table of contents


Edward Feigenbaum on Artificial Intelligence

Edward Feigenbaum is Professor Emeritus of Computer Science at Stanford University, where he was also co-director of the Knowledge Systems Laboratory. He received his PhD from Carnegie Institute of Technology (now Carnegie Mellon University) in 1960, working under the supervision of Herbert Simon and developing EPAM, “Elementary Perceiver and Memorizer.” He is considered one of […]

download transcript [vtt]
This is KZSU Stanford.
Welcome to entitled opinions.
My name is Robert Harrison.
We're coming to you from the Stanford campus.
I'm joined in the studio by a very special guest who is going to share some thoughts with us today about cognitive science and artificial intelligence.
In title opinions specializes in embodied existential, historical,
eminently human intelligence so it should be very interesting to hear all about both the history, the present, and the long-term future of machine intelligence.
No one is more qualified to speak about this topic than my guest Edward Feigumbow, Professor Emeritus in the Department of Computer Science of Stanford.
Professor Feigumbow received his PhD at 1960 at Carnegie Institute of Technology, now known as the Carnegie Mellon University.
His thesis advisor was Herbert A. Simon, one of the founding fathers of cognitive science, artificial intelligence, and information processing.
In his thesis, Ed Feigumbow developed E-PAM, or Elementary Perceiver and Memorizer, which, if I understand it correctly, is a psychological theory of learning and memory implemented as a computer program.
Designed along with Herbert Simon, E-PAM was originally intended to simulate verbal learning but was subsequently expanded to account for data on the psychology of expertise and concept formation.
Will he be hearing more about that shortly?
In 1994, Ed Feigumbow, along with Raj Reddy, received the most prestigious award in Computer Science, the ACM Turing Award for pioneering the design and construction of large-scale artificial intelligence systems, demonstrating the practical importance and potential, commercial impact of artificial intelligence technology.
A former chief scientist of the Air Force, Ed Feigumbow, received the U.S. Air Force exceptional civilian service in 1997.
In 2011, he was inducted into the IEE, Intelligent Systems, which is artificial intelligence's Hall of Fame for significant contributions to the field of AI and intelligence systems.
Professor Feigumbow was also chairman of the Computer Science Department and director of the Computer Center at Stanford University.
He is the founder of Stanford's Knowledge Systems Laboratory.
In 2012, he was made a fellow of the Computer History Museum for his pioneering work in artificial intelligence and expert systems, and this is just a sampling of his very impressive bio, which will be posting on our website.
But I for one am eager to get him talking, so let me welcome into the program. Ed, I'm glad that you could enjoy us here on entitled opinions today.
Robert, thanks for having me. It's really a privilege to be here, and it's an opportunity to thank you for the really great work that you've done in giving us so many interesting programs.
And so I'm happy to contribute to that.
Well, these interesting programs are in large part dependent on having guests like you who are willing to come and share their expertise with me on air.
And so that's why I'm looking forward to this topic in particular because it's not a domain that I tend to venture into a lot, artificial intelligence, but it's something that clearly affects all our lives, those of us who belong to the time that we live in,
and it's going to convince going to have a huge effect on all our future.
So we're going to talk about the history as present as well as future of artificial intelligence and cognitive science, and you have a very fascinating story to tell about that, both the history and the future.
Could I ask you just to give us a kind of basic version of that history and share with our listeners some of the things that you have done also personally to contribute to that history?
Here's a plan.
I'd like to focus some time on what I call the Big Bang, which took place in 1956, and thereabouts, lasted from 1956 to about 1965, but there's a very interesting,
story of what happened before the Big Bang. That's what people always ask physicists.
What happened before the Big Bang? I'll tell you, in artificial intelligence.
And then after that I'd like to focus on the stream of work that led up to, through machine learning, up to the present day, and some rather sensational applications, and also some of the maturation of what came to the first day.
So let me begin in the prehistory, which is absolutely fascinating.
In mid-century, actually in the 1930s, there was a young and brilliant logician in Britain named Alan Turing, for which the Turing Award is named. He's regarded as the father of computer science.
He proved a special theorem about a rather abstract kind of device that we now call a computer. It wasn't a real computer, but it had all the properties of what we now call a computer.
But that didn't last very long because World War II came along, and Alan Turing, although he volunteered for some other duty, was actually assigned to a then-secret and now very famous
cryptographic laboratory, a cryptanalysis laboratory called "Bletzley Park," which and Turing's work with his colleagues at "Bletzley Park" resulted in breaking the enigma code.
You've probably seen the play breaking the code. There was a movie made of it, and there was a TV program made of it.
So Turing, in order to do that, code breaking is very hard to do manually. That's why it's used for secrets.
So Turing had in mind the idea that you could design a computer, an information processing machine, not doing a lot of calculation, but just doing a lot of manipulation of symbols that you could do that.
And he designed a computer of its time, pre-electronic, electro-mechanical. It's happened to be called the bomb, BOMBE, and it was later replaced by a real electronic computer called the Colossus, which is really the first major electronic computer that came into existence, although the British intelligence service never let it be known to the world.
Such a thing was Colossus developed during World War II.
Yes, it was developed during World War II in 1943 and 1944.
By Turing.
By a group followed on from Turing's work, another group, in a famous person who should have become famous had this not been a secret.
A post office research employee named Tommy Flowers actually designed the Colossus machine, which came into existence because the Germans had changed their coding machine from the enigma to a much more complex machine and the British were in the dark for about nine months until there was this electronic machine that was about a thousand times faster than Turing's machine.
So Tommy Flowers is an unsung hero in this pre-history?
A bitter. Well, he's dead now, but he was a bitter and unsung hero because the British didn't make this public until the 1980s when everyone else had claimed credit for the invention of the computer of the electronic computer.
But anyway, during the war, in the early 1940s, Turing was talking to the young people that were working with him.
In particular, one of them named Donald Mickey, who later became the pioneer of artificial intelligence in the UK, was working for Turing at the time.
And Donald told me many times about the Turing stories of Turing talking about these computers ultimately will be capable of thinking.
In 1949, when Turing was leaving the employ of the British governmental lab called the National Physical Laboratory, he wrote a report for the boss at the National Physical Lab on this board.
And he was on this subject.
And many people urged him to rewrite this article for public use in a journal.
And he did. And the journal was the journal mined.
The article has some lengthy title, but most of us have shortened that title to be kind of machine-think.
And in that paper, Turing lays out, first of all, a point of view and second of all, a test, which has become known as the Turing test for intelligence.
But the point of view is what is most important, which is thinking is, as thinking does.
Thinking is just a word that we use to cover a great deal of behavior. From perceptual behavior all the way to the deepest kind of cognitive behavior that you would find involved in Einstein's thinking and Shakespeare's thinking and the kind of stuff we do on this show.
And the kind of stuff. And even language, just the conversation that we're doing right now.
And so, Turing introduced the idea of a behavioral definition of thinking.
Let's not be philosophical about it. Let's not be theological about it.
Let's just see if a machine can do what people can do.
And hide the person and hide the machine and have a human judge decide at a level better than chance, which is the human and which is the machine.
So that's the famous Turing test where a human judge would not know whether it's speaking to a computer or a human being.
And if it's not able to determine that, then it's irrelevant whether this kind of thinking is equivalent.
Yeah, let's just give it the same word we give it when we talk about people, namely thinking.
So nothing much happened in the field from the publication of that paper until 1956.
In 1955, John McCarthy, who was a professor here until he died two or three years ago, decided to convene a conference, which was held at Dartmouth, famous conference, the 1956 summer,
in which about a dozen people were brought together to try to frame up what kind of scientific issues were involved in actually producing computer programs that would do that.
In the meantime, computers themselves were growing up from basically nothing to the first things that you could sell to businesses and you could use for scientific work.
By 1956, something really existed that you could run these programs on.
In mid-December of 1955, if you read the autobiography of Herbert Simon, my thesis advisor who later won the Nobel Prize in economics in the '70s, he was a genius in many disciplines.
He's one of the biggest prizes in political science, and psychology, and computer science, and economics.
Herbert Simon and a very advanced graduate student of his at the time who later became one of the most famous computer scientists, Alan Newell, really got down to business.
And they actually took a task, which was, people thought was a difficult task, and proceeded to write, I'll tell you what that is in a minute, and proceeded to write a computer program which solved problems within that task.
In the early part of the 20th century, Bertrand Russell and Alfred North Whitehead had, in investigating the foundations of mathematics, had written a tome called "principia mathematica," which dealt with what is called propositional logic.
And by propositional, it means that the elements of these theorems were simply P's, Q's, R's, S's, etc. They were not objects in the real world.
And that was a brilliant choice by Simon and Newell to do that, because we didn't know how to connect up these programs to anything meaningful in the real world.
And it didn't quite follow. Simon and Newell were trying to implement a computer program to solve some of the problems that Bertrand Russell was of propositional logic.
And the program was called the LT or the logic theorists, and it proved theorems in Chapter 2 of "principia mathematica."
In fact, proved all the theorems in Chapter 2, and proved one of the theorems that was given a rather lengthy proof by Whitehead and Russell.
It proved it in a few lines, a spectacular short proof, that in fact, Simon wrote up and sent to Russell, and Russell wrote back and was amazed.
And you can see the Russell's letter back to Simon is quoted in Simon's autobiography.
As Russell was saying, "Whitehead and I waste our time on this, I'm now prepared to believe that a machine can do anything."
Was this a mathematical computation that the machine had done?
Very interesting question, Robert, because you used the word computation.
Computation up till that point was thought to be a numerical thing.
You would compute a formula, and there would be numbers coming out.
Numbers went in and numbers came out.
This was pure symbolic manipulation.
These were theorems, there were no numbers involved.
You'd start with a theorem, and you'd end with an axiom, and just like a mathematician does, you would say QED.
And so that kind of computation became known as symbolic computation.
The key thought was that what people did when they went through the behavior that we call thinking is purely symbolic information processing.
In physics, when you construct a theory, a physicist normally aims at expressing that theory in mathematical terms.
But that turns out to be difficult and fruitless for something as complex as the behaviors that we call human thinking.
It turned out exactly the right language was the language of information processing that is reflected in the programming languages of computers.
You don't have to think of them as adding machines or multiplying machines.
You can think of them as simply devices that take a symbol out of memory, put it somewhere else, look at it, test it, manipulate it, do something with it, infer something from it, and put it back into memory.
So what is a symbol here, could it be a word, or is it an abstract?
All of the above.
Down at the machine level, it's represented as a string of zeros and ones, which are choices.
They're not thought of as numbers. They're ons and offs.
But they're just a way of coding something.
I already mentioned that in connection with Turing's work on the German messages.
They weren't treated as numbers. They were treated as symbols.
So is this the Big Bang we're talking about?
Yeah, so this is the Big Bang. This is the big idea.
And McCarthy gave the field its name, Artificial Intelligence, by naming the conference that was held in 1956.
In addition to the logic theory program, one of the attendees at that conference, Herbert Colerinter of IBM,
did a program very much like the logic theorist to prove theorems in Euclidean geometry, high school geometry.
And that program, the Geometry Therm-Pruing program, was the only taker of the New York State region's exam in plain geometry,
and I believe the year 1959. The only taker of that exam that scored 100%.
In addition to that, it actually proved one of the more famous of Euclide's Theorems in a completely novel and interesting way,
which had not been noticed in a couple of 2,500 years of 2,000 years of examination by all of the people who have looked at Euclide, including all the teachers and professors and students and so on.
That was really the Big Bang. There are a couple concepts there that are very important in carry on into the future.
The first is the concept that the essence of problem solving is search.
That's search really is, that problem solving really is decision making in a very large space of possibilities.
That are implicit in the symbols that you're manipulating, but the whole trick is to find your way through that space of possibilities, to find the right answer or a good enough answer, actually in Simon's terminology, he used the word "satisficing" to contrast it with maximizing.
You're not trying to maximize anything, you're trying to find a solution which is good enough.
So by search, do you mean that the machine has to, if you want to use this figure, eliminate all the noise, that is irrelevant and zero in on what is most pertinent to the problem?
Yeah, and it's not necessarily noise. It could be just worse solutions, solutions that are not as good.
So for example, you can imagine starting off the search in a good chess move.
If I move here and he moves there and then if he does that, I have these possibilities and if I do that, he has those possibilities.
If you work that all out to the end of the game, that number of avenues of looking for a good move is a one followed by 120 zeros.
Well, there's no way. I mean, that's bigger than the number of particles in the universe.
And search is a way of narrowing down? Yeah.
And what you use to narrow it down is what we call heuristics. Heuristics constitute what I call the art of good guessing.
But it's based on knowledge. It's based on either real hard knowledge about the world and you can therefore cut off some avenues of search because they're not in a, because we know better than that.
Or it's just based on experience. We know that certain things work in the past, other things didn't work in the past. We have some common sense, particularly in the area of, in specialties where we're expert,
we have really good expert common sense about the area. If you're a doctor and a patient comes in who's, if the patient is a man, then you rule out all things that have to do with symptoms having to do with pregnancy.
So that's just good, that's factual knowledge about the world and good common sense.
So heuristic programming became the essence of the Big Bang.
And it was carried on later into the second phase, which I'm going to tell you about right now.
When I got to Stanford in 1965, I had been, I had been at Berkeley for four years, completing, actually I worked with Herb Simon long distance.
We didn't have the internet at the time. We didn't have any network at the time. The interaction with Herb Simon was via the post office and telephone.
But the field had divided in the Big Bang. It had divided into two pieces.
The people who wanted to construct machines that were as smart as possible, whether or not they were doing anything like what people were doing.
And those are the people who used the banner artificial intelligence.
And then there were the people who wanted to construct realistic and testable models of human information processing.
That is, they wished to behave like psychologists.
That their models were constructed not out of intricate verbal arguments, such as you had find in the book by Bruner Goodnow and Austin, a study of thinking, which was the hot book at the time in the 1950s.
And not like mathematical expressions like you had find in the work of one of our most famous psychologists Gordon Bauer here at Stanford or Dick Atkinson, who later became president of the UC University California system.
They were doing mathematical or Pat Supeys. They were doing mathematical psychology, which turned out to be a dead end.
But these were information processing models, which you could write programs for and test.
So for example, my-
You're working now.
Yeah, for example, my work on human verbal learning, that line of research had started with Ebbinghouse back around the turn of the 20th century.
And psychology. And psychology. And they were very- there was a very large amount of stable experimental data about experiments that had been taken place in learning verbal materials, particularly those verbal materials that were not tied to real world meaning.
We call them nonsense syllables. And in my thesis, I constructed, I hypothesized and then programmed and then tested mechanisms, information processing mechanisms that would account for those experimental results.
And it became the best tested theory of verbal learning behavior at the time, quite well known.
And it's lasted right up to the present day. There are still people doing research using the e-pam model of verbal learning behavior.
That became known as cognitive science. Originally known as simulation of cognitive processes.
So when I edited with my colleague Julian Feldman edited the first book in AI, which was called Computers and Thought. It was an anthology.
We divided that anthology into two parts. Simulation of cognitive processes, which were basically psychology oriented papers, where the original publication were in places like the journal psychology,
the psychological review, and other work was artificial intelligence, which published in computer journals essentially.
So that's the work I was doing, but I decided when I moved to Stanford to move from the psychology end of the field, which I was not too happy with, to the artificial intelligence field, which had been my dream of the science field,
which had been my dream from day one to do what, again, one of Turing's young people, IJ Good, turned out to be a famous statistician.
He called the ultra intelligent computer. He had a famous paper called "Toward the ultra intelligent computer."
The problem that I was interested in was the problem of induction, not theorem proving, not discovery of proofs, not puzzle solving, which is a constituted many of the tasks that were done in the Big Bang period, but rather the problem of inducing hypotheses from a lot of data.
And my thought was, this is what we all do, almost all of the time, data is pouring into us through our senses, and what we're doing is forming and continually reformulating a model of our world. What's going on around us?
Well, the people who are supposedly experts at this particular area are the people we call scientists. That's their job to take in data and form hypotheses about what's going on from that data in their field of science.
So I expressed that interest to a newly found Stanford colleague, Joshua Letterberg.
Josh had gotten, he was chairman of the genetics department and a Nobel Prize winner in medicine and physiology, 1958, I believe, or '59.
But Josh had gotten interested in computing in 1964, had written a rather sensational program which could serve as the basis for one of these programs that I've, these induction programs that I was describing.
And so we got, he said, "Ed, I've got just the problem for you. It's the induction of hypotheses about how organic chemical molecules are structured, where the atoms are, and how they're connected to other atoms. Given a whole bunch of data that comes from an instrument, rather modern instrument at that time, called a mass spectrometer.
I won't go into what that is, but it was a new instrument. It was particularly interesting to Letterberg because he was helping to design the instrumentation that would go on the first planet probe that landed on Mars, looking for signs of life.
And in particular, looking for life precursor molecules called amino acids.
It's hard to do wet chemistry in a lab that lands on Mars, so you wanted some kind of instrument that would be able to do everything electronically and report back the data.
Oh, yes, I have seen this kind of a molecule. That's how Letterberg got started in that particular problem.
We began work on that, and as I mentioned, the essence of the artificial intelligence work is what I called heuristic programming, the application of heuristics to tasks like that.
The program that Letterberg had written to elucidate structures gave vast amounts of structures for even simpler molecules.
So what we wanted to do was to cut that down, given the knowledge that we had of chemistry, given the knowledge that we had of the specific expertise in mass spectrometry, what's the best candidate?
And then we would give that to the chemist.
It turned out that we ran out of what Letterberg knew pretty quickly, because he was working on amino acids only.
So what we did was recruit the expertise of the head of the Stanford mass spectrometry laboratory, turned out to be another genius, Carl Gerassi, the inventor of the birth control pill.
And Letterberg and Gerassi and I and our colleagues in the computer science department and in the mass spectrometry lab did a program called Dendral, or heuristic Dendral, that actually achieved levels of performance on these tasks better than PhD level postdoc level problem solving in Gerassi's lab.
And that gave us the courage to move on to do the same task for medical diagnosis, for example, in diagnosis of blood infections, diagnosis of pulmonary disease.
One of those programs is called Myson, the other was called Puff.
We did it for several engineering problems.
We even did it for a problem that our sponsor, the Defense Department sponsor, ARPA, same people who paid for the ARPA net, which was the precursor to the internet.
They wanted a secret one done to detect Soviet submarines off the West Coast.
And we did it for that and eventually it became a, it led to the formation of companies in Silicon Valley that did that kind of work.
Indeed, one of those projects that Letterberg and I did was called Mylgen, said for molecular genetics, it was the first program that did what we now call computational molecular biology.
Remember, for the molecular biologists, the key things are letters, ACGNT, in the DNA, not numbers.
And so we were able to do symbolic manipulation that was a powerful assistance to molecular biologists and started a company called Intelli Genetics, which did that for the pharmaceutical industry and other biologists.
So that's basically the Phase 2 story.
Well, that's great. So we'll take just a short little musical break and we'll be right back to get the rest of it. Stay tuned.
We're back on air with Professor Edward Feigumban, continuing our story about artificial intelligence.
So we went through the prehistory, we went through the Big Bang and then spoke about your own work about induction and I'm eager to hear what comes next.
Well, the work on induction, as I said, led to a great body of work which dominated artificial intelligence for a period of maybe 10 or 15 years called expert systems, which are these programs that perform at high levels of intelligence when we give them a great deal of knowledge about some particular area of expertise.
And that's excuse me for interrupting of his set where your famous phrase you're known for that through knowledge comes power.
Is that what it is? Is that the more knowledge you put in there, the more power you get from the...
Yeah, that's what I would call the second Big Bang. In 1968 when, later Bergen and my colleagues and I were going over the results of the original work that we did together,
we had a set of experiments which showed definitively that the program's behavior became better and better in terms of quality of chemical work.
The more knowledge we gave it of the domain in which we were working, namely chemistry and mass spectrometry.
So that's when the, of course, is as famous phrase, "knowledge is power." It wasn't used in anything having to do with technical language when it was when Bacon first used it, but so I changed it into "in the knowledge lies the power."
That was in contrast to the work that happened during the Big Bang where people were concerned about methods, search methods, heuristic methods, but they weren't concerned about knowledge.
So by the mid-70s, two of the famous researchers at MIT wrote a paper in which they talked about AI's shift to the knowledge-based paradigm.
That shift was initiated here at Sanford and it still dominates today.
So there actually is a brilliant counter-example to that which I like to cite because it satisfied one of the early predictions of Herb Simon actually,
that a such a program would beat the world's chess champion. He had predicted that it would happen in ten years, but it actually took about 40 years and that was a program called Deep Blue that used rather
a small amount of knowledge about chess, but did a tremendous amount of searching in that maze I was telling you about. If I do this and he does that and I do this, it looked at 250 million alternatives per move, roughly.
That was in the 80s? Yeah, that was in 1997 or 8, Deep Blue beat Gary Kasparov, the world's chess champion.
Now Kasparov didn't search anything like he searched maybe a few hundred or a few thousand alternatives, but he started off with some superb pattern recognition, some experience that he had that said that certain patterns were great patterns.
And he didn't have to do all that amount of searching, but the searching beat him. And that was a counter-intuitive thing to the people like me who were believing in the knowledge's power principle.
That whole set of ideas is called the knowledge versus search spectrum, and it's very important in artificial intelligence because we've been able to beat the search game by putting tremendous amounts of knowledge in about particular specialties of the world, or large bodies of common sense knowledge about the world.
But what we haven't done is to exploit large amounts of computation, which are now available and weren't available then. So that really is part of the future of artificial intelligence is to look down avenues like that for combinations of symbolic concepts that are novel and interesting that induce a sense of awe in the knowledge of knowledge.
And the sense of awe in the human audience who sees it and our new demand kind, just like you had in that move that Kasparov, where Deep Blue beat Kasparov, there was actually in the audience there was among the grandmasters and masters watching it, there was a sense of awe.
And Kasparov, who himself is not a religious man, said it was like looking into the mind of God to see that move emerge.
So the relationship between knowledge and search, so initially you would put as much human knowledge, let's say, into the machine and that the more knowledge you would put in the more it would arrive at the desired results.
But then quickly enough, I suppose that the machine can generate knowledge rather than just merely receive it in process, is that correct?
That's a really great question. One of the parts of the field that received some attention, but not a lot of attention in both of those early periods, a big bang period and the expert system period, was a part of the field called machine learning.
Some people thought of that as the critical element for an intelligent program or an intelligent human being, that it learns from experience.
But our efforts up until 1980 were rather minimal in my own project with Josh Lederberg and Bruce Buchanan and my team.
We were able to learn, excuse me, our programs were able to learn the knowledge necessary about certain families of molecules that was necessary to solve problems in particular.
We thought of that as a great achievement. That was the first time that any new knowledge about a field got published in a journal of its own discipline, not a computer journal, but in this case, the journal of the American Chemical Society.
So we thought that was fantastic, but it turns out machine learning didn't go in that direction. It went in the direction of statistical pattern recognition, which means that the AI people join forces with the statisticians, and it's very hard now to distinguish a machine learning theorist from a statistician.
It's at the big data stuff that we hear someone get out. Exactly.
Now, the people who do inference from big data don't always use machine learning algorithms, but a great deal of the time, if you read the Forbes magazine or New York Times or something, you'll see the terms come up together, big data and machine learning come up together.
Now, the field developed extremely sophisticated statistical methods for doing pattern recognition starting with that work in 1980 and going right up to the present.
Now, if you do that, you're not doing a great deal of deep cognition. You're doing surface cognition. You're doing recognition as opposed to deep cognition.
So there actually, a world famous psychologist named Daniel Kahneman wrote a book in the last two or three years ago called Thinking Fast Thinking Slow, Wonderful Book.
And he distinguishes between the two kinds of thinking. The thinking fast is the perceptual recognition of things. And that's the kind of thing that AI programs are doing very well now using these machine learning statistical methods that hook up to big data, as you said.
So the thinking slow is when we really get down to thinking hard about a difficult problem that takes us minutes, hours, weeks and months, or in the case of Einstein years to solve.
The thinking fast programs have led us to intelligent sensing devices that are spectacular.
For example, the sensing programs, sensors and programs that help guide the Google self-driving car. The Google self-driving car is only now getting into which you might call the thinking part of driving.
But it's been sensational at the recognition part of driving, recognizing pedestrians, recognizing bicyclists, recognizing traffic lights.
Then you've driven in one hour. Oh yeah, I took a ride in a Google self-driving car about three weeks ago through the sit-through mountain view, up shoreline Boulevard and up across California Street over to Ringstorf and back to the freeway.
It was really amazing, but one of the things that the Google people right now are doing is exercising this knowledge principle that in the knowledge lies the power. They have to know a great deal about the streets of mountain view in order to drive carefully in mountain view.
You have to know the traffic light configurations, so you know where the red light really should be, so you can look there. When it stops outside of a zone that says, "Keep clear," I knew damn well that it wasn't reading K-E-E-P-C-L-E-A-R.
It's that the team had marked that as a clear zone, and it knew enough about that not to stop inside the clear zone. In fact, it's not allowed to break any of the rules of the California automobile driver's handbook manual.
It has to know a great deal about mountain view, and it will have to know a great deal about San Francisco, if it's going to drive in San Francisco.
And any of the cities that it drives in, hence we arrive at what AI people call the knowledge bottleneck. It's going to have to learn those things.
In some way, otherwise it's going to have to be told everything about those. And right now, the Google car is being told all of those things about mountain view.
So, but it works very efficiently in mountain view because it has such an extensive knowledge of the street. It's actually spectacular. But it's wildly challenging to get that level of expert knowledge about all the cities in the United States.
But I imagine if anyone thinks big it's Google. Yeah, it made me think that there's an opportunity for a wonderful business there, which is to provide other manufacturers for it and Toyota and Audi with that body of knowledge for all, and Google isn't planning to go into the car business, but I don't think.
But anyway, I told that to someone at Google and they said, "Of course, we know that. That's what we're aiming at."
Just to own and sell that body of knowledge.
So, is a car like that add something that excites you as is it also something that you consider to be one of the fruits of your own research over the years?
I see the, or is it only fast thinking and that you're more interested in slow thinking?
It turns out that for me personally, I'm interested in slow thinking. But what is very important for the artificial intelligence field is to have what are called integrated systems that combine all the elements of fast thinking and slow thinking into a working system.
One another way to put it is that computer science mostly and artificial intelligence in particular are experimental disciplines, and we learn by watching what we do and what mistakes our programs make, and we learn how to make them, we humans learn how to make them better, and we learn how to build learning programs to make them better.
But we can only do that if we have integrated systems that do the complete job, albeit badly, even if it does it badly, it will get better.
Something like the sire of the iPhone.
Yeah, and so how does that, such an integrated system, is it a combination of fast and slow thinking?
Well, Siri, it's called Siri, the sire, if you pick out the letters sRI, it means that the little company that Apple bought came was founded by people at sri.
And then we Stanford Research Institute.
What used to be Stanford Research Institute was spun off from Stanford during the Vietnam War demonstration era.
So they added the letter I to make it pronounceable, and Siri was developed as a consequence of an absolutely huge DARPA program, the same defense department sponsorship, defense advanced research project agency, it was a $300 million project.
It's a personal assistant, and Siri was the spinoff of that project, further developed by Apple, of course, as a commercial effort.
Now Siri does not do a lot of perception, so it doesn't, it's machine learning is not machine learning of sensations, because the Apple iPhone really doesn't have that many sensory issues.
It doesn't have that many senses on it.
But what Siri does is to, has done a lot of machine learning on natural language text that exists out on the internet.
The whole field of natural language processing, which is central to the artificial intelligence field, because language is so central to intelligence, has really switched to the artificial intelligence field, because language is so central to intelligence.
Has really switched over to become statistical natural language processing, in which the meaning of words is implicit in the statistical juxtaposition of certain adjectives with certain nouns, at certain adverb, that certain distances from each other, and the big machines that we have today can process billions, and I say billions, you know, the Carl Sagan, Billions, and you know,
the Carl Sagan, Billions, and Billions, and Billions, actually in the case of IBM's Watson program, another spectacular program will talk about in a minute.
They claim a trillion pieces of writing, and so Siri uses that, it also uses huge bodies of knowledge.
For example, it's linked up to one of the largest knowledge bases in the world. That's a thing called Wolfram Alpha, to which you can ask an absolutely enormous number of factual questions and get factual answers.
So Siri is an integrated system for coupling voice understanding, which itself was an artificial intelligence problem that was gradually solved over the years to be almost perfect now.
Voice understanding with textual understanding, trying to infer what is your meaning in the question, and generate an answer for you.
Now it doesn't always generate the right answer in just the same way that Watson, when it did its Jeopardy program, challenging the Jeopardy champions of all time, and Jeopardy didn't quite always get the right answers.
Sometimes it got really silly answers. Sometimes that's the case with Siri also, but that's the way it is with people too.
Sometimes people don't understand what you're saying, and they give you the wrong answer, and they have access to wrong knowledge.
So Siri is regarded by the artificial intelligence community as a real triumph. It's not only Siri, it's Siri and it's Microsoft competitor, and it's Google competitor by now.
So there are three of them out there. So how far away are we from a data who is a character in the Star Trek next generation?
He's the Android who not only has all the enhanced capacities of Siri exponentially multiplied, but it's also a sentient thinking moving syntheticization.
Synthetic human being in every respect except for the fact that he was not biologically generated.
Yeah. Great question. Let me answer the question by challenging the in every respect.
That's the key thing about data. Mr. Data.
He is not, we can tell right away he is not human like because Mr. Data is like an expert system.
He really knows a lot of specialties. What we love about Mr. Data, what makes him so cute on the program, is there's all that stuff he doesn't know that all the other humans on the ship know, but Mr. Data is so narrow.
He only knows about his specialties. Now with the exception of fluid natural language processing, Mr. Data is what we have now in our expert systems.
We can construct expert systems that do the things Mr. Data does when he does his specialty work.
What we haven't been able to replicate yet is the fluidity of the natural language processing, but we're getting there.
So now when you ask about the other characters on that program, you see real human beings.
Those human beings know an enormous amount more than our programs today know.
And when I say no, I don't mean no in the sense of Wolfram Alpha specific hard knowledge about the world.
I mean experiential knowledge, the knowledge of everyday life, common sense knowledge.
Well there's also this phenomenon that I'm curious about which is in philosophy, it's called attunement.
To be human is always to be attuned to the world, to your environment, to others.
It's almost a musical metaphor that you're in tune with things and that sort of attunement which on that Star Trek show you see,
the human beings being attuned to each other, so many things are just natural, taken for granted.
And Data doesn't, Mr. Data doesn't seem to have that attunement, although he's getting close.
Well Robert, there we go back, you're exactly right, and there we go back to that principle, the knowledge principle, in the knowledge lies to power.
Data never got that knowledge.
And what amuses us in the program is the occasional times when Data begins to dawn on him, that there's something else that he doesn't know, that it's a world outside of what he knows.
And we giga, and we say oh yeah of course, Data doesn't know that.
But isn't that also the origin of consciousness in a certain way where they're Aristotle through a bunch of other people have this notion.
Have this notion that thinking begins in wonder, it's when you become aware of the fact that you don't know something, that there's some incomprehensibility about the world or something that you're aware of, and that it activates this process of thinking about things.
Well, I think that there are two, my own feeling is that there are two separate questions there.
The first is something that we're really on top of, the research community and artificial intelligence, and that's the question of how can we know when it is we don't know, so that we can set up the learning process, so that our programs can set up the learning process to begin to figure out what it is it doesn't know.
That we have a reasonable handle on in artificial intelligence, we're moving in that direction, and that indeed is a big research direction.
Knowing what it is you don't know to be the start, the framework of machine learning.
The question of consciousness is an entirely different question that ranges from what you might call the "excessor"
what you might call the "excessor" some people view as the extremely fundamental and other people view like me view as an epiphenomenon.
The epiphenomenon view says that if you look inside some of the programs that I've described, particularly the one, the early one I described, E-PAM, you find components in there like the immediate memory.
We know what that is, that's heavily research in psychology, it's the very most short-term memory, plus or minus seven symbols, seven plus or minus two symbols.
We have the working memory, which is in the order of 300 to 1000 symbols, and we have the long-term memory.
We know that items in the working memory are passed into the long-term memory overnight, usually, or over a longer period of time through a brain mechanism called the hippocampus, the transfers memories from working memory to long-term memory.
During that period of accumulating symbols in these different working memories, we have an internal glimpse of the world that we're dealing with.
It's not 100% glimpse because there's all this unconscious work that's going on in the long-term memory that we don't have access to because those symbols are not available to us for working.
So the working memory is the consciousness, that's the EPI phenomenon view. It just says, yes, of course these machines have a consciousness, but that consciousness consists of only the things that they're working on.
And when we say that we're thinking of something and it takes place, oh, we woke up in the morning and suddenly we knew it.
That's because other processes were working on our behalf, but they were working outside of the working memory to deliver a result.
Now other people think that consciousness is ingrained in nature itself. You can read physicists who think that consciousness is actually a quantum phenomenon.
You can read the book by Sir Roger Penrose, who thinks that consciousness is really built into the quantum universe.
And there are all shades in between. But the great book to me, the single book someone should read if they're interested in the subject is by the philosopher Daniel Dennett of Tuft University called Consciousness Explained.
That is the single best book I would read if I were interested in that subject.
Well, here at Can I read you something I get from a paper that you yourself published in the 90s and where you're talking about the far side of the dream of artificial intelligence.
And it's a view from the future looking back at our present from Professor Marvin Minsky of MIT, who says, can you imagine that they used to have libraries where the books didn't talk to each other?
And you go on to write the libraries of today are warehouses of passive objects, the books and journals sit on shelves waiting for us to use.
Our intelligence to find them, read them, interpret them and cause them finally to divulge their stored knowledge.
Electronic libraries of today are no better. Their pages are pages of data files, but the electronic pages are equally passive.
And then you go on to ask us to imagine the library as an active intelligent knowledge server where these books would actually talk to each other.
So two questions here. One is as literary critics or historians, but especially literary critics, we look at the history of literature and all these books and part of our work, a great deal of our work is get these books to speak to each other.
Through our own hermeneutics and analysis and interpretation, making connections between them and coming up with a history of literature or interpretation.
So that would be one question, is that something that we've been doing natural?
I mean, we've been doing humanly in any case? And the other question is, how do you envision this future,
what you call here the intelligent knowledge server that would get these books in the library to speak to each other independently of our human agency?
Back though, in those days, that was with the minor exception of a few disciplines like I mentioned some of them, a few chemical disciplines, a few medical disciplines, few engineering disciplines.
The knowledge was all dead stuff. It was words printed or formulas printed on paper stored in libraries.
And that's in one way, that's a miracle that we were able to do that. We as human beings, we don't find animals below our level, writing up their stuff, writing down symbols, and transferring their knowledge, one baboon does not transfer its knowledge to another
baboon through writing and books and storing that material and passing it on to generation after generation by the act of reading.
Teaching reading. So it's miraculous that we, that's the essence of what we do as a culture.
Not all of our knowledge is passed through books. Some of it is experiential, but that is mostly a person by person. One-on-one, we gain experience and we get some knowledge that way, but as a culture, we pass our knowledge along through these books.
Now, we're not passing along dead symbols, ink on paper. We're passing along thoughts, concepts that are represented by the ink on paper.
The problem with the traditional library is it's dead. It sits there as ink on paper, not as actionable thoughts where you could pose a question and then those actionable thoughts would deliver,
an answer or a concept or a suggestion. None of that is available through a dead book.
So the quote that you have there says, let's try to make all of those books live in that symbolic sense of tying itself to other knowledge, tying itself to language, tying itself to the ability to communicate to human beings and
to other sources of knowledge that relate to it, and analogical ones, or related in some other way, metaphorical ones.
So, how close are we to that? Well, in one sense, I mean, there is a sense in which we are amazingly close to it.
Anybody who's used Google and uses it with any level of skill knows that the answer that you're looking for from these, as I said, billions and billions and billions of pages usually comes out right on the first page.
That's amazing. On the other hand, the Google people will be the first people to tell you that there's no understanding that's involved in that.
There's excellent keyword search and ranking of other people's search queries. So, realizing that, I remember one of the founders of Google who gave a talk at Stanford once saying that Google is an artificial intelligence company that just doesn't realize it's yet.
Well, they began to realize it, and they actually bought a couple artificial intelligence companies that had some very large knowledge bases, and they incorporated that material and build on it in a thing called the knowledge graph.
And now, if you use Google, you'll see the knowledge graph of what you asked, come up on the right-hand side where there used to be ads. Now you find a great deal of knowledge about the subject that you were talking about. So, things are just beginning to become knowledge rather than just symbols that are dead living on old trees.
Well, it's been quite a ride over the last hour, and you've taken us through so many different disciplines. I'm starting to realize that the work that you do over there with your colleagues in artificial intelligence really engages almost all our disciplines here in the university. There's psychology. There's language learning and symbolic systems, obviously chemistry.
And every sort of thing, so it seems like you can take the whole knowledge that we are involved within the university and put it to work in your artificial intelligence programs.
But we're all of us who teach here are proud of Stanford being one of the world's most influential and most excellent interdisciplinary centers of thought.
And in this case, that magic has been worked over the last two decades. We've had an undergraduate program called the program in symbolic systems, which cuts across psychology,
linguistics, and computer science. And it's been one of Stanford's most successful interdisciplinary programs, and student interest has been very high in the symbolic systems program.
A huge amount of majors, I gather. Yes. You can have, that's right, you can major in symbolic systems.
And I agree, I think that Stanford, something about this university, promotes interdisciplinary, and it's the kind of place where I would imagine, not every place could bring together these things away, the way we do it here.
Well, at Stanford, interdisciplinarity, if you can use that word, is highly rewarded in the faculties of the different departments, as opposed to other places where in solidarity, publishing as many papers as you can in the same journal, is unfortunately, that's what's rewarded.
That's great. Well, Ed, thanks for coming on. I want to remind our listeners, we've been speaking with Professor Edward Feigumbaum from the Department of Computer Science, and you can check out his entire bio on our website. It's very extensive. I mentioned it parts of it at the beginning of the show. So thanks again, Ed, for coming on, and thank you all for listening to entitled opinions.