instrumental– whatever that is– visual, tactile, spatial She’s thinking all these things What if I pulled out that bottom block? You can’t see the tower very well Should I help him or knock is tower down? How would he react? I forgot where I left the arch-shaped block That was real It’s somewhere over here But I don’t think we could– maybe it’s that I don’t know I remember, when it happened, she mentioned that she reached around, and it wasn’t where she thought it was So commonsense thinking involves this– in most cases, I think, several representations I don’t know if it’s as many as seven, or maybe 20, or what But that’s the kind of thing we want to know how to do OK, I think I’ll stop, and we’ll discuss things But in the next lecture, I’ll talk about a model of how I think thinking works What’s the difference between us and our ancestors? We know we have a larger brain But if you think about it, if you took the brain that you already had and say– trying to remember the name of the little monkey that looks like a squirrel, jumps around in trees Anybody know what– AUDIENCE: Spider monkey? PROFESSOR: What? AUDIENCE: Spider monkey? PROFESSOR: Maybe It’s a squirrel-like thing You wouldn’t know it was a monkey till you took a close look AUDIENCE: Lemur, maybe? PROFESSOR: Maybe Lemur? I don’t– I forget I’ll have to– anyway, if you just made the brain bigger, then the poor animal would be slower, and heavier, and would need more food, and take longer to reproduce The joke about difficulty to give birth– I don’t know if any animal has the problem that humans have A lot of people die and so on So how did we evolve new ways to think and so forth? And my first book, The Society of Mind, had this theory that maybe we evolved in a series of higher and higher levels, or management structures, built on the earlier ones And this particular picture suggests that I got this idea from Sigmund Freud’s early theories There’s been a lot of Freud bashing recently You can look on the web I forget the authors But there are a couple of books saying that he made up all his data, and there’s no evidence that he ever cured anyone, and that he lied about all the data mentioned in his 30 or 40 books, and so forth AUDIENCE: Also [INAUDIBLE] PROFESSOR: Yes, right But the funny part is that if you look at his first major book– 1895– called The Interpretation of Dreams, it sort of outlines his theory that most of thinking is unconscious, and it’s processes you can’t get access to And it has a little bit about sex,

but that’s not a major feature And it’s just full of great ideas that the cognitive psychologists finally began to get in the 1960s again and never give credit to Freud So he may well have made up his data But if you have a very good theory and nobody will listen to you, what can you do? His friend Rudolf Fliess listened to him And there was another paper on how the neurons might be involved in thinking, which was also written around 1895, but never got published till 1950 by– forget who– called “Project for a scientific psychology.” And it’s full of ideas that, if they had been published, might have changed everything Because– anyway, what’s on your mind? Who has– what would you like to hear about? Or who has another theory? AUDIENCE: I’ve got a question PROFESSOR: Great AUDIENCE: So earlier, you talked a little bit about how we don’t really see the neuroscience, all of these things like K-lines, et cetera Do you think it’s because they’re just really hard to find, or no-one’s actually looking for them? PROFESSOR: Well, Restak’s review says, he uses vague, ill-defined terms like K-line, and microneme, and a couple of others, and frame, and so forth They’re very well-defined They’re defined better– I mean, when he talks about neurotransmitters, it’s as though he thinks that chemical has some real significance Any chemical would have the same function as any other one provided there’s another receptor that causes something to happen in the cell membrane So you don’t want to regard acetylcholine or epinephrine as having a mental significance It’s just a– it’s just another pulse, but very low-resolution And yes, a neurochemical might affect all the neurons a little bit, and raise the average amount of activity of some big population of cells, and reduce the average activity of some others But that’s nothing like thinking That’s like saying, in order to understand how a car works– what’s the most insulting thing I could say? Or to understand how a computer works, you have to understand the arsenic, and phosphorus, and/or– what’s the other one? You have to understand these atoms that are– what? AUDIENCE: Germanium PROFESSOR: Yeah, well, that’s the matrix So there are these one part in a million impurities And that’s what’s important about a computer, isn’t it, the fact that the transistor has gain and so forth Well, no, the trouble with the computer is the transistors That’s why practically every transistor in a computer is mated to another one in opposite phase to form a flip-flop whose properties are exactly the same, except one in a quadrillion times In other words, everything chemical about a computer is irrelevant And I suspect that almost everything chemical about the brain is unimportant except that it causes– it helps to make the columns in the cortex, which are complicated arrangements of several hundred cells, work reliably Whereas the neuroscientist is looking for the secret in the sodium When a neuron fires, the important thing is that that lets the sodium in and the potassium out or vise-versa– I forget which– at 500 millivolts– really quite a colossal event But it has no significant It’s only when it’s attached to a flip-flop, or to something like a K-line, which has an encoder and decoder of a digital sort every few microns of its length

that you get something functional So the trouble is, the poor neuroscientists started out with too much knowledge about the wrong thing The chemistry of the neuron firing is very interesting, and complicated, and cute And in the case of the electric eel, you know what happened there The neuron synapse, it got rid of the next neuron And it just– in the electric eel, you have a bunch of synapses, or motor end plates they’re called, in series So instead of half a volt, if you have 300 of those, you get 150 volts I think the electric shock that a electric eel can give you is about 300 volts And this can cause you to drown promptly if you are in the wrong wave when it happens to bump into you I don’t know why I’m rambling this way You’re welcome to study neuroscience But please try to help them instead of learn from them [LAUGHTER] Yeah? They just don’t know what a K-line is And that’s a paper that’s been widely read It’s published in 1980, and Restak says ill-defined And I guess he couldn’t understand it Yep? Yeah? AUDIENCE: Why is there no trying to make the neuroscientists trying to find this in the human mind? Why don’t we just, as computer scientists, program the K-lines and try to [INAUDIBLE]? This is the human mind, and we can reproduce it Why is not– is that not widespread into the computer scientist field? PROFESSOR: Well, there are– I’m surprised how little has been done There’s– Mike Travers has a thesis, Tony Hearn There are three master’s theses on K-lines They sort of got them to work to solve some simple problems But I’d go further I’ve never met a neuroscientist who knows the pioneering work of Newell and Simon in the late 1950s So there’s something wrong with that community They’re just ignorant They’re proud of it Oh, well I spent some time learning neuroscience when I was– I once had a great stroke of luck When I was a– I guess I was a junior at Harvard And there was a great new biology building that was just constructed You probably know, it’s a great, big thing with two rhinoceroses What are those– what are those two huge animals? So this building was just finished and half occupied, because it was made with a future So I wandered over there, and I met a professor named John Welsh And I said, I’d like to learn neurology And he said, great, well, I have an extra lab Why don’t you– why don’t you study the crayfish claw? I said, great So he gave me this lab, which had four rooms, and a dark room, and a lot of equipment, and nobody there And he had worked on crayfish So there was somebody who went, every week, up to Walden Pond or somewhere, and caught crayfish, and bring them back And I was a radio amateur hacker at the time So I was good at electronics So I got my crayfish And Welsh showed me how to– the great thing about this preparation is you can take the crayfish, and if you– claw– and if you hold it just right, it goes, snap It comes off Grows another one– takes a couple of years And then there’s this white thing hanging out, which is the nerve And it turns out it’s six nerves, one big one and a few little ones And if you keep it in Ringer’s solution, whatever that is, it can live for several days So I got a lot of switches, and little inductors, and things, and made a gadget, and mounted this thing

with six wires going to these nerves And then I programmed it to reach down and pick up a pencil like that and wave it around Well, that’s obviously completely trivial And all the neuroscientists came around, and gasped, and said, that’s incredible How did you do that? [LAUGHTER] They had never thought of putting the thing back together and making it work Anyway, it was– I’m always reminding myself that I’m the luckiest person in the world Because every time I wanted to do something, I just happened to find the right person And they’d give me a lab I got an idea for a microscope And it was this great professor, Purcell, who got the Nobel Prize after a while And he said, that sounds like it would work Why don’t you take this lab? It was in the Jefferson Anyway– yeah? AUDIENCE: I think part of the reason that you don’t see experimental neuroscience on things like K-lines is that neurons are long and thin So if you want to do an experiment to actually measure a real neural network, you have to trace structures with, roughly, maybe tens of nanometer resolution But you need to trace them over what might be a couple, or even tens, of millimeters to start to– and you need to do this for thousands and thousands of neurons before you could get to the point of seeing something like a K-line and understanding it So it’s just a massive data acquisition and processing problem PROFESSOR: Oh, but they’re doing that AUDIENCE: They’re starting to try to PROFESSOR: But they don’t have– they don’t know what to look for Maybe you don’t have to do so much Maybe you just have to do a few sections here and there and say, well, look, there were 400 of these here Now there’s only 200 It looks like this is the same kind Maybe you don’t have to do the whole brain AUDIENCE: No, but I mean even getting a single neuron is big Because it might get down to– you need to be looking at electron micrographs of brains that are sliced at about 30-millimeter– sorry, excuse me, 30-nanometer slices So even just having a single person reconstruct a single neuron takes– might take weeks PROFESSOR: Well, I don’t know Maybe a bundle of K-lines is half a millimeter thick AUDIENCE: Oh, so If you actually do some larger-scale structure to start off looking at, yeah PROFESSOR: Why not? I just think they have no idea what to look for I could give you 20 of those in five minutes, but nobody’s listening AUDIENCE: So I guess you need to know what it looks like before you can look for it at that scale PROFESSOR: What scale? AUDIENCE: I don’t know I mean, they know what neurons look like So you know– PROFESSOR: Yeah AUDIENCE: You know what to look for if you’re saying a neural net level PROFESSOR: I’m saying you may only have to look at the white matter AUDIENCE: Oh, all right PROFESSOR: Ignore the neurons Because the point of K-lines is, where do these go? And what goes into them and out? I don’t know It’s just this idea, let’s map the whole brain, 100 billion things And then people like Restak says, oh, and there’s 1,000 supporting cells for each neuron He’s just glorying in the obscurity of it rather than trying to contribute something Anyway, if you run into him, give him my regards [LAUGHTER] I really wonder how somebody can write something like that Yes? AUDIENCE: Excuse my ignorance, but what is a K-line? PROFESSOR: The idea is that– suppose one part of the brain is doing something, and it’s in some particular state that’s very important, like– I don’t know, that– like I’ve just seen a glass of water Then another part of the brain would

like to know there’s a glass of water in the environment And I’ve been looking for one So I should try to take over and do something about that Now at the moment, there is no theory of what happens in different parts of the brain for a simple thing like that to happen, no theory at all, except they use the word “association.” Or they talk about, what are the purposeful neurons? Goal– forget OK, so my theory is that there are a bunch of things which are massive collections of nerve fibers, maybe a few hundred or a few thousand And when the visual system sees an apple, it turns on 50 of those wires And when it sees a pear, it turns on a different 100 or 50 of those wires But about 20 of them are the same, so forth In other words, it’s like the edge of a punched card Have you ever seen a card-based retrieval system? If you have a book that has– suppose it’s about physics, and biology, and Sumatra And a typical 5 by 8 card has 80 holes in the top edge So what do you do is, if it’s Sumatra, you punch eight of these holes at random, particular set They’re assigned to Sumatra And then if it’s– I forget what my first two examples were But you punch eight or 10 holes for each of the other two words So now there are 24 punches Only probably four or five of them are duplicates So you’re punching about 20 holes And now, if something is looking for the cards that have– were punched for those three things, even if there are 30 or 40 other holes punched in the card, you stick your 20 wires through the whole deck and lift it up And only cards fall out that had those three categories punched for So you see, even though you had 80 holes, you could punch combinations of up to a million different categories into that And if you have to put a bunch of wires through, you’ll get all of the ones that were punched for those cate– the categories you’re looking for And you might get three or four other cards that will come down also Because all of the eight holes were punched for some category by accident Do you get the picture? I’ll send you a reference It was invented by a– in the 19– early 1940s by a Cambridge scientist here named Calvin Mooers and was widely used in libraries for information retrieval until computers came along But anyway, that’s the sort of thing that you could look for in a brain if you had the concept in your head of Zastocoding But I’ve never met a neuroscientist who ever heard of such a thing So you have this whole community which doesn’t have a set of very clear ideas about different ways that knowledge or symbols could be represented in neural activity So good luck to them when they get their big map They’ll still have to say, what do I do with 100 billion of these interconnections? Yeah? AUDIENCE: What are your thoughts about the current artificial intelligence research at MIT, such as Winston’s genesis project? PROFESSOR: I think Winston is just about one of the best ones in the whole world I don’t know any other projects that are trying to do things on that higher level of commonsense knowledge He’s just lost a lot of funding So one problem is, how do you support a project like that? Have you followed it? I don’t know if there’s a recent summary of what they’re doing AUDIENCE: [INAUDIBLE] PROFESSOR: We used to write a big– a new book every year called the progress report

The nice thing is that we never wrote– we had a very good form of support from ARPA, or DARPA, which was, every year, we’d– every year, we’d tell them what we had done They didn’t– they didn’t want to hear what we wanted to do And things have turned the opposite So what would happen is, every year, we’d say, we did these great things, and we might do some more Went on for about 20 years And it was– and then it fell apart One thing– it’s a nice story– there was a great liberal senator, Mike Mansfield And unfortunately, he got the idea that the defense department was getting too big and influential So he got Congress to pass a law that ARPA shouldn’t be allowed to support anything that didn’t have direct military application And Congress went for this And all of a sudden, a lot of research disappeared, basic research It didn’t bother us much Because we made up applications and said, well, this will make a military robot that will go out and do something bad I don’t remember ever writing anything at all Because– but anyway, around 1980, the funding for that sort of thing just dried up because of this political accident It was just an accident that ARPA, mainly through the Office of Naval Research, was funding basic research And that was a bit of history If you look back at the year 1900 or so, you see people like Einstein making these nice theories But Einstein wasn’t a very abstract mathematician So he had a mathematician named Herman Weyl polishing his tensors and things for him And Herman Weyl’s son, Joe, was at the Office of Naval Research in my early time And that office had spent a lot of secret money getting scientists out of Europe while Hitler was marching around and sending them to places like Princeton and other forms of heaven in the– in Cambridge And again, one of the reasons I was lucky is that I was here And all these– you know, if you had a mathematical question, you could find the best mathematician in the world down the block somewhere And Joe Weyl was partly responsible for that And the ONR was piping all that money to us for work on early AI So it was a very sad thing of the– maybe the most influential liberal in the US government actually ruined everything by accident ARPA changed its name to DARPA It was Advanced Research Projects Agency And it had to call itself Defense Advanced Research Projects Agency Yeah? AUDIENCE: My question is, do you think the achievement of artificial intelligence is inevitable? Or is there an obstacle that we’re just never going to be able to overcome? PROFESSOR: Well, Christianity wiped out science That might happen tomorrow Only choose your religion AUDIENCE: Even favorable circumstances PROFESSOR: It’s a hard problem The number of people working on advanced ideas in AI

has gotten smaller and smaller as the– f right now, the– around 1980, rule-based systems became popular And there were lots of things to do Right now, statistical-based inference systems are becoming popular And as I said, these things are tremendously useful But the problem is, if you have a statistical system, the important part is guessing what are the plausible hypotheses and then making up the– then finding out how many instances of that are correlated with such and such So it’s a nice idea But the hard problem is the abstract, symbolic problem of, what sets of variables are worth considering at all when there are a lot of them? So to me, the most exciting projects are the kind that Winston is developing for reasoning about real-life situations And the one that Henry Lieberman– would you stand up, Henry? Lieberman runs a world-class group that’s working on commonsense knowledge and informal reasoning And it seems to me that that’s the critical thing that all the other systems will need In the meantime, there are people working on logical inference, which has the same problem that statistical inference has– namely, how do you guess which combinations of variables are worth thinking about? Then it seems to me that the statistics isn’t so important In fact, there’s a great researcher named Douglas Lenat in Austin, Texas who once made an interesting AI system that was good at making predictions and guessing explanations for things And it was sort of like a probabilistic system It had a lot of hypotheses And every time one of them was useful in solving a problem, it moved it up one on the list So Lenat’s thing never used any numbers It didn’t say, this is successful 0.73 of the time, and now it’s successful 0.7364825 of the time What it would do is, if something was useful, it would move it up past another hypothesis Every now and then, it would put a new one in Well, if you’re doing– if you’re trying to solve the problem, what do you need to know? You want to know, what’s the most use– what’s the most likely to be useful one, and try that You don’t care how likely it is to be useful as long as it’s the most, right? I mean, if it’s one in a million, maybe you should say, I’m getting out of here I don’t– I shouldn’t be working in this field at all, or get a better problem But Lenat’s thing did rather wonderfully at making theories by just changing the ranking of the hypotheses that was considered– no numbers It did something very cute He gave it examples of arithmetic And it actually– it was a rather long effort And it actually learned to do some arithmetic And it invented the idea of division and the idea of prime number, which was some number that wasn’t divisible by anything It decided that 9 was a prime– didn’t do much harm And it crept along And it got better and better And it invented modular arithmetic by accident at some point And it’s a PhD thesis A lot of people didn’t believe this PhD thesis, because Lenat lost the program tape [LAUGHTER] So he was under some cloud of suspicion for people thinking he might have faked it But who cares? Anyway, I think there’s a lesson there, which is that, let’s start with something that works And then, if it’s really good, then hire a mathematician who might be able to optimize it

a little But the important thing was the order And a good statistical one might waste a lot of time Because here’s this one that’s 0.78 And here’s this one that’s 0.56 And it’s the next one down And you get a lot of experience And it goes up to 0.57 and 0.58 And it never– you know, might be a long time before it gets past the other one, because you’re doing arithmetic Whereas in Lenat’s, it would just pop up past the other one Then it would get tried right away And if it were no good, it would get knocked down again Who cares? So it’s a real question of– I don’t know Mathematics is great And I love it And a lot of you do But there should be a name for when it’s actually slowing you down and wasting your time because there’s a better way that’s not formal Yeah? AUDIENCE: Isn’t there a saying there are people who know the price of everything and the value of nothing PROFESSOR: That’s very nice Yeah? AUDIENCE: I know you’re also a musician So I have a music-related question What do you think is the role of music? Like, why do all cultures have it? PROFESSOR: I have a paper about it Oh, OK I’ve been trying to revise it, actually But it’s a strange question, because there is music everywhere On the other hand, I have several friends who are amusical And so I have this theory that music is a way of teaching you to represent things in an orderly fashion and stuff like that Well, I have three of my colleagues who aren’t musical, but they dance So they may– it may be that– I don’t know the answer It’s interesting The theory– the first theory in my paper is that when you have a lot of complicated things happening, then the only way to learn is to represent things that happen, and then look at the differences between things that are similar, and then try to explain the differences, right? I mean, what else is there? Maybe there’s something else So in order to become intelligent and understand things, you have to be able to compare things And to me, the most important feature of what’s called music is that it’s divided into measures– bah, bah, bah, bah, bah, bah, bah, bah, bah, bup, bah, bah And measures are the same number of beats, or whatever they are And so now you can say da, da, da, da, da, da, da, da, da, da, da, da, da What’s the different? You changed the eighth notes in the second one, the last four eighth notes– no, the two before last– to a quarter note So you’re taking things that were in different times, and you’re superimposing those times And now you can see the difference And the reason you can see the difference is that you have things called measures And the measures have things called beats And so things get locked into very good frames Now there’s some Indian music which has 14 measures for a phrase And some of the measures go seven and five And I can make no sense of that stuff whatever And I’ve tried fairly hard, but not very So I don’t understand how Indians can think Any of you can handle Indian music? AUDIENCE: I guess, just to add on what you said about this, my favorite quote from your paper on music, mind, and meaning is the one about what good is music, about how kids play with blocks to learn about space, and people play with music to learn about time

And I think, in that sense, both music and dance are different ways that people can arrange things in time And in a sense, improvisatory and improvisatory movement are both ways of– different blocks, if you will, in time as opposed to space PROFESSOR: Mm-hmm Yeah, my friends who seem amusical, they probably– maybe there’s something different about their cochlea Or maybe they have absolute pitch in some sense, which is a bad thing to have Because if you’re listening to a piece composed by a composer who doesn’t have absolute pitch, then you’re reading all sorts of things into the music that shouldn’t be there And if you’re– and the opposite would be true I read music criticism sometimes And maybe the reviewer says, and after the second and third movement, he finally returns to the initial key of E flat major– what a relief Well, I once had absolute pitch for a couple of weeks, because I ran a tuning fork in my room for a month And I didn’t like it [LAUGHTER] Because you can’t listen to Bach anymore Oh, well It’s a good question, why do people like music And I don’t know any other paper like mine If you ever find one, I’d like to see it Because if you go to a big library, there are thousands of books about music And if you open one, it’s mostly Berlioz complaining that somebody wouldn’t give him enough money to hire a big enough chorus But I’ve found very few books about music itself Yes? AUDIENCE: It’s not about music anymore Is that all right? OK Do you think that having a body is a necessary component of having a mind? Could you do just as well just as a– you know, sort of a simulated creature? PROFESSOR: Oh, sure, you could– AUDIENCE: And have all the things? PROFESSOR: Simulation– I think a mind that’s not in a world wouldn’t have much to do It would have to invent the world And I don’t see why it couldn’t But you might have to give it a start, like the idea of three or four dimensions But can’t you– what happens if you sit back and just think for a while? You wouldn’t know if your body had disappeared for– would you? There also is a strange idea about existence and– why do you think there’s a world? One of the things that bugs me is people say, well, who created it? And that can’t make any sense Because this is just a possible world Suppose there are a whole lot of possible worlds, and there’s one real one How could you ever– how could you possibly know which one you’re in? And then you could say, well, didn’t someone have to make it? And what’s the next thing you’d ask? Well, who made the maker? So the body-mind thing– seems to me that once you have a computer, it can be its own world It just can sit The program can spend half the time simulating a world and half the time thinking about what it’s like to be in it Yeah? AUDIENCE: Do you have a current theory of existence? PROFESSOR: Yes, it’s an empty concept It’s all right to say this bottle exists, because that’s saying this bottle is in this universe But what would it mean to say the universe exists? The universe is in the universe? So there’s something wrong with thinking about–

so there are only possible worlds There’s no– it doesn’t make any sense to pick one of them out and say that’s the real one Yeah? AUDIENCE: So it’s existence is relative? PROFESSOR: Yeah, you don’t say, this is the world I’m in But you shouldn’t say– that doesn’t mean it exists Like, 2 is in the set of even numbers But what’s the set of even numbers in? It doesn’t stop anywhere Yeah? Lots of worlds AUDIENCE: So is mathematics [INAUDIBLE] worlds But physics, it only explains the current world? Or I don’t know how you compare these two subjects PROFESSOR: Well, you can’t tell Because five minutes from now, everything might change So nothing ever explains anything You just have to take what you’ve got and make the best of it Yeah? AUDIENCE: So this relationship means systems knowledge in artificial intelligence? PROFESSOR: Which knowledge? AUDIENCE: Systems, basically, in general PROFESSOR: Well, there are people who talk about systems theory, but I’m not sure that it’s well-defined AUDIENCE: Right, exactly PROFESSOR: Artificial intelligence means, to me, making a system that is resourceful and doesn’t get stuck And so if you have a system– and also, it’s an– how do you put it? Some definitions are not stationary, like what’s popular Popular is what’s popular now There isn’t any such thing as popular music in terms of the music So I know there were– there was once a little department called systems analysis at Tufts, which had a couple of rather good philosophers try to make general theory of everything And they were writing nice little papers And it got a– it moved along But then there was this Senator McCarthy you’ve probably heard of And he announced that he had evidence that the– one of the principal investigators had slept with his wife before they were married Well, Tufts was very frightened at this and abolished that department And Bill Schutz went to California, and started Esalen, and had a good time for the next 50 years– more stories Yeah? AUDIENCE: Kind of as a extension of the body-mind question, it seems to me we– as humans, we– learn a lot from just interacting with the environment Like language, we hear it being spoken We speak it We see things We touch things But as far as I know, a lot of the efforts in artificial intelligence so far have been confined to the computer that does not go out into the real world, interact Doesn’t [INAUDIBLE] to see, learn new think PROFESSOR: Well, here’s the problem– I look over at Carnegie Mellon, and there are some nice projects And the most popular one is robot soccer And here are these little robots kicking a ball around They’re Sony– what are they called? AUDIENCE: AIBOs PROFESSOR: Yes, the Sony AIBOs Sony stopped making the AIBOs But it respected Carnegie, and it made a little stash, secret stash, of AIBOs to send to Carnegie when the present ones break But my impression of AI projects that have robots is that they do less, less, less than projects that don’t The reason is, if you have a robot like ASIMO, made by Sony? No AUDIENCE: Honda PROFESSOR: Honda ASIMO can get in the back seat of a car with some effort, and usually falls over

However, if you simulate a stick figure in a computer getting into a stick figure of a car, then you can make it learn to do that and get better and better And so all AI projects without robots are way ahead of all AI projects with robots And the profound reason is that robots are usually expensive, and they’re always being fixed So if you have five students, and the robot is being fixed– I don’t know what they’re doing– but they have to wait Whereas if you have a stick figure robot, then you can just run it on this, although it might be a little slower than your mainframe– probably not Yeah? AUDIENCE: So then back to the idea of the mind and the body, here’s a theory that I just thought of– the idea of the body as a seeming abstract, basically a mechanism for input-output It’s a set of sensors from which our brains can get information about the world, and instead of actuators, in which we can display our state So in that light, it’s almost as if our brains are really independent on our body itself It can adapt to any sort of body if we happen to hook it up that way And it just so happens that we’ve been hooked up to this body since birth, that we have such good mental models of how to use this body And I guess an example of– from experiments that support this theory might be how, when people have limbs amputated, it takes them a while to forget that they have the limb Because their mental model still exists The mental models don’t go away overnight And also, I guess, they train monkeys to control robot arms with their brains PROFESSOR: Sure Well, but it just seems to me that a large amount of our brain is involved with highly evolved locomotion mechanisms And as I said, when you’re sitting back with your eyes closed in a chair thinking about something, then it’s not clear how much of that machinery is important But it might be that– I have a strange paper on– I don’t know if it’s– trying to remember its name It’s called– do you think I can actually get a– I can’t remember the name of the title Oh, I give up The idea is that maybe– in the older theories of psychology, everything is learned by experience in the real world– so conditioning, and reinforcement, and so forth In this theory I call internal grounding, I make a conjecture Suppose the brain has a little piece of nerve tissue which consists of a few neurons arranged to make not a flip-flop, but a– what would you call a three- or four-flop, a flip-flop with three or four states? Let’s say three So when you put a certain input, it goes from– couldn’t find the chalk So here are three states And here’s a certain input that means if you’re in that state, you go to this And if you pop that input again, it does this And if you, say, go counterclockwise, it goes

So three of them get you back where you were But if I go this, this, and that, that would mean to go like this, this, and back So this would be– that means that’s equivalent to just going one Get the idea? In other words, imagine that there’s a little world inside your brain, which is very small and only has three states And you have actions that you can perform on it And you have an inner eye which can see which of the three points of that triangle you’re on Then you could learn, by experience, that if you go left, left, left, you’re back where you were But if you go left, right, left, right, you’re back where you are And if you go left, left, right, that’s like going one left In other words, you could imagine a brain that starts out– before it connects itself to the real world, it starts by having the top level of the brain connected to a little internal world which just has three or four states And you get very good at manipulating that You add more sensory systems to the outer world And you get to get– learn ways to get around in the real world So I call that the internal grounding hypothesis And my suggestion is, maybe somewhere in the human brain, there’s a little structure that’s somewhat like that, which is used by the frontal part of the cortex to make very abstract ideas You understand? The more abstract an idea is, the simpler, and more stupid, and elementary it is Abstract doesn’t mean hard Abstract means stupid Real things like this are infinitely complicated So we might have– and I wouldn’t dare suggest this to a neuroscientist There might be some little brain center somewhere near the frontal cortex that allows the frontal cortex to do some predicting, and planning, and induction about a very simple– a few simple, finite-state arrangements Who knows? Would you look for it? Well, if you were a neuroscientist, you could say, oh, that’s completely different from anything I ever heard Let’s look for it And if you’re wrong, you have wasted a year And if you’re right, then you become the new Ramon y Cajal or someone Who’s the best new– who’s the currently best neuroscientist? Maybe it’s late One more question One last question This is Cynthia Solomon, who’s one of the great developers of the Logo language Yay Yes? AUDIENCE: So maybe it’s a bad question for the end I will ask anyway What do you think about theories such as Rodney Brooks’ theories that can speak of no central [INAUDIBLE]?? PROFESSOR: Completely weird Obviously, those theories have nothing to do with human thinking But they’re very good for making stupid robots And the vacuum cleaner is one of the great achievements of the century However, his projects– what was it called? Cog– disappeared without a trace That theory was so wrong that it got a national award And it corrupted AI research in Japan for several years I can’t understand– Brooks became popular because he said, maybe the important things about thinking is that there’s no internal representation You’re just reacting to situations And you have a big library of how to react to each situation Well, David Hume had that idea

And he was a popular philosopher for hundreds of years But it went nowhere, and it’s gone, and so is Rod However, he is one of the great robot designers And he may be instrumental in fixing the great Japanese nuclear meltdown Because they’re shipping some of his robots out there The problem is, can it open the door? So far, no robot can open the door even though it’s not locked [APPLAUSE] Thank you