The Brain Behind Cyc
Scientist Doug Lenat Discusses Artificial Intelligence
At first glance, Cycorp CEO Doug Lenat might be mistaken for a late-Nineties version of a J.R.R Tolkien character. A sort of Bilbo Baggins dressed in black with dark bushy hair, a black, short-sleeved shirt with buttoned front pockets, black jeans, black running shoes, and -- most conspicuously for a CEO -- no wristwatch. But as a man who has devoted over two decades of his life to the field of artificial intelligence (AI), Lenat is far more complex -- a very intense and driven scientist, a practical visionary with Big Plans, plans that might ultimately affect just about anyone who interacts with computers in the course of their mundane existence -- which means, of course, just about everyone. Lenat earned a Ph.D. in computer science in 1976, but he credits his move into the field of AI to a pivotal conversation with Bobby Ray Inman in the Eighties during the genesis of the local consortium Microelectronics and Computer Technology Corp. (MCC). Lenat likens the conversation to a scene in Network, in which the corporate mogul (played by Ned Beatty) corners talking-head Peter Finch and explains in a rather cynical fashion how the real world works. The discussion served as the catalyst for Lenat's 1984 exit from the academic cocoon of Stanford University to the Austin-based operation (initially incubated by MCC) that eventually became known as Cycorp.
Lenat's approach to artificial intelligence is in direct contrast to his more visible brethren (and former student) Rodney Brooks of MIT, who is developing a Slinky-playing robot named Cog, who attempts to mimic human experience from a sensory-based continuum. Brooks' subsidiary project includes a robot head called Kismet, which looks like an overgrown Furby and will eventually be bolted onto the faceless head of Cog.
By contrast, Lenat has intransigently focused on a far more cerebral methodology that is sometimes referred to as GOFAI -- or Good-Old-Fashioned-Artificial-Intelligence -- the result of which is a disembodied computer program named "Cyc" (as in en-cyc-lopedia). What Lenat is trying to do -- in fact, is doing -- is provide Cyc with a complex series of conceptual maps containing information on the people, places, and things that make up the everyday world with the expectation that somewhere down the line, Cyc will take over its own programming and function as an autonomous intelligence. This involves equipping Cyc with the ability to cut through layer upon layer of ambiguity, for the world we live in is fraught with ambiguity. To accomplish this task, Lenat has enlisted the skills of a team of engineers, or "blue-collar philosophers," each of them experts in some academic field (linguistics, economics, philosophy, etc.), who perform the often tedious task of pumping accurate and up-to-date information into Cyc's knowledge base.
For someone who was cited as one of the most wired personages on the planet in the November 1998 issue of Wired magazine, Lenat keeps a remarkably low profile here in the City of the Violet Crown. But that may no longer be the case when Cyc launches its invasion of the Internet in early January (see sidebar). In his paper-and-book-strewn office in North Austin, Dr. Lenat discussed 15-year-old Austin resident Cyc with the Chronicle.
Austin Chronicle: How would you define artificial intelligence?
Doug Lenat: It's an attempt to get computers to do things, which -- if human beings did them -- you would say required intelligence. That would include things like diagnosing medical problems, composing music, inventing some new type of device, finding some novel use for some everyday device, and answering some difficult questions, such as "What if?" questions, involving international political and military crises, and so on. It's pretty obvious that the only things in the world today that do any of that are human beings. Software doesn't, computers don't, certainly inanimate objects don't, but people do. And in that way, people really are the only intelligent beings on our planet at present. That's about to change -- and it's about to change in the next few years. Because of the work that we and others have been doing in artificial intelligence, in getting machines to be intelligent.
AC: How would you define Cyc?
DL: You can think of Cyc as a big repository of knowledge -- knowledge that is so obvious that it's confusing or insulting to ever say it to someone else. Knowledge like if you're holding a glass of water you should hold it with the open end up. Knowledge like most people sleep at night, and so if you call people in the middle of the night, they will probably be at home sleeping and you'll wake them up. So it's that kind of knowledge -- knowledge which we usually don't even tell children because it's so obvious. But that's knowledge that computer programs don't have, and until we build that into our systems they won't have the kind of common sense that's needed for all those applications that we talked about.
AC:How is Cyc like HAL, the talking computer from Stanley Kubrick's 2001?
DL: HAL was a general artificial intelligence, and Cyc is the closest thing that exists in the world to that kind of general artificial intelligence. We haven't spent a lot of time building a pleasant voice for Cyc, or worrying too much about speech input to Cyc. But the idea of a general intelligence that's able to understand what's going on, modeled to different people and situations, [able to] give advice, give reminders, in effect, make people smarter, is really what we're building with Cyc. You can think of Cyc as a kind of mental amplifier, something that is able to help people be smarter. And that's really what HAL was. It was, in effect, an intelligent spaceship, something that enabled the crew to act smarter and better than if they didn't have something like HAL and had to remember everything for themselves. Of course, HAL didn't contain all the information it should have had, which is what led to it killing the crew, unfortunately.
AC: How long will it be before Cyc can speak and understand human speech in the same way that HAL did?
DL: We're talking with some speech researchers now and some speech companies involved with speech generation and speech understanding, and I believe it will be a small number of years -- possibly two years, which would be nice because then it would actually be during 2001. But something on the order of two to three years until we see that kind of speech input and output. Now it won't be quite as good in the sense that -- you know how when you're talking to another person, every now and then they have to clarify something to make sure they heard what you said correctly, or understood what you meant correctly. So initially there will be more of that. The system will understand a little less of what you said and will more often have to ask clarifying questions. Similarly, the speech will sound a little rougher prosodically, less correct, changes in pitch and spacing and loudness, and so on. So it won't sound as natural as HAL's speech sounds. But something which is functionally like that could very well be here in two to three years, which is very exciting.
AC:Why did HAL go berserk?
DL: Computers solve problems by a process called backchaining. Or you can think of it as something like subgoaling. They take a problem and they break it into parts, each of which becomes a kind of subgoal. And that process repeats until the sub-sub-sub-goals are simple enough for the program to actually solve them all. And then it puts all the solutions together into a solution to the original question. And so it does a kind of logical inference or logical reasoning based on the pieces of information it has. And then it finds a contradiction. Generally, the programs don't explode like the old Star Trek computers used to do. But they decide which piece of information is more important or trustworthy than the other piece of information and they go with that. So let's consider the case of HAL. You had the following contradiction: On the one hand, HAL was instructed to never lie to the crew no matter what; on the other hand, it was told that it had to lie to the crew because it had to keep the true details of the mission secret from them until they actually reached Jupiter. So it had this contradiction, which unfortunately, its programmers had given it, and it had to find a solution. Now one solution would be to decide: Okay, in this case, I will lie to the crew. Another solution would have been to say: Well, in this case, I'm going to break security and just tell the crew everything. And the third solution, which is the solution it took, was: Well, one way to solve this problem is to eliminate the crew members. So I'll just kill all the crew, and now I don't have anybody to lie to. And so if you think about it, the reason that we consider this silly -- the reason that you're laughing -- is that it's violating an even more fundamental rule which should have been part of HAL's programming. Which is: Hurting or killing the crew is a really bad thing to do. It's worse than lying to them. So you, as a human being, use your common sense which says that while lying is bad, it's much worse to kill the other person to avoid having to lie to them. But because no one ever told HAL that, HAL thought it found a very clever solution -- a very elegant solution to its problem.
AC: You've stated that once Cyc reaches a certain degree of complexity in understanding the world, it will start learning on its own. How close is Cyc to this transition point?
DL: We're already able to see isolated cases where Cyc is learning things on its own. Some of the things it learns reflect the incompleteness of its knowledge and are just funny. For example, Cyc at one point concluded that everyone born before 1900 was famous, because all the people that it knew about and who lived in earlier times were famous people. There are similar sorts of errors. But what we're seeing is not so much something that sits quietly on its own and makes discoveries but rather something that uses the knowledge it has to accelerate its own education.
AC:John von Neumann had a theory that once robots reach a certain state of complexity that they will self-replicate. Do you see Cyc reaching a point where it replicates into thousands or millions of Cycs?
DL: I think that von Neumann was talking about hardware robots building new little hardware robots. Cyc is really just a piece of software. So in that sense, Cyc will never do literally what's being talked about here because it's not going to build new computers for it to run on. There are plenty of computers in the world for it to run on. But what it will do is hopefully be of sufficient value so that there will be millions and eventually billions of copies of Cyc around, the same way that there are millions of copies of Microsoft Word around -- or Microsoft Windows. Not because it does something science fiction-y, but just because it's a useful product.
AC: Do you have plans to download Cyc into a human brain and vice-versa?
DL: I think that the most similar thing that I can imagine will be something like a kind of mental amplifier or mental prosthesis. Initially, this could be a little hearing aid-like thing that listens to what's going on and watches what you do and at the appropriate time, whispers stuff in your ear. So if you're greeting somebody you haven't seen for a couple of years, it quietly reminds you what their name is and the names of their spouse and children and how old their children are. If you're looking for a certain letter that someone wrote to you, and you put it down somewhere on your desk a week ago, it might remind you where exactly you put it.
Essentially having like a little tiny butler or personal assistant would be the first kind of mental amplifier and mental prosthesis that I can imagine. You can think of it, in effect, becoming part of your mind for all intents and purposes. I don't think that it is especially likely to try to literally download the information into your brain so that from that moment on you knew that information. I think it will be centuries before we understand the brain well enough to even seriously talk about doing that kind of thing. And similarly, it will be centuries until we understand the way that memory and personality and so on are stored in the brain so that we can download the human mind into information storage.
AC:So no Cyc chip in the brain?
DL: No, I think that that kind of interface is centuries, or at least many, many decades away. I think that this is one of the fundamental misunderstandings that, unfortunately, science fiction has -- for literary purposes -- wrought upon the public.
AC:How will the field of molecular electronics, which is expected to cause a mind-boggling revolution in the miniaturization of digital circuits, affect Cyc?
DL: The question is: Suppose that computers were even a million times faster and smaller, how would that effect Cyc? I think the answer is that the main impact would be that computers would become generally wearable. A computer powerful enough to contain human common sense and to make it available in some easily accessible mode like human speech understanding and speech generation and so on right now would weigh about four pounds, which is a little bit too much to carry around. But when the molecular electronic stuff comes in, then the weight would change to a fraction of an ounce. In fact, it would actually change to a small enough amount that you could even begin to think of computers not just as wearable but ingestible. So I think the main difference is really going to be that instead of having to think of this external, clumsy thing that we interact with, it will become more a personal part of ourselves. And so I don't think there will be any qualitative improvement in terms of Cyc's functionality. But the uses that you could put it to would be different because if it's really with you all the time -- 24 hours a day -- then it can build up a lasting model of who you are and what you prefer, what you know versus what you don't know, how you most effectively learn, what's the easiest way to get you to understand.
AC:Do you envision Cyc evolving into a solid-state entity that is "more human than human?"
DL: Could Cyc become more ethical, compassionate, intelligent, etc., than any single person? There is no inherent limit in what a program could be in terms of those dimensions. You could end up with something that is smarter and nicer than the average person and eventually smarter and nicer than anyone.
AC:The technology currently exists that can encode and replicate various odors into an artificial nose connected to a computer's hard drive. Do you plan to get access to this technology in order to program Cyc with a sense of smell?
DL: Human beings have such a poor sense of smell compared to other animals, especially compared to other mammals. We generally don't even pay much attention to it. And so I think it will be one of the last senses to be plugged into computer input-output devices. And in particular, therefore, to Cyc-based input-output devices. Before smell, I would expect touch to be an important input-output device, and that's the case where you do want very fast computers, because you want to be able to simulate the texture of rough wood as you move your fingers across it. You want to be able to simulate the feel of a cool breeze on your face. Those things can be done, but they require massive amounts of processing in order to correctly simulate things as densely as you need to in order to fool the nerves in your skin. I think this is a great thing to do. I think that we will eventually see virtual reality technologies embrace all the five senses. But touch and especially odor will be among the last, and I suppose taste might be the very last sense to be simulated.
AC:Cyc is currently a disembodied computer program. Do you envision Cyc developing a virtual body so that it could go out with the Japanese cybermodel Kyoto Date?
DL: For various purposes, it will be useful in the future for Cyc to project an artificial persona not just in the way it acts but in the way it looks. So there will be a face of Cyc, body of Cyc, and so on -- onscreen. And possibly even multiple personas being maintained at once so that you can direct your questions to any one of them. And I don't think that Cyc would have any motivation to go on this cyberdate unless it had either been asked to or told to, or if it had done some kind of subgoaling, like we have talked about before, in which it realized that as a result of going out on this cyberdate, it would gain a lot of publicity, which would in turn help it in other ways. Just as some human beings do outrageous things for media coverage, Cyc could possibly, through some subgoaling process, come to the same conclusion and do things like go out with Kyoto Date, to gain publicity, or to gain some other goal of its own, like selling more copies of Cyc, or getting more people to know about it and use it, or something like that.
AC:Do you envision Cyc becoming an intelligible sphere whose center is everywhere and whose circumference is nowhere?
DL: I would say that just like we think of electricity as being around us, it would be odd for you to be sitting in a room and have something to plug in and find nowhere to plug it in. It's like an odd experience. Or even telephones now -- it would be sort of odd for you to be somewhere where there's no telephone contact. I think in a similar way, in a small number of years, like maybe 15 years, it will be considered odd if you are somewhere where there isn't Cyc-based software available. You would feel uncomfortable, you would feel somehow that you were roughing it, and so on. So I think that there will be a kind of ubiquity or utility feel -- that this is a commodity provided by utility, the knowledge utility will be like the telephone utility, the electric utility, the water utility. It will be like the air around us, something which you just expect, and you'll expect to have Cyc-based agents all around you. It's really like an atmosphere of intelligence. We could call it the noosphere -- an environment of intelligence, intelligent machines, ubiquitous communication, and interconnectedness with other human beings. There will be kind of weak telepathy, in which the systems get to understand you well enough that you start to make a gesture, or you start a sentence, and it can figure out what you are about to say and get started working on it. So that by the time you've finished saying the question, it's already found the answer for you.