Interview With Marvin Minsky, 1990

we're any of you along with the moravec expedition did you film him okay now in your book society mind you you start off by saying or in the first page the century ago we had no way to explain the human mind now it's interesting to me and one of the things I want our viewers to understand that a machine which came into the world around the 1950s why should a machine at the computer be any helping and in this way did it form the basis to focus about a new intellectual movement artificial intelligence the development of computer science produced a new kind of thinking a new style of explanation for which there was really no precedent and the idea was how do you describe a process for example mathematics is good at describing things geometry it's good at describing curves and and it's good at describing mechanics like the differential equation for a solar system but it's not good for describing even a simple machine whose behavior depends on things that happen as they come in mathematically the mathematical world is to me sort of frozen but once we had the idea of a computer we could start saying how can I make something that learns how can I make something that adapts to its environment how can I make something that stores responses to a million different situations and changes and adapts and there wasn't any way to think about this before 1950 looking in our own lifetime so all of philosophy to me is simply bad psychology because the Philosopher's are struggling with the idea of trying to describe these complicated processes that we're made of in terms of simple things like logic and geometry static low-level crude course VOC descriptions compared to the subtlety of any process that has a million parts and a million bits of memory and responds to a million different conditions in different ways there just was no science of such a thing before even to this day no one knows how the nerve cells in the central nervous system worked very well there are a few theories but I think any sensible person would agree that there's only a small chance that the first few theories are right and so most of what we think we know about the human brain today is by analogy with what we really know from the computers I think there are some very different views of the same thing circulating today because some people say the brain is a biological thing it's very different from a computer it's soft computers are hard it's cold people are warm the shapes of the cells are variable and curved whereas the shapes of machines are a square but this is missing the point I think the brain is different from the rest of biology in a very important way because in most of biology there are a lot of interactions and things are rather complicated what was discovered around 1990 it's only a century old is that each brain cell is more or less separate from the others it's a separate machine it's certainly in a bath of chemicals but those chemicals are very carefully controlled and so when certain conditions occur around the periphery of a nerve cell then it suddenly goes bang and that's absolute it's very simple and clean just like the parts in the computer the signal flows along the fiber goes to the end make some conditions for another neuron so you might say that the cells of the nervous system are the closest thing in the universe to the transistors and gates of a computer so I find it very funny that some people will say oh it's so different and then they mentioned all the superficial features when they missed the important feature which is that unlike the liver and the heart and all the other parts of the body which are sort of continuous at analog it's the nerve cells which are digital they're on or off certainly they're very complicated digital devices and they're the thing in the universe closest to computers so it's funny to emphasize the difference instead of the similarity well one of the problems with the brain and the reason why today in 1990 we still don't know how a cortical granules cell works most of the brain is cortex it has believed to have a hundred billion or more cells and what we know about the cells of the brain we yet from our knowledge about the crayfish and the sea anemone because they stay alive for days in cool water whereas you know that mammalian brains die in five or ten minutes if they're disconnected and so it seems to me that the idea that you can't study a brain the way you could study a typewriter or an electric circuit it's very hard to study it because it's so delicate but that's not a matter of principle that's a matter of bad luck if the brain were a thousand times larger you could crawl around and attach clip leads and study it just like any other machine and so the general belief has grown that there's something different about it and again people are confusing essentials with accidents there's a reason why brain why your brain cell is a million times smaller than you or a billion times because in order to be intelligent you have to have a billion brain cells so you have to be a billion times larger so I think all animals in the universe wherever they may be will have a little trouble at first understanding their own brains because you need so many cells to be but we're developing instruments and I think another 10 or 20 years we'll be able to connect things to brain cells and find out just how they work and it won't be so hard after that maybe the first important idea about how machines could be smart was to notice that machines were tireless and they were becoming fast and so if you had a certain kind of problem which might be hard for a person it would be easy for a computer to solve it if the computer could simply try every possibility now only certain problems are small enough for this to happen but I remember once there was a wonderful puzzle it's called pentominoes I should have a picture of it but anyway it's a rectangle which I think is 6 by 12 or 8 by 12 of little squares like it most of a checkerboard and they're a bunch of little pieces each of which are made of five squares it's 5 in a row and there's corner shaped 1 1 2 3 4 5 and an L and a zigzag like a W I think there are 12 different pentominoes that you can make and it's a wonderful puzzle for somebody to put all of these pieces into one rectangle people find it very hard they fuss around some people fuss for an hour can't do it well we once Stewart Nelson was one of the hackers here once said well that's easy wrote a little computer program that tried all possibilities and it generated all million solutions in a couple of minutes so there's an example where the computer wasn't very smart but it could simply try after I put this piece in there are eight ways to put the next one in six ways to put the next one in five ways and it just did all those possible ways so if you have a small enough problem and a way to tell when it's solved then you can use what we call the exhaustive search and many of the early problems did they fall into that category yes and now a few problems fell into that category and worked very well but one of the things we wanted to do was get a machine to play chess and checkers because people liked those games and turned out for those games there's too many possibilities because I make a move in chess there about 30 things you can do typically and then the other person can make about 30 replies and so in two layers that's 30 times 30 it's about a thousand for layers it's a million and six layers you're up to a billion possibilities and a computer today could do a billion chess moves in in a minute or two but in those days it would take weeks or more so you're sort of getting a tree each limb has 30 branches in each branch has 30 twigs in each twig has that sort of thing and so you have to prune the tree so a lot of AI research in the 1950s was finding wonderful new ideas about how to reduce these search trees so that you could solve harder problems in the same time and I'd say there was 10 years of progressive and interesting discoveries about that presented this paradox that these were among many of the things we human most human beings considered very difficult to do you mentioned a wonderful example was and I think he was my second graduate student Manuel Bloom was the first Jim Slagel decided to write a program which would try to solve the kinds of problems that MIT students do in the first year calculus mathematics and the problem was doing what we call symbolic integrals and he wrote a program which consisted of more or less a hundred kinds of rules or suggestions let's say if you see one minus x squared in a math problem the mathematicians know it's very tempting to say x equals a sine Y or some trigonometric thing the reason has to do with the fact that in a circle the equation is X squared plus y squared equals one is very familiar to mathematicians very alien to everyone else but anyway Slagel wrote down about 20 rules of thumb suggestions for how to solve a calculus problem another 20 or 30 rules for how to do high school algebra because you can't do calculus without algebra and then a profound set of another dozen or so suggestions about how to tell when a problems getting too hard and you should try another one well you see in calculus to solve one of these things there are many things you can try like chest you could try moving this pawn or that Knight and in calculus you could try using a logarithm here or a trigonometric or a sine or a cosine or or just multiplying or dividing and there all sorts of alternatives when you're doing the algebra so he wrote down these about a hundred rules altogether you would the machine would try various ones and then it would use the special set of rules I called them rules of fear if if the thing got too complicated would say that's no good it's too complicated but if it seemed to be getting simpler it would follow it further with the amazing thing and this was 1960 just a couple of years after we started the thing got an a on the MIT exam and it was frightening it was doing as well as the average student or maybe slightly better I couldn't do some problems could do others and the kinds of problems it solved were pretty much like the kind of students did well of course this calculus problem is a very purely formal problem it's it's all in equations and mathematical expressions and although those are frightening to some people they're simpler than words Danny Bob Roe in 1964 started a program to do the same sort of algebra except that it would be word problems problems like John is twice as old as Mary was when she was the same age as you know those those are things that high school students find very difficult actually although the algebra is terribly easy but the problem is that the words have too many meanings and if you look at mathematical symbols in algebra maybe it's like a language of only 20 words plus minus x-squared such as equal such small vocabulary in pure mathematics but in the vocabulary of ordinary language 10,000 for a normal person hundred thousand for a very articulate person and so the kinds of things that every little child learns to do like talk with ten thousand words is much much harder we found to our surprise than solving the kinds of things that a PhD mathematician mathematician would do in a world of expertise and this went on for twenty years from the middle 1950s to the middle 1970s we found it was became easier and easier to do to get machines to do things that people admire it as expertise but it was very hard to creep down words from the college level to the adolescent to the child to the infant and see if we could get a machine to learn the kinds of things that everybody considers it's perfectly natural and simple and obvious yes very strange field because it had this backwards regression we started out in the 1950s Newell Simon and Shaw at in Pittsburgh wrote a program that did very amusing things with mathematical logic and it proved theorems and mathematical logic they were pretty hard they did they found a proof that was better than the best one that Russell and white had found and Bertrand Russell was sort of impressed and you know the machines were starting to play chess and do calculus and that sort of thing everybody's very impressed because machines were doing hard things but what we began to see is that the things that people think are hard are actually rather easy and the things that people think are easy are very we could do the calculus with just a few hundred pieces of program but to learn language to recognize faces to walk and to put your clothes on and do the kinds of things we expect every child to do we still can't do with the robots of a eyes of 1990 there are several ways to explain this a humorous way would be that it took animals took about three billion years for the first to go from the first cells to the vertebrates the fish and the amphibia and the reptiles and then four hundred million years to go from the first animals to the chimpanzee and then it's just four hundred million years you see and then it's just four or five million years to go from the chimpanzee to man so you might expect in that sense that the kinds of things that the chimpanzee or the child can do are very hard and the difference between the chimp and the man which is playing chess and doing calculus and things like that difference between a kid and a adult would be relatively simple in a sense first you had to get the basic brain that's able to learn complicated things the complicated things themselves or nothing so that's one reason I think the other reason is that why was it easy to build these expert systems and this is my own theory that if you look at the expert systems out there today that do such good things like chess each one is based on a certain way of representing the world we call this representation of knowledge or a model of the world or something and these wonderful high-powered programs each use one way of representing the world and one way of representing knowledge but in language each thing we do uses I suspect three or four major different kinds of representations and maybe twenty or thirty minor ones and so everything that an ordinary person does in ordinary life is a is consist of maybe 20 different ways of proceeding and all their relations between them that's much more complicated than the kind of precise narrow thing that an expert does for example when I see a dog I recognize it as a physical object and part of my brain says oh that interesting thing weighs about 4 pounds and it's has this color and so forth another part says it seems to want something so I have an emotional not emotional but I have a social reaction to it in terms of social communication and maybe their defense mechanism I have to treat this as a threatening situation is it going to bite when you meet a person you're discussing a particular topic you're wondering how you're getting along with them you're trying to cope with cultural differences if I meet somebody they say where do you live if they're a foreigner I say I live in Boston or I live in the east coast of the United States if there's somebody from this area I say Brookline but I know that strangers don't know where Brookline is they might have heard of Cambridge and so every time a word comes in the way I react to it depends on many different other kinds of knowledge and I don't think these problems are unsolvable at all effect I in the society of mind I propose some theories of it but I feel that the research community working on artificial intelligence got so addicted to its success with expert systems that almost everybody in the community is saying if we just get exactly the right representation we can solve all problems and I think the reason why it's hard to get a machine to behave like a child is that it's not finding the right representation that's important at all it's finding six or ten representations and discovering how to manage the relations between them I don't think that's a very hard problem but for some reason no one works on it it's it's outside the scope of what people consider their job your work on blocks world isn't it I'm wondering I mean many of the ideas were developed during that I wonder if you could tell me how the blocks the blocks project started and can't remember there's a story that's told anyway there's a proper form which says they did this was a this was conceived as a holiday project is that there's a garbled story about Gerry Sussman ya know what was true is that we had a vision project trying to get the machine to see these blocks and other things and it wasn't working very well and there was a very smart freshman named Gerald Sussman so I decided that the reason the visions project wasn't working very well was that everybody must be on the wrong track and so I put him in charge of it for one summer and because he had had many ideas of his own and I thought it might be a good chance to see if a beginner would do better and he didn't do worse than the others the legend changed into I've seen it written that Minsky put a graduate student and in charge of the project but that just shows how conservative people are I was actually a freshman and he's now a professor he was a rather good freshman well we decided we wanted to make a machine that interacted with the real world and a nice way to do that would be to give it eyes and hands and so I decided we should try to get the computer to be able to see things and when it sees something it should be able to do something with it pick it up and that turned out to be very complicated indeed because when you try to recognize it up it's easy enough to get a picture into a computer we made circuits that use things like television cameras which we're just beginning to be usable and the trouble is that a blocked or is different you move it this way it's a different shape and so you almost never see the same thing twice sometimes there's shadows on it sometimes it's darker or lighter different boxes have different surfaces sometimes there are things written on it so that even though to you or me or a child the idea of seeing a block seems simple it's actually very very complicated it's out of focus sometimes if the light on two sides is just the same intensity you cannot see the edge there are plenty of problems and it turned out that somebody would write a computer program to locate a block and it would work on three blocks out of ten or five blocks out of ten it just wouldn't find the others and another person would write another program to find blocks with a different idea and it would work on different blocks it wasn't that you could say each program got a score of 40 percent or 80 percent or something each different program would see would be better or worse at different jobs and what I began to sense was that we should stop looking for a very good vision program and this concept came to Peppard and me around the same time we were working together the idea was all right let's see if we could get 10 pretty good vision programs and get them and manage them and see which of them seemed to be working in different situations so the idea of the society of mind was that in the brain or in the mind or in a computer you shouldn't look for perfection and you shouldn't go around trying to debug programs and find the best possible way to do something you should find a lot of different ways and have different resources and then you should make managers that can decide under which circumstances to use which ones now I still think this is the way to make big programs work better but no one does it and so this idea which is now from about the late 1960s and now it's the early 1990s in spite of how simple and clearly correct this idea is it hasn't caught on and I'm very disappointed in my colleagues and people in the field of AI in general for some reason they've gotten fixed on the idea let's get it right and that's wrong there is no right in the world just we go there's one sense that you said before that it had been possible to capture the knowledge of an expert how was it taking something like a child stacking blocks or listening to stories what was different in the kind of things that they knew but what happens in understanding a simple story my friend Roger shank at Yale now at Northwestern had many of his students work on the problem of getting a machine to understand a simple children's story we did that here and a few other places and that's different from doing calculus or playing chess because in chess there are a few rules it's pretty clear what to do it's very hard to know what to how to play as well as a human and nobody's figured it out yet although the program's now play better than most humans maybe very much better but they don't do it the same way and they're still using a lot of search but when you understand a story you come across a word like boat and what is boat mean well that's a bad word but when you look at you know so much in calculus if you see a sine function you only need to have a few rules at least two calculus but for a boat you have to know different kinds of boats for different kinds of water they're kind of abandoned the damn thing I think well a very simple one in linguistics is the word take John took a trip to Mexico very strange word sometimes take means to obtain a physical object to take it away from someone else it's a social thing taking a trip I don't know what that means take a look you see there are if you look in the dictionary you'll find 40 or 50 different entries and so here's a set of meanings a set of processes to be applied to the rest of the words and you have hundreds of them in your head because the ones in the dictionary are just the families of these you have to use the other clues to decide which of these hundreds of different mental procedures to apply to the rest of the words in each of the words has the same thing well what I'm saying is that to understand even a simple sentence you have to know thousands of different things and my example before is to get an A in that part of freshman calculus you only have to know 100 things now I'm not saying that it passed the whole calculus course I don't want to you know oversell this thing yet but it did a certain large part of it the formal integration and that's part that people considered expert well to get it understand a simple child story you have to know so many hundreds and thousands of things and there are different kinds of things in a little child story some word we'll talk about the geometric shape or the nature of space another one we'll talk about time another thing we'll engage social relations and a normal person's fear of the unknown or greed or acquisitiveness or territorial defense and just to begin to talk to a four-year-old you have to know all those things that's what I'm saying is for a beginner to play chess you have to know 100 rules and if you do a little exhaustive search you can avoid the simplest disasters to talk to a little child maybe you have to know a hundred thousand things so what I'm saying is these simple things like understanding a little story seems to me maybe a thousand times more complicated than at least beginning to approach human competence in narrow expert domains well most people in every field end up in a few clusters of establishment I don't know the statistics but I'd say half of the certainly in the applied area a large proportion of people use rule-based expert systems in the world of research a large number of people use languages like Prolog which make it easy to work with rule-based systems in the America maybe most of the world when it comes to representing knowledge by far the majority of people I think use something related to mathematical logic the others use frames and scripts and rules but the most popular ways of doing things are always the ones that are the best established from 20 years before everyone everyone agrees that that you can't have an ignorant but brilliant machine be very good at solving problems because in order to solve a problem you'd better know something about the subject otherwise you have to make all the evolutionary mistakes but what they don't agree on is how to represent the knowledge and I'm afraid that mostly they fight about which representation is best and I feel that we have a dozen pretty good representations and I wish there were a hundred people working on the managers because I'm pretty sure in the brain that things in the visual cortex are represented one way maybe by sort of two dimensional structures and in the auditory cortex there's two of them this one maybe uses rule-based stuff this one uses something called semantic nets other parts use frames and all the different representations that the researchers and artificial intelligence have developed I don't know one of the dozen or so popular representations that isn't better at something and so I suspect that the brain has evolved lots of knowledge representations the exciting problem is how to coordinate them I think the expression common sense knowledge has a couple of flavors they're almost contradictory maybe the literal meaning is common sense knowledge is the knowledge that everybody share and you can trace that back for example to childhood every child knows what a parent is except one that doesn't and we don't care much about exceptions here so every child knows something about a family and every child knows things about social relations that if if you hit somebody they make an expression suggesting annoy and serve pain and if it's another child it might fight with you it's just so many thousands of little things everyone knows that if you hold something and release your grip it falls they don't know about gravity but they know that this is common sense there's no person that you can communicate with who doesn't know the same things you do about space and time and social relations and geometry and language and and what not how large is this database that we all share I suspect it's about 10 million items or units whatever units are but nobody you know depends on your representation of knowledge now there's another thing when we say common sense common sense reasoning it's as though there's a kind of thinking which is very simple and obvious and everyone has it and that's a I think a bit of an illusion the kinds of reasoning we find most simple Harper HAP's the most complicated and highly evolved ones in our brains what do you mean by comments an example of common sense knowledge common sense reasoning sorry as opposed to common sense knowledge I think an example of common sense reasoning is that if you see something move then you say well either it's an animal and it moved for a reason of its own or it's a physical object that's inanimate and it must have been pushed or blown or something like that so that we all share kinds of reasoning when we see something happen we make explanations and these I think are rather complicated and very important and nobody knows very much about them or did say that these things came from growing up with a body and the idea of a disembodied was therefore in question because of this I think that the there was some criticism that somehow you couldn't separate the brain from the body and the world but I never could understand what the critics who were talking about that had in mind because of course if a machine has to learn something it has to have some environment from which to learn it and if the machine is going to be competent in dealing with the physical world then either it will need a body to experiment in the world or else it will need a little computer that simulates the kind of physics that you need for a world like an airplane simulator but there's one other thing that the philosophy that that those critics didn't understand which is that if you could program into the Machine and the same knowledge that you would get my experience without learning then it would understand it just as well as if it had learned it and so there's a lot of confusion between the present state of a person suppose that you have a normal person they become paralyzed will they still understand the world they're not interacting with it so it was nice to have a controllable body for learning through but it's not philosophically or technically important it's just convenient and there are other ways that it could learn the brain actually after all isn't in the world it's imprisoned in the skull in this dark moist quiet place only connected to the world by video cables it seems to me that the reason people are as smart as they are is that they have several ways of representing knowledge if you have just a single way to represent knowledge say as strings of words the chances are that you might get stuck not be you try every way you know of solving a problem they don't work there's nothing else to do if you have a visual way to represent the world and an auditory way and a logical way and a possession away in a political way and so forth then whenever you're trying to solve a problem and you get stuck you can shift to another way and so the more modalities it's not that you have more senses it's that each part of the brain connected to a sense organ has actually evolved a different kind of hardware and so the person who's born deaf is a little bit handicapped because they don't have access to a kind of one way of dealing with the world now in fact if they're well educated they may become better at solving most kinds of problems than hearing people or sighted people because they can overcompensate and I believe every person has a dozen ways of representing knowledge and if you're blind you lose a couple of them you've still got eight left and but if you lose all but one if you lost all but the sense of touch then you might might be very difficult Helen Keller is a person who I think she got meningitis after she and she had some memories of seeing and hearing and it's much harder with babies who were born with no senses at all I don't know if there are there was an example we found Oliver Sacks has one of a person born with nine with cerebral palsy blind and deaf not deaf no no me nothing but but still it's remarkable that most of what she knew about the world until she was 80 or something she was read to her out of novels no they're still very restricted mm-hmm and she could talk but when you think about it if you're reading a novel then you're reading knowledge that it's been processed by adults and you're much better off than learning it yourself through a baby's brain project it's an interesting example which you've been talking about and he's having to give well the psyche machine is disabled in the most profound sense of all which is that it doesn't learn so wouldn't matter if it could see or hear it's basically a knowledge base that is not able to acquire knowledge on its own and it's the first attempt to try to put in one machine many different kinds of knowledge and I expect I would like to see ten other such attempts around the world it's a shame that we have all our eggs in one basket and Leonard and his group have many wonderful ideas some of them might be badly wrong I think it's hard to learn if you don't know a lot of things and if you don't know how to learn the trouble is that no one in AI knows how to tell the Machine much about how to learn because there hasn't been a really enough research on it there are quite a number of early stage machine learning projects around the world but I'm afraid they're a very small minority of the general investment in artificial intelligence compared to building practical performing expert systems the we really don't know very much about machine learning to this day the learning would seem to me to be a like a basic thing why has it proved so intractable or are you saying that just not very much good work yes I not sure that learning is intractable but the number of people who have tried to make machines learn is pretty small it hasn't been a high priority project in order to learn I suspect that you have to go through stages that is unless you have a certain set of concepts that and processes for using them it's very hard to understand the next step it's the way Piaget the great Swiss psychologist described the development of children he didn't have much of a theory of how they learn actually that may be in the earliest stages of infancy but he pointed out that before you can understand the idea of conservation of matter or energy or something like that you have already have to or he thought you had to have the idea of distinguishing between actions which are reversible and irreversible for example if I take a piece of clay a ball of clay and I flatten it out it looks much bigger but the older child knows that the flattening out process was reversible and we can just roll it up again and so the child knows well if the operation is irreversible then in some sense the greater extent is not essential it's just a momentary feature and gradually the child accumulates a number of ideas which amount to that there's a certain quantity of substance and what Piaget is saying is that you probably can't learn that concept in its full power and all of its facets until you've got these other ideas first so there we're saying that you can't learn the law idea of conservation until you know the idea of reversibility well maybe there's another route maybe there isn't nobody's thought of a plausible one but it might be that even to learn very simple things that we take for granted for example how do I learn that if I take an object and release my grasp it goes down I have to have the concept of down I have to have the concept of of intentionally releasing it as opposed to something else I have to have the idea of of support in this case support from the top that's different from support from the bottom mouse to the same thing so there's so many ideas you need before you could even look at the world and make explanations and we have people working on what we call explanation based learning and to me that's one of the very most promising ideas in modern AI research today explanation based learning you look at a situation you don't just describe the bits or the pixels of the picture you describe the objects and their apparent relationships but the relations come from you it's not that there's a hand there and a piece of paper it's that the hand is grasping the paper that grasping is not there in the world actually it's something that comes from my own knowledge of the scene so maybe for psych or any knowledge based program to learn it's going to be very slow nearly impossible until we prime it with the with just the right sorts of concepts so it can start going rapidly I'll bet that a human infant is born with a surprisingly large collection of built-in procedures and mysterious pieces of hardware and and reaction schemes that make it easy for the little infant to learn we just don't know what those are yet I suspect that once once you get a pretty good system for learning then they're a couple more stages maybe the child can then starts experimenting or the machine with variations on ways to learn because after all some people might take five examples of something before they say well maybe that's the general rule they're more impulsive people they see something happen once and say oh whenever there's a dis that'll happen that people go so far you'd call them superstitious and so each person may tune things and I think the greatest breakthrough of all will be when you get smart enough that you can invent new ways to learn and try them out and see which work and then the thing can invent better ways to improve itself and so forth and just take off and what both Leonard and I and many other people agree is that some threshold if you don't get up to that threshold the machine just won't get better maybe it'll get worse if you let it learn there was an example in 1957 when Arthur Samuel made a program that learned from experience to play checkers and if it played with a good player it got better and better if it played with a bad player it got worse and worse and we don't want that and but I suspect that once you get up to a certain threshold you could say my goodness I've been learning from this experience and I'm getting worse I'll turn it off that's a very simple piece of knowledge but if you don't have it you might ruin yourself and children who make bad friends that's we as parents our greatest fear is what happens it's it's not our influence on the children it's who are their real friends who are going to elevate them or ruin them because we know our children don't know how to learn to learn they're going to copy well it's difficult for anything to learn something unless it's it's the right machine for it I built the first neural network learning machine in fact before I started to work on symbolic approaches and I got annoyed with the thing because my particular machine learned very quickly at first and then it got slower and slower as it filled up and in order to make a new distinction it would have to forget an old ones rather small machine and I got the feeling that that it just didn't have enough organization to learn hard things now the modern connectionists are are in a very strange level of science I would say right now because you can see hundreds of papers somebody says look I got this machine to learn to pronounce words from spelling surely this is a very hard problem some human programmer took a couple of years and to do this by hand and his program is only a little better than mine that sort of thing well the trouble is we don't know how hard that program is in an absolute sense if that human programmer managed to write a program I still don't know what it is he understood in doing that and I don't understand what the neural net did as far as I know until they get some more science we just have to look at these anecdotes people say I got a neural net to do this I got a neural net to do that sometimes you hear somebody say I'm trying to get it to do this but I can't of course people don't publish what it won't do and so we don't learn much from this because they're just anecdotes so people are angry at me in that field because my feeling is yes if a neural net did that it shows that probably the problem that it was solving was easier than they thought and they get very angry but instead of getting angry of course what they should do is come up with a theory to show me that that problem was in some technical sense hard the trouble with the field right now is that there are good theories of classifying problems into levels of difficulty in other parts of computer science there's been some progress on that there's a what we call the theories of algorithmic complexity but they're still not very good and so the progress of psychology in general and particularly connectionism as a science is going to depend on the invention of better mathematical theories of how difficult problems are just as in physics physics couldn't progress until even after newton until we had more theories of the characteristics of different kinds of differential equations right every now and somebody will solve a problem somebody else ought not solve a problem we won't know whether they were lucky or all the problems they're solving or easy or they are making really profound discoveries so it's a little muddy until you get a theory but most sciences proceed fifty or a hundred years with the experiments ahead of good solid theory so I'm not complaining that much your views of diverged someone how would you describe the difference between he does he still believe you can get at these things through logic and it's it's rather tricky to describe just where we agree and disagree we both agree very much and always have from the beginning that in order to for a machine to be smart it would have to have common-sense knowledge where we differed I think was on how that common-sense knowledge would best be represented and on what are the reasoning processes that use it now he's maintained that it would be good to have a uniform logical reasoning process but in order to do that you have to find ways of dealing with exceptions and suppositions and things like that and he's been working on technical subproblems of that sort for some thirty years how do you make a logical system how do you have an axiom and tolerate a few exceptions how can you do reasoning of the form what if this were not true for a moment what could I learn from it it's very difficult and my feeling is that there are other ways to reason by analogy using frames and defaults that are more lifelike and more productive and that you don't have to struggle quite so hard with these logical difficulties if you start with a more flexible system in the long run though it would be nice if we were using these other informal kinds of reasoning to have theorists come along and clean them up and say well certain places we can replace it by much more efficient perfect procedure I suspect that most situations that can never be done but it doesn't matter so we differ on what problems to work on John likes to McCarthy likes to prove things get them settled if you have a good theorem it lasts a lifetime if you have a practical theory you just never know what its status is from one year to the next the society vine theory is basically that in order to make a machine with the kind of versatility and resourcefulness that we take for granted in people a good way to do that is to package into that machine a lot of different ways to represent it not represent knowledge and a lot of different ways to exploit it and this leads to a certain difficulty is there a central place in this mechanical brain that's in charge of everything and knows everything and I think what I show in the book is that that really can't be because of different kinds of knowledge I represent in different ways than the parts of the brain the parts of the machine that's doing all this really can't communicate with each other very well and so you get a very different picture of identity and I can't explain it briefly but it's a 300 page book and in it I think I show all sorts of new ways to explain problems that have bothered psychologists and philosophers for a long time like what does it mean for a machine to be conscious and what I argue very much as Freud did is that this is not so difficult to problem as people think because the phenomenon of consciousness is overrated people if you talk to people they act as though they know what they're thinking and they know what's out in the world and so forth but in fact you don't know where you got the next sentence that you speak I don't know where my words are coming from what made me think of them so that there's a little speech machine which has a little bit of memory of what it did a moment ago and I don't see any great difficulty in simulating that sort of thing in a computer the hard part are the maybe 400 different sub machines that are computing different aspects of how to solve various problems there are lists of goals that I have and machines interpreting those goals maybe one of the goals is expressed verbally but it's talking about physical things and there's a misunderstanding between this part of my mind and that part of my mind and it's a big mess well I think that the only way to make sense of the weird phenomenon that baffle psychologists and philosophers is to build a machine that works this way and as far as I can see judging by the failures of for example connectionist machines to learn to talk there's a big difference between learning to understand a sentence and learning to pronounce a word and logical machines learning to solve the simplest common-sense problems and so forth seems to be the way to proceed is to find ways to do everything build them all together find ways to manage them and then study what kind of phenomena you get when you assemble that machine my prediction is with a little little work you'll find the machine saying that it's conscious and denying that it's a machine and and having all sorts of beliefs of unscientific kind that every normal common-sense reasoning person ends up with the mind is not a centralized thing it's a whole collection of different parts and we see that in brain surgery somebody has an accident loses a piece of brain there's still a person there it's not the same person it's missing some trait it can't recognize faces it can't think of the we see injuries where person can't think of the names of animals peculiar kinds of aphasias and ain't know me as and so forth and if these are small injuries this apparent person still functions it's somewhat like the original person it's missing some things sometimes it adapts and rebuilds and find substitutes but to me the person is not a person in normal sense a person is a wonderful package of interrelated traits and ideologies and things it's learned and pieces of hardware and it's a wonderful concept even if realistic I guess I don't understand oh yeah computers are now getting so fast pretty soon you'll be able to buy a little box that computes at a rate of 100 million operations a second and by the year 2000 or 2010 it'll be doing 10 billion operations a second in a desktop machine by that time I wouldn't be surprised if that's enough hardware that you could make everything a human brain does or simulate everything a human brain does in some sort of software maybe it'll be 10 or 20 years after that maybe so nerds from the view of history 10 or 20 years is a blink in in in evolutionary time so we shouldn't be worried about what what day it happens but surely in a hundred years there will be machines this big that have more capacity than the brain Hans Moravec thinks it's 40 years I think you should do what I'm doing namely start with the most interesting aspects of mental activity and try to figure out how they work and simulate parts of them I think if you start by simulating the early stages of evolution then you'll spend a long time discovering the obvious UI which I would start as Leonard does with simple phenomena of natural language and work both ways take take some level of performance which is meaningful and easy to understand and respectable and work down to say how could it be learned and work up to say how could this turn into something like an adult but I wouldn't start at one extreme or the other if I had unlimited resources I would duplicate myself and just stay home and think and after a long time I'd come out and tell people what I concluded I'm not serious I get most of my ideas by arguing with people who don't agree and then going home and working on the details and then when I get stuck coming out and arguing again I'm not interested in a big project because well I'd like to have lots of people thinking about how to combine different approaches it's not the different approaches themselves it's why aren't there more people making a machine that uses three different representations of knowledge and crosses over that's a very specific kind of research project and I see no one doing it so that's to me that's the missing link didn't see a thing yeah to me foreign countries are places where the people can't talk right I get no insect friend what is it - I mean limited vocabulary in with them I haven't seen a project I can't speak Japanese so I don't with continuous speech when did they demonstrate this I've seen a film it's kinda Carbonell and it's in collaboration with masters this has been throughout a controversial field I wanted to get some of your reflections on why you think sort of you know some people whether it's what do you think I think the the field is controversial because we live in a spiritual spiritualist culture when Pastor argued that living things were just chemistry and that was unacceptable and because people said there's a real difference between things that are alive and things that are dead you think you don't even apply the word dead to rocks they're not worthy of it what we say alive and inanimate and so in the 19th century until Pasteur roughly this was considered to be a very important distinction now in science there's no distinction at all nobody considers living things to be any different from other things except that it happens to have certain processes going on well I think the same thing is people think that we live from a tradition from Plato on which is that there's a mental world it's a spiritual world that that the body is a mechanical thing and the mind or even the soul is something else and so AI is challenging that in a religious culture we would be heretics to be burned or or whatever because but I don't see that has anything to do with artificial intelligence it's that to to most people in our culture we're saying there are no souls there are no spirits and so this is a religious controversy not a technical one no technical person to me then equality thinks that there is such a thing as a living thing there are just things that move around because they have myosin and the mechanisms are sort of understood pretty well and we understand that you can't have something that's half alive because it takes a lot of stuff to cap this thing keep going and repairing itself and fueling it because it's a rather crummy structure anyway it needs a lot of continual repairs all the time so the living things certainly are identifiable and they're different from the other things but there's nothing special about them I think the same thing is the case with mentality that is if you have enough knowledge and enough processes and enough add up other processes to keep it in contact with what it's doing then you get in mind and I don't see it as something to argue about but if somebody thinks that we have a spirit and an inherent value which is different from the stuff we're made of then of course it's a threat but it's a religious argument not a technical one they don't know they're religious religious is blue to me religious is the superstitious belief in spirits that don't exist and so anyone who says there's something in a man that's not in the machine is religious in the sense that they're saying there's a spiritual quality I can't explain no matter what you say I refuse to believe that I don't have it that's faith that's not they're not saying anything that it does that that technically they can show we can't do or whenever you've done something the the problems being redefined so because clearly still today people from Attica there's a way people say well that's just mathematics well it's nice you gave that example because Slagel was the first program to do formal integration then Joel Moses four years later wrote another one which is somewhat better than he and Karl angleman Bill Martin a number of people worked on it then Bobbe cavaness and robert rish came in they added more thematics it got better now it's better than any mathematician in the world and so now that it's that good it's not considered experimental or controversial so it's out of AI and typically as machine gets better and better at something it gets its own identity it depends on what you're looking for in a funny way in physics for example that's a game for young people because it's very hard the new theory of physics comes in it it's more complicated and wipes out the old one maybe it's simpler it's hard to keep up in AI people are are so it's so controversial that it's still easy for me at my age to make up new theories like so in a kind of selfish personal way it's very enjoyable that there's this hostility there's still only a handful of us and all these wonderful problems it's it's like having all the children's blocks you want and the other kids don't come and take them away from you but I think it's too bad that more people don't understand how much more we could do if people would sort of try new ways and cooperate and try to combine these methods instead of always arguing I want the best one my method is better than yours I think it was a bit of big surprise that the things children do was so much the things that seem harder easy and the things that are easy seem hard other than that it's hard to dissect that because I never tended to think in terms of how long things would take or how hard they were it was more saying if it's easy then I don't want to bother with if it's too hard I don't want to work on it now I don't know if I imagined when I think everyone was surprised when the machines got twice as fast in memory got twice as cheap so rapidly but when I was a little kid I read HG Wells and he Smith and Isaac Asimov was great pleasure meeting him and keeping up with him now because how often do you get to meet your gods and so I read science fiction more than anything else I don't read ordinary literature at all I read some technical things and I read science fiction novels and nothing surprises me except why doesn't everybody see that this is the right thing and work on it the science fiction is like any field most of its bad but it's full of a dozen people who I think they're great philosophers of our time Asimov and Fred Paul and Arthur Clarke and now Greg Benford and David Brin and now that I started I feel Vernor Vinge II I feel everyone I don't mention Harry Harrison it's being left out but these are the great writers but the publishers have got them in this niche but when I see Norman Mailer or or someone like that that's trash why he's no better than Harris dolphinese he's writing again about the human condition and people screwing each other and people betraying and being attracted and infatuated it's the same old stuff but in science fiction people say what if something we're actually different and general literature is what if things were the same again it's too boring robotics proved to be complicated well he didn't say when Oh Robert I was so entranced by Robert Heinlein's book in 1940 about remote manipulators we still don't have those in inequality so some things are unaccountably slow I tried for years to get people to build robots with five fingers just like hands I said no it's too hard after a long time they started making ones with three fingers and then four why don't they just bite the bullet because I want a five-fingered one so I can slip into the glove and get an output and I don't understand people it's only 20% more than four figures it's not as though it were twice as hard I'd say they were dark ages and then the Enlightenment and it came in 1950 rather than 1350 they'll just move the transition the computer is when people started understanding processes instead of just static things and so philosophically that was a great difference before 1950 there was no way to describe something that was changing a new way of thinking that there were procedures and that in computer science you make a procedure you say here's the procedure it's on this disk this little package I'm going to take this procedure and that one and put them together and I'm gonna attach this one here on the side so that what happened in 1950 was that we could think of processes with the same mental equipment that we could think of things before I can everybody is known for ten thousand years that you can build something higher by stacking one thing on top of another now we know about subroutines and recursions and tael recursions and there's a hundred words that the average person doesn't know which are just important as the old word like beside and on top of so most people don't know that what happened in 1950 that man for the first time learned to talk we didn't have everybody says well we learned speech sometimes 30,000 years ago nobody knows when but what I'm saying is a thousand years from now it'll be 1950 when when this animal learned to talk the stuff before was just emotional utterances because it couldn't describe processes he could just describe no there's a thing there computer scientists were the worst of our enemies was the computer scientists who were telling the public it can only add fast I had so many friends artists and I tell them we're going to be able to do this and they say how will it work and I would tell my scientists and say it'll do this they said it's just a fast adding machine can't do any of those things so through cue tyranny people who know too much but not enough had great advantage because when I came to the field of little college student I meet Warren McCulloch and John von Neumann and these people different world they were called cybernetics and I was just very fortunate I landed in this these are the people who century from now will be the philosophers of our time how many people know the name Warren McCulloch the greatest philosopher of the twentieth century he's unknown but that's my prediction 100 years from now they'll say those people were so lucky to have known Warren right and those the people who are thinking of processing processes as stuff and so when I sort of appeared as a child I got into that culture the Macy conference cybernetics inner circle never fell out of it again okay thank you very much thank you could get some real time yeah

Channel: GBH

Views: 22,703

Rating: 4.8225255 out of 5

Keywords: Marvin Minsky, 1990, Interview Marvin Minsky, WGBH, WGBH Open Vault, WGBH Media Library, Media Library, WGBH Archives, Archival Footage

Id: DrmnH0xkzQ8

Length: 72min 11sec (4331 seconds)

Published: Mon Apr 29 2019