November/December 2001

An Evening with Ray Kurzweil and Howard Gardner

Every month, the Harvard Graduate School of Education invites educators, researchers, community activists, and policymakers from across the country to talk about key issues in schools and school reform. We are pleased to be able to provide you with an edited transcript of some of these forums. Below is an edited transcript of a talk given at the Harvard Graduate School of Education, May 15, 2001.

For easier reading, we have divided the transcript into the following sections:

Introduction by Howard Gardner, Professor, Harvard Graduate School of Education
Discussion between Howard Gardner and Ray Kurzweil

You can also scroll right through the transcript without clicking on the above links.  

HGSE Forums Home Page
Transcripts of Past HGSE Forums
HGSE feature on the Kurzweil/Gardner Forum, May, 2001

Introduction by Howard Gardner, the John H. and Elizabeth A. Hobbs Professor of Cognition and Education at the Harvard Graduate School of Education

Tonight we're privileged to have one of the leading thinkers and makers of our time. I say "maker" in the sense that Ray Kurzweil is one of a handful of individuals who have really helped to define our era. Ray is an inventor and has invented many important practices, objects, and even fields of future study that we'll talk about tonight.

Ray is known also as an author of books. His most recent book is The Age of Spiritual Machines, and I'm sure it's known to many of you. He also has written a book about improving your health, and he has a new book in preparation on singularity, which he'll probably talk about this evening.

I wanted to mention some of his inventions because they're quite dazzling. Ray was really the first person to do serious speech recognition and to do reading for the blind, where you have texts that are read by machines so that blind people can know what's in them. He has invented musical instruments [synthesizers] that are incredibly real in sound, so much so that people can't tell which is the piano and which is the Kurzweil. He's been involved in software that writes poetry. He's been involved in medicine with the creation of medical reports, and also education for physicians that allows them to work on virtual patients.

We're all used to hearing that the people who speak at the Forum have lots of honors. I think Ray has won just about every honor, every award given to people who are inventors. It's really quite dazzling. And these are from the most prestigious academies, universities, and institutions.

Those of you who have read the newspaper recently know that he won probably the most prestigious award, the Lemelson-MIT prize, some weeks ago. I'm sure there will be others, but that's a very special punctuation mark in a distinguished career. And everybody remarked when he walked in how much younger he is than we all thought he would be. But that's part of the secret of youth, which we'll also talk about later this evening.

Discussion between Howard Gardner and Ray Kurzweil

HOWARD GARDNER: We [Ray and I] never met until half an hour or so ago, so I'm very excited about the chance to ask him some questions.

RAY KURZWEIL: We did meet in auditory virtual reality, which is to say on the telephone.

GARDNER: Correct. I'd like you to reflect a bit about what I would have seen you doing on a typical day [when you were] around the age of 12.

KURZWEIL: Probably hanging around Canal Street in Manhattan, which is where you could pick up used electronic equipment, carting that back to my home in Queens, and then taking it apart and reassembling it.

GARDNER: When did you first learn about computers?

KURZWEIL: At the age of 12, which was 1960. They were somewhat less ubiquitous than they are today. I had a part-time job doing analysis of variance, which is a psychological test, for Head Start. I did research on the Head Start program on these old Fried electromechanical calculators, where you punch in numbers, and then the machine would go cachug, cachug, cachug, until it did the calculations. It was actually carrying out the algorithms electromechanically. They had scores of people doing these tabulations of sums of squares and so on. So I found an IBM 1620 computer that they had access to, and we programmed that test for the computer.

GARDNER: So you programmed a statistical test?

KURZWEIL: That was the first little venture, yes.

GARDNER: But then you went on "I've Got a Secret," which will also separate out the old timers from the rest.

KURZWEIL: I came on and played a piece of music on an old upright piano. Then they showed the secret to the audience, not to the celebrity panelists. The secret was that I built my own computer, which back then was considered unusual. Steve Allen said, "That's impressive. But what does it have to do with the piece of music?" Then the second part of the secret was that the computer had composed the music.

GARDNER: Tell us about the computer and how it composed music. This, again, is in the early Sixties?

KURZWEIL: Yes, this was in 1963 or 1964. That was my first venture into pattern recognition, which is my primary technical area of interest, which I believe is the foundation of human intelligence, or at least the various different intelligences we have, the different skills that go into those intelligences--

GARDNER: I paid him a lot to say that.

KURZWEIL: The 8.5 intelligences that we have are all using a paradigm of being highly parallel and self-organizing and following some variation of what we would call a pattern recognition paradigm. This is being able to organize a lot of diverse elements and finding patterns that are not previously programmed. We are realizing that the bulk of human intelligence follows that type of paradigm. So I've been very interested in pattern recognition.

So I fed in compositions of classical composers that would build a pattern recognition model of the patterns apparent in those compositions, and then create original melodies with those same patterns. And it worked well enough that you could identify which composer it had been based on. So it would sound like a second-rate student of Mozart or Chopin.

GARDNER: That's pretty amazing.

KURZWEIL: It was interesting, and it showed the power of pattern recognition. It had a certain mathematical model of the kinds of patterns you'd find in musical melodies. It covered kind of some superficial qualities, I think.

David Cope's work, which is in more sophisticated experiments in musical intelligence, follows a similar paradigm, but really does quite an outstanding job of replicating these composers. For at least a brief period of time, Cope's program would fool you into thinking you were listening to some new Chopin prelude. But if you listened to it long enough, and if you know Chopin well enough, that illusion would dispel itself. We haven't passed the turning test yet in musical pattern recognition.

GARDNER: In our virtual meeting yesterday, we talked about speech recognition. I'd like you to give us a feeling for what you've been trying to achieve over the last 15 to 20 years, what the steps of progress have been, what the software could do initially, and what it can do now.

KURZWEIL: Well, speech recognition is an interesting model, because it incorporates, first of all, multiple levels of knowledge. Speech or language is not one model that we can just feed into one neural net. If you just feed speech into one big neural net, it won't do a successful job because language is a hierarchical phenomenon that has knowledge embedded at different levels. There is knowledge of how the vocal chord is constructed and the kinds of sounds it makes.

GARDNER: We might be thinking about this as we listen to Ray, because essentially we're machines that do know how to recognize speech.

KURZWEIL: Right. And then we form different phonemes, which are the letters of speech. When we burst our lips, we make a plosive sound. If the vocal chords are moving, the plosive sound is voiced. When we let air go through our vocal tract and just resonate, we create vowels like ah and e. So we form the different speech sounds, but everybody has a different way of articulating them.

We put them together in different sequences according to the different dialects of language to form words. And then the words form concepts, and there are basic syntactic rules that vary from language to language. And then at the next level we have semantics. And to really understand language, we realize that you really need to know something about the knowledge that's being expressed.

To understand language at a human level, you really have to embody human knowledge and human intelligence, at least as far as language is concerned. We're not able to embody our machines with all human intelligences, but we have to give [a machine] enough semantic knowledge so that it can disambiguate words that are homonyms, for example. It needs to be able to figure out what a word means from context. So we provide knowledge at many different levels.

GARDNER: Ray, you've given us such a vertical picture that it seems hopelessly complicated. When you first began doing this, what were the bets that you made that you could at least break into the problem of speech recognition?

KURZWEIL: Well, we realized we would have to identify applications that could match the limitations of early speech recognition. So we built systems that had limited vocabularies. With the first systems, you had to train every word. We used them for medical reporting because we could structure the reports, and medical reports have a lot of value if they are accurate. So we found those early applications. But we realized that in order to make linear progress in speech recognition, we needed exponential gains in computing power.

GARDNER: Translate that, because I know that's very important.

KURZWEIL: Well, as the vocabulary size gets larger and larger, the number of combinations of words that you can put together becomes astronomical. And if you have continuous speech, which is the way I'm speaking now, as opposed to putting pauses between words, then you don't know where the words begin and end. So there is a tremendous amount of ambiguity. The number of possibilities becomes quite vast.

[There's the need] to explain vocabulary, increase accuracy, handle phenomena like continuous speech, handle speaker independence, which means you don't know who's speaking. Also, everybody has a different dialect and a different way of forming the phonemes. The amount of variability becomes quite enormous.

So you need to make exponential gains. For example, you need to double computing power just to make a small linear gain in accuracy. We actually have models of this. If you multiply computing power by a factor of ten, which means you have ten times as much computing power, you might add 2 percent to the accuracy of a system like this.

But the fact is, we are making exponential gains in computing power, so that we're able to make linear gains in speech recognition. We still don't have the Holy Grail, which is extremely accurate systems that are fully speaker-independent, that can handle a large vocabulary and continuous speech. But anyway, it's a good sort of classic example of pattern recognition. And it's dealing with human language, which embodies the full richness of human knowledge. And we won't really have systems that can fully understand human language until they're operating at a human level. That's a whole other discussion.

GARDNER: Help us understand this. You're working with other technical people to build machines that can perform functions, whether it's a fairly limited vocabulary or whether it's getting closer to ordinary human discourse. Meanwhile, there are people at places like MIT studying linguistics, and learning a great deal about the properties of language qua language. And there are people all over studying the brain, including the auditory system, and its regions and so on. What is the interplay between what you do as somebody who is really trying to create the best possible product of a sort, and the insights and discoveries, and false leads and blind allies, that are going on in these more pure sciences?

KURZWEIL: That's a good question. I have an interest, both as a technology creator [and as someone who is] trying to develop theories about future technology and human intelligence, in the extent to which we'll be able to model that in nonbiological intelligence. One connection is that we are further along than people realize in reverse-engineering the human brain.

GARDNER: Explain that term please.

KURZWEIL: It means understanding how the brain works, in a detailed fashion, from brain scans and from detailed mathematical modeling of human neurons. There are many different technologies that are growing exponentially. You are probably all familiar with the exponential growth of computing. Sometimes it's called Moore's Law. Moore's Law is actually just one paradigm, which is the shrinking of transistors on an integrated circuit. Every two years you can put twice as many transistors on integrated circuits, you have twice as much stuff, and they run twice as fast, so it's a quadrupling of computing power.

It turns out that that's only one method for bringing exponential growth to computing. It's actually the fifth different paradigm. We had electromechanical calculators, relay-based computers, vacuum tubes, transistors, and then integrated circuits. And integrated circuits will run out of steam within 10 or 12 years, when the key features are only a few atoms in width. But that won't be the end of the exponential growth of computing, because then we'll go to three-dimensional molecular computing. Our brain is organized in three dimensions. We might as well build our computing circuits that way.

So computing has been growing exponentially, and it's not limited to Moore's Law. But it turns out many other technologies, particularly if they relate to information, are also growing exponentially. Communication speeds are growing exponentially. The size of technology is shrinking exponentially.

DNA sequencing is also growing exponentially. When the Genome Project started, it was actually not considered a mainstream project. It was considered kind of a flaky, radical project, and mainstream critics said, "There is no way you're going to be able to do this, because at the rate at which we can sequence DNA it's going to take 10,000 years to get through the genome." But the speed of that, the cost-effectiveness, has grown exponentially. We went from $10.00 per base pair to a tenth of a cent today in ten years.

Another technology that's growing exponentially is brain scanning. The speed, the resolution, the bandwidth of being able to see inside the human brain and actually map what's going on there is growing exponentially. I've got various charts that people can look up on our website, which I might mention later, that show the exponential growth of these different technologies. So we are learning about what's going on inside the human brain.

The brain, as we were talking about earlier, is not one big tabula rasa, one big neural net that just kind of organizes itself. It's actually hundreds of different regions. The different regions do different things. They have different ways of representing information. They have different ways of processing the information. So it's hundreds of different information-processing organs in a very intricate, somewhat ornate architecture. And we're actually in the process of scanning that, understanding the detailed wiring of the brain, and understanding those methods.

Now, we're in the early stage of that, but we have shown the feasibility of the project. If you take the analogy to the Genome Project, we have shown earlier, some years ago, the feasibility of scanning base pairs and getting the raw data of the genome. And we've also shown that we can understand the genome. We don't understand the whole thing yet, but we understand some of the genes and how they express themselves in proteins. We're beginning to understand how the proteins work. We are in the somewhat early stage of that process, too.

One interesting example of how far along brain reverse-engineering is, which is simply understanding and modeling the brain: There is some pioneering work done by the scientist Lloyd Watts out in California, where he has taken mathematical models of the different types of neurons in 15 regions in the auditory cortex, and the wiring information, and has actually built a model. He has understood the algorithms, which are not algorithms like those that would run on your notebook computer, because they're not sequential, they're not digital. They're digital-controlled analog. They're highly parallel. They're self-organizing. And there are actually different ways in which the information is processed and represented in these 15 different regions.

He has a model of that, and he's built that model into a computer simulation. And that simulation closely matches experimentation on the human auditory system in terms of our ability to do things like localize sounds and identify the sources.

GARDNER: I take it you wouldn't have been saying all of this if this kind of information didn't inform what you do.

KURZWEIL: Well, we took some of this information and used it in our speech recognition, even without understanding it fully or understanding why the brain does this. So we're already beginning to utilize some of this knowledge in our practical applications. But it also does fuel some of the theories I have about the future, in terms of projecting into the future the likely course of our understanding the human brain and being able to replicate it. How powerful a computer would it take to run a process like the hundreds of regions of the human brain, and when will we understand that? So it also feeds into that interest.

GARDNER: We will talk about your incredible, provocative ideas about the relationship between human beings and machines, but before we go into that, all of us have heard about Moore's Law and about the notion of things getting faster and cheaper, and so on. And you've said that these are principles that are at work everywhere.

I don't want to take this question more deeply than you want to take it, but why is that so? Why should it be the case that these laws occur in different domains? Is it the nature of technology? Is it the nature of human inventiveness? Is it the quantum level kind of thing? How does one think about that? Why won't it just stop at a certain point?

KURZWEIL: There are forces driving it forward. For one thing, it's an economic imperative that comes out of economic competition, as well as our quest for understanding, and just the human drive for innovation.

It's interesting. I've done a lot of studying of technology trends, which originally came out of my interest in creating technology. Because if you create technology, you want your project to be relevant, and appropriate, and on the cutting edge, and at the right point in time when it's finished, not when you start the project. And you want the market-enabling forces to be there, and you want the enabling technologies to be in place.

Most projects fail, not because the creators of the technology fail to actually fulfill their mission, but they are either too early, which is usually the case, and all of the various market forces aren't in place . . .

GARDNER: What would be an example that we might know of that?

KURZWEIL: Well, usually there are many different generations of a technology before something hits. Like electronic books, there have been several waves of those. And the display technology, which is a required enabling technology, isn't there. It doesn't have the contrast, the viewing angle, the resolution, that you need to replace paper.

There are obviously a lot of advantages. You can download books, and you can do a lot of things with the electronic information that you can't do in a paper book. But it's not pleasant to read if the enabling technology isn't there. But it's going to happen. So that's an example of something that is premature at this time.

So I became an avid student of trends, because you really want to know, if a project is going to take three or four years, what is the world going to be like in three or four years? It's not going to be like it is today. And that kind of took on a life of its own.

GARDNER: Talk about the life of its own that it's taken on. I think one of the most striking things in your writings is this notion that we all assume that things are going to change at a certain rate.

KURZWEIL: Why is this happening is a good question. One thing I've noticed-and it's probably the most important point I try to make-is the exponential increase in the rate of progress itself. And it does pertain to the issue of why this is happening. It's a human desire to improve things in a geometric way. We try to make things twice as good. We don't try to add to them linearly. And technology inherently grows exponentially for that reason.

We also use one generation of tools to create the next. Right now computer design is creating the next generation of computers, but they have very powerful computer-assisted design tools to do that. And they'll specify parameters at a high level, and the computers will generate ten different levels of stages of design automatically. Then other computerized systems will lay out the chips, and other computerized systems will manufacture the chips, and other computerized systems will automatically develop printed circuit boards to put the chips on, and put the chips on there automatically. There is this tremendous amount of automation, which means that a whole new generation of computers can be designed in one year, and it can be that much more powerful.

When I put different technologies or paradigm shifts on a double exponential chart, it's interesting to see that technological evolution follows seamlessly from the biological evolution that created the technology-creating species in the first place.

Things speed up with each new generation. The first stage in technology took tens of thousands of years: the fire, the wheel, stone tools. A thousand years ago a paradigm shift took only centuries. At the beginning of this century, a paradigm shift took only decades. And now something like the World Wide Web, which is clearly a powerful, transforming paradigm shift, took only a few years.

I put many different paradigm shifts, biological ones and technological ones, on this exponential chart, and it forms a seamless straight line. People say, "But your assumptions are wrong. This paradigm shift in the Cambrian explosion didn't take 10 million years. That took 30 million years. It's not right to say the Web took five years. That took ten years."

It's still a straight line, because it just doesn't change the line very much. The Cambrian explosion didn't happen in ten years, and the Web didn't take 50 million years to evolve.

GARDNER: Of course, one could accept your general point, which is an intriguing one, but still wonder about two questions. One, whether you're parsing it correctly, because when we talk about explosions and paradigm shifts and so on, that's a human invention, and somebody else might parse it differently. One could also ask the question, "Is this smooth, or is it more, in [Stephen Jay] Gould's terms, punctuated with periods?" Because when you think about cultural evolution, I would say it's much more punctuated. You have Rome and Greece. You have Florence . . .

KURZWEIL: In the technological evolution, what I find is a cascade of S-curves. A particular paradigm tends to start out very slowly, and it grows exponentially, then it finally reaches its potential, and then it levels off.

And as I gather more and more data on the many, many examples of technologies, rather than just seeing one continuous exponential, you see these little ripples of these S-curves, and you can identify them with specific paradigms.

So as a particular technology tends to level off, it generates a whole interest in research in how to overcome the limitations of that technology. And usually it comes from a completely different source. They were shrinking vacuum tubes to make them smaller and more efficient. They finally couldn't shrink them any more and keep the vacuums, and a whole different paradigm came along, which is the transistor, which is not just a miniature vacuum tube-it's a whole different paradigm.

The important point I wanted to make before is that according to the models I have, we're doubling the paradigm shift rate every ten years. There is a lot of data behind the model, and there is also a theoretical model behind how one stage of innovation provides much more powerful tools so that the next stage can go faster.

You can also sense it, just the amount of progress that's being made. But I have measured this, and that's a very important point. Because most people, even people who are thoughtful about the future, have a linear view of how things will unfold.

I had a debate actually a few blocks from here with Bill Joy recently. We'll probably get to some of the issues he's brought up. Very often Bill Joy and I are paired where I'm the optimist and he's the pessimist-although I will say that Bill Joy got his concerns from a bar room discussion that he and I had in Lake Tahoe a few years ago.

GARDNER: A footnote: Bill Joy is a major scientist at Sun Microsystems, inventor of Java and other languages. He caused quite a stir about two years ago by writing an article called "Why the Future Doesn't Need Us," in which he was Cassandra-like about what might happen to human beings because of nano-technology, and robotics, and genetic engineering, and so on.

KURZWEIL: So this one Nobel Prize-winning biologist said that these dangers are never going to happen, at least not for a long time. We won't see self-replicating nano-technology entities for a hundred years. That's actually a good estimate of the amount of technical progress required at today's rate of progress to achieve that particular milestone. Most people have a linear view of the future. You will very often see thoughtful scientists say of some particular biological technology, "That will take 75 years, 150 years. We won't see that for hundreds of years." They are assuming a linear model. They have an intuition that is pretty accurate about how things progress, and they're just projecting today's rate of progress.

But because of the power of exponential growth, we're doubling the rate of progress every ten years. So we'll achieve 100 years of technical progress at today's rate of progress in 25 years, which is really Bill Joy's and my estimate of these dangers. It's a very different view of the future. In the 21st century, we'll see 20,000 years of progress at today's rate of progress. We saw only 25 years of progress at today's rate of progress in the 20th century because we've been speeding up to this point, to this rate. If you see the 21st century as 100 years of change, it's an interesting century. If you see it exponentially, there's a much more profound century ahead.

GARDNER: Just out of curiosity and to switch the subject for a minute, is the stock market transaction software that you've created based in part on this notion of the speed of change and figuring out which kinds of companies are likely to crest and so on?

KURZWEIL: Not exactly that. We're analyzing data more in [terms of] looking for specific patterns. We feed in everything that we can get our hands on in the sense of quantifiable data. But we're not actually trying to make long-term opinions about companies. We're actually making short-term predictions about what will happen four hours from now, or a day from now. And it doesn't have to be perfectly accurate.

If you can predict whether a stock will go up or down tomorrow with, say, 65 percent accuracy where the chance is 50, you can definitely make money because it puts you in the position of being like the house in the casino, where the house can lose any bet, but over 50,000 bets, the probability is 99.999 that it will make a lot of money, because the odds are slightly in its favor. So if you can put the odds slightly in your favor, then if you place lots of bets you can make money. That's the theory.

GARDNER: You were asking me earlier about my theory of intelligence, and how confident I was that there were eight or nine kinds of intelligence.

KURZWEIL: Eight and a half at last count.

GARDNER: Right. And my gamble, my bet, my wager is that it's useful to think about that many different kinds of pattern recognition. Now, I'm well aware that once you get into any kind of a sensory system, you find many more fine-grained kinds of distinctions. So the real question is not, in some absolute way, whether eight or nine is the correct number, but how much mileage do you get for various purposes by thinking about it that way?

KURZWEIL: You put the theory out some time ago, over ten years ago. It's interesting. It's impressive how well it's held up. Because at the one extreme you have intelligence as one sort of model . . .

GARDNER: Ten years is becoming much longer, we're learning.

KURZWEIL: Yes, intelligence is one monolithic ability. And then you have, at the other extreme, the concept of hundreds of different specific skills, and for that matter you can break them down quite finely. The ability to play heavy metal music is probably somewhat different than playing a Chopin prelude. But there are many different specific skills that you can identify.

And you've postulated an intermediate organizational paradigm on the order of eight different ways in which these intelligences or skills aggregate into something we would call an intelligence, by your theory, despite the fact that there are many different skills.

Somebody who's a great musician may be a mediocre writer. Even within music, there are many different paradigms. Somebody could be good at jazz and not good at classical music, and so on, and even with jazz, there are dozens of different schools of thought. Yet the evidence does support the idea that there is some kind of resonance in each of these eight areas, which is interesting. It will be interesting to see the perspective we have as we learn more about the brain and how it's organized-the hundreds of different regions and what they do to information, the areas that deal with emotion, and what's really going on when somebody plays music, or deals with language, and so on.

GARDNER: What I'd like to do now, Ray, is talk a bit about the age of spiritual machines.

KURZWEIL: Didn't we cover that?

GARDNER: Well, what's gotten a lot of attention and a lot of people very excited are some of your specific predictions about 10 years and 30 years from now. (To audience) They are quite arresting, if you haven't heard about them before.

So why don't you take us forward ten years. Most of us can probably think about ten years. Thirty years is harder for most of us, and it's also much more extreme in terms of your predictions. But give us those movies.

KURZWEIL: Well, as everyone can see, computers are getting smaller and smaller, and they're going from notebooks down to palm tops. The display technology is getting higher and higher resolution. All of these technologies are growing exponentially.

And if we look at what will be feasible ten years from now, we won't carry displays. Images will be written directly to our retina, or eyeglasses and contact lenses. There are systems like this that you can try today. They're not quite high enough resolution. They're still too cumbersome and too expensive. But ten years from now, this will be quite ubiquitous.

So that could create a full-immersion visual environment that can either overlay real reality or replace it. We'll have extremely high bandwidth connection to the Internet at all times. Pretty high-level bandwidth, pretty high bandwidth wireless connection, is coming to Europe this year. And for political reasons we are two years behind them, but in about ten years we'll have extremely high bandwidth connection to the Internet.

The electronics for all of this will be so tiny, it will disappear. It will be in our clothing, it will be in our eyeglasses. And this will enable us, first of all, to have augmented visual reality at all times. As we walk around, we'll be online all of the time. If you look at a building, you'll immediately see little avatars floating in your visual field of view that will interact with you and tell you what you need to know about that building.

You'll be able to visit with other people in full-immersion, visual, auditory virtual reality. What we're doing now, or what we're all doing, is primarily visual and auditory communications. Just as an aside, the telephone was viewed as auditory virtual reality by observers in the 19th century. They thought it was amazing that you could actually be with someone else even though you were hundreds of miles apart, because it was the first time in human history that you could hold a conversation with somebody without being together.

But there is a lot that we communicate visually. There is a lot of the sense of presence we have by being able to use our bodies, even if we're not touching. So we will have full-immersion visual virtual reality, and it will feel like, or at least look like, we're together. There are crude prototypes of this called tele-immersion which you can try today, but this will be ubiquitous and routine by the end of the decade.

We will be interacting with our computers the way we interact with people, primarily through virtual personalities. So we'll look human. If you go to my website, the website of my company KurzweilAI.net, you can download Ramona, my female alter-ego. She's your hostess to the site. She's an example of a coming trend, which will be quite ubiquitous by the end of the decade: that you'll be interacting with lifelike, human-looking personalities by 2010. You won't mistake them for humans, but they will have reasonable facility with human language and will have very robust, narrow AI, which is artificial intelligence that can do a wide range of human tasks within limited domains, tasks that if a human does them, we would consider that to be intelligent behavior.

We have many examples today. If you get an electrocardiogram, the machine will generate a diagnosis and will actually write on the print-out where it finds arrhythmias. That's something that used to require a skilled physician. They play chess. They can understand human language within limited domains. This will be much more robust by 2010.

GARDNER: Let me interrupt for a second. Because one of the things that has been predicted for at least 50 or 100 years is that when technology reaches a certain level, there will be huge amounts of unemployment. And of course that's never happened. Let's stick with ten years from now. To what extent will the jobs that we have now already be anachronistic?

KURZWEIL: Well, there is constant turning of our jobs. Very early automation occurred in the English textile industry around 1800 in England, and that led to the Luddite movement, who were weavers who were thrown out of work. And it seemed quite obvious to them that if one person with these new machines could do the work of ten weavers or spinners or cloth makers . . . There were many different machines at different levels of the textile industry coming into existence through the early 19th century. It seemed apparent that fairly soon employment would be the domain of just an elite few.

What happened, though, was that whole new industries were created to create these machines. The demand for clothing went up. The common person could have well-made clothing for the first time. People didn't just want one shirt, they wanted a whole wardrobe. Fashion became an area of interest.

And if you look at the type of work that people do today, most of the job categories would be unrecognizable half a century, certainly one century ago. If you look at the statistics, we have 130 million jobs in the United States. We had one-tenth that number 120 years ago. The jobs pay six times as much in constant dollars. We spend ten times as much in constant dollars on elementary and secondary education. We have six million college students versus 50,000 in 1870.

The whole progression has been with constantly automating jobs at the bottom of the skill ladder, many different jobs, and creating new jobs at the top of the skill ladder. So the skill ladder is going up, and we've dealt with that by investing more in education. Those numbers are dramatic: from 50,000 to six million college students in a little over a century. So it requires more skill, and a lot of the investment has gone into education to keep up with the skill ladder. Ultimately, we're going to have to actually augment human intelligence. Not by 2010, but if we're talking 2030, 2040, then we're actually going to have to increase human intelligence to keep up with that paradigm. But so far, the impact on employment has been a positive one. I mean, more people are gratified by their jobs today. Maybe not everybody, but if you go back 200 years, most human employment was very difficult labor, and not something [from which] people defined themselves, or drew their identity from, or found particularly creative or gratifying.

The kind of scenarios that become feasible in the year 2030 are the following. We will be able to send, and we will routinely send, little nanobots into our bloodstream. There are in fact applications on the drawing board today to do that for medical purposes. But we can send billions of these little microscopic robots the size of blood cells into our bloodstream, and they will be able to do a number of things.

First of all, they could scan the brain from inside. We have scanning technology today. If you put the scanning tip right near the neural features, it can scan those neural features with very high resolution. So you could scan my brain today and actually see the interneuronal connections, and in some cases the neuro-transmitter concentrations, if you move the scanning tip right next to all of the neural features.

If you want to do that without making a mess of things, you want to send the scanners inside the brain. The capillaries travel by every neural feature, because every neural feature requires blood to provide nutrition. So, it will be able to actually scan from inside with very high resolution and create these extremely high-resolution maps of the brain for this brain reverse-engineering project, which will provide us the designs, one set of knowledge from which we can build intelligent machines.

Another technology that has been demonstrated today is called a neuron transistor, which is a piece of electronics which, if it's right near a neuron, can actually communicate wirelessly with a biological neuron in both directions. If the neuron fires, it detects that electromagnetic pulse. And it also, conversely, has the ability to suppress a neuron from firing, or cause it to fire, which is communication in the other direction. Scientists have actually controlled the movement of animals, for example, using this type of technology.

So we send billions of nanobots into the human brain. They are all communicating with each other. They are on a wireless local area network, which is technology we understand today; not at the scale of billions of these elements, but we're talking 2030.

They are, of course, all on the Internet. And they can communicate wirelessly with our biological neurons through a modern version of the neuron transistor I just mentioned.

So let's take the virtual reality scenario. If I want to be in real reality, the nanobots sit there and do nothing. If I want to go into virtual reality, the nanobots suppress the signals coming from my real senses and replace them with the signals you would be receiving.

GARDNER: But you are preserving ego? [This] is what I was asking.

KURZWEIL: Yes. My brain will be getting signals from the virtual environment, and it thinks it's in that environment because it's getting signals as if they were coming from my senses, but the nanobots are [actually] providing them. They are suppressing the signals from my real senses and replacing them with the signals I would be receiving if I were in the virtual environment. So then it feels like I'm in the virtual environment, and I can be an actor in that environment. I can go to move my hand, but it suppresses my real hand from moving and instead moves my virtual hand. So my brain-I-think I'm in that virtual environment.

But you and I can go there together. Two people can go there. A whole group of people can go there. You can meet other people in these virtual environments. And we'll be able to do this in 2010 without the nanobots, just with full-immersion virtual reality from such intimate things as our eyeglasses and contact lenses.

But in 2030 it's going inside the brain, inside the nervous system. It can be much more realistic, much more compelling, and involve all of the senses. So you can have any kind of experience with anyone, from sexual and sensual experiences to business negotiations, or dialogues like this, in virtual environments.

And going to a website will mean entering a virtual reality environment. There will be recreations of earthly environments like this hall, or [you'll be able to] take a walk through a virtual Mozambique game preserve, or fantastic, imaginary environments that have no correlate in the real world.

Experienced beamers will beam their entire flow of sensory experience and neurological correlates of their emotions, or secondary reactions to the sensory experiences like sexual pleasure, onto the Web the way people now beam their images from their Web cams. And you can plug in and experience what it's like to be someone else, all of the thought-the concept of Being John Malkovich. Or you can relive archived experiences.

I have a friend who is deaf, but I can hold a telephone conversation with him because of his neural implant. There are implants for Parkinson's patients. We're beginning to use neural implants, particularly for disabilities and conditions caused by disease, but they are invasive. They require surgery to implant.

The nanobots can go inside the brain non-invasively and augment both our experiences through this kind of virtual reality technology, as well as ultimately augment human intelligence itself. The bulk of our thinking takes place in the interneuronal connections. And as I mentioned, we've actually modeled that in some regions of the brain. And we're limited to a mere hundred trillion connections. And I don't know, I find that pretty limited.

GARDNER: I saw some reflexive hands going up.

KURZWEIL: Just one more point. We'll be able to add new connections, and ultimately have 200 trillion, or 100 trillion times a thousand, or times a million, and explain human intelligence.

GARDNER: Since you've been cast as the optimist vis-à-vis Bill Joy, who I think would make some of the same predictions as you, do you worry about distopias, about things going wrong?

KURZWEIL: I think my writings about the future are neither utopian nor distopian. I think Joy sometimes bridges on distopian visions, and also his solutions are somewhat totalitarian. He says, "I'm not anti-technology. Technology does a lot of good things. Let's keep those good applications, but those potentially destructive technologies, you know, the ones like nano-technology, let's just not do those."

In my view that's unrealistic, because those potentially disruptive technologies are the same ones that are helpful. I mean, the same biotechnology that we're going to develop and are in the process of developing that will save millions of lives from cancer and disease are the same tools that a terrorist could [use to] create a bioengineered pathogen, and it's the same for these other technologies. It's the same knowledge. It empowers both our creative and destructive impulses.

If you read some of what Joy writes, the potential applications in some of these technologies sound scary, and they are. But he is sort of presenting them as if the destructive technologies of 2020 were suddenly presented to us today in a world that's completely unprepared to deal with them. And that's not how we're going to get there.

We're going to get there with our ability to protect ourselves also growing more and more empowered, and dealing with the dangers and the potential benefits as these technologies evolve. I think something that can give us more comfort is how we've dealt with computer viruses, a new form of pathogen. They're self-replicating, and all of these dangers have to do with self-replication. They self-replicate in a particular medium, which is computer networks, and they are human-made or initiated.

When these first emerged, observers said, "The first computer viruses are primitive. But when they get more sophisticated, they're going to completely shut down and destroy computer networks." Indeed, they have gotten more and more sophisticated over time. But we have relegated them to a relative nuisance level. They have, and they do, from time to time cause millions of dollars of damage. But the damage is a fraction of a percent of the benefit we get from computer networks. Someone could say, in fact Bill Joy did say this in a dialogue, "But software viruses, Ray, don't usually kill people." And I said, "Well, if they get into an intensive care unit software, they might." Largely that's true, but I think that really only strengthens my observation. Because the fact that software viruses are not generally lethal today means that our response is fairly lackadaisical. Lots of people put them out because they think they're practical jokes. If they thought they were killing people, they wouldn't do that.

We're not that diligent about using our anti-viral software. Law enforcement kind of has a reaction, but it's not that intense. When we get some of these future technologies that are potentially lethal, the law enforcement response, the ethical guidelines, the attempt at technological safeguards and technological immune systems, will be 100 times more intense. So the fact that with a fairly lackadaisical effort we've kept this new form of self-replicating, human-made pathogen to a nuisance level is encouraging.

GARDNER: Of course, there are assumptions about human good and evil, and how they change, which takes us in a totally different sphere.

The last official question. You are at a school of education. Many people here will either themselves be teaching kids in 10 or 15 years or will be training people who are going to be teaching kids in 10 or 15 years. To what extent should we educators be thinking differently about what we're doing? We talked a bit yesterday about what knowledge is and how it's being affected by what's going on. So if you could give us a few thoughts about that, that would be a great tone in which to conclude.

KURZWEIL: Well, if I think about my own childhood, I learned the most when I was doing something that I felt passionately about. So I would encourage kids to follow their passion, and to be experimental, to actually do their own projects. I would teach kids it's okay to fail. It's okay for adults to fail, too. Edison tried hundreds of different filaments before he found one that worked. And we don't remember those hundreds of failures; we do remember the one experiment that came to fruition. So I'd like to say that failure is just success deferred.

But I think it's an important lesson to teach kids: to be experimental, to try things out. And if it's possible for a kid to actually find something that he or she has a passion about, whether it's in a particular area like music, or technology, or literature . . . you follow those passions, and then you learn, as a side-product of trying to achieve some vision. And I think kids can do that.

As technology explodes and has more and more diverse capabilities, as knowledge and science become more and more diverse, there are more and more ways of putting this knowledge together. So I would encourage kids to be experimental, but to provide some discipline and have them define what they're trying to do. But the idea of having a rote education world where kids just learn specific areas of knowledge and specific skill sets, [as in the] emphasis towards testing . . . I mean, the extent to which we do allow kids to really try out the world of knowledge and use it in creative ways at young ages, we're much better off than trying to have every kid meet predefined skill sets. Though I will agree that there are certain bases for an educated society in terms of reading ability and so on, and we should establish some basic levels of capability.

GARDNER: Your education vision is a nice blend of very traditional values about motivation and passion, but played out in the field of expanding opportunities and areas of knowledge, very much like what you've said about the occupational spectrum of the future.

I'm sure everybody who was here this evening, who remembers this discussion for ten years, or for 30 years . . . maybe we'll come back and we'll do a little checklist. Thank you very much, Ray.