The starting point, of course, is the digital computer. Everyone knows that computers don't think, they just perform a bunch of chores that used to require thinking. But thinking is as thinking does, and once you've got a machine like that, it makes you stop and think -- er, that is, it prompts you to inquire: exactly what is the distinction between what we used to call thinking and what this machine does? Is the distinction all that great? Might the machine shed any light on how the human brain works? Might we improve the machine to "think" for real?
The solder was scarcely set on the first few electronic computers when A. M. Turing, in 1950, suggested a famous test for an "intelligent" machine. You put a man (we'll call him Alfred, the Answerer) in one room and a computer in another. Another man (Quincy, the Questioner) sits in yet another room, where he can see neither of them, and asks questions by teletype. He can address either room, but isn't told which is which. Now, the test is this: the computer must pretend it's Alfred, while Alfred tries to scotch this lie and Quincy does his best to get at the truth.
Since Quincy can't see the two contestants, and can communicate only by teletype, the only clues he gets are the answers to his questions. He has to devise some line of questioning that will make the computer give itself away, or enable Alfred to distinguish himself. Presumably, if there is any difference that matters between the "thought" processes of the machine and those of a human being, Quincy should be able to put a question that takes advantage of it.
Turing didn't say how often the machine would have to win in order to pass the test, or against what class of competition, but you get the point: if Quincy and Alfred are both reasonably clever, and if the computer can fool Quincy at least as often as Alfred can put him right, then it's a pretty sharp cookie, and we might be justified in calling it an "artificial intelligence".
We might, but personally I would propose more stringent requirements. Turing himself livened up the discussion, when he set forth this proposal, by first describing a similar game between a man and a woman. Both the man and the woman would claim to be the woman; the interrogator was to figure our questions that would flush out the man or help the woman to make herself known. Well, I ask you, now: suppose the man were to win fifty percent of the time or better, would that make him an "artificial woman"?
Anyone shopping for an artificial woman is likely to demand a couple of features besides feminine reactions to questions. In fact, he is likely to make a sharp and heated distinction between an "artificial woman" and a "female impersonator."
I submit that, in like vein, an "artificial intelligence" should have some additional abilities besides answering questions. In particular, it should be able to manipulate a robot and get it about in the world, subject to imperatives such as keeping the robot fueled and constraints such as never wrecking any vital parts. Above all, it should show awareness and comprehension of what it is doing -- to the satisfaction of reasonable critics. (But try demonstrating your awareness and comprehension to somebody who is really intent on disbelieving.) Any lesser measure of competence invites the epithet, "intelligence impersonator," not "artificial intelligence."
Be that as it may, Turing made the right point: to say it again, intelligence is as intelligence does. Or, if you like, intelligence is in the mind of the beholder. If a machine dressed up like a human being can fool you indefinitely, then it is passing as much of a test as you require of your friends. It is a ticklish question indeed whether the "intelligence" that your attribute to yourself and your friends includes anything over and above the programmed bag of tricks that would enable a machine to carry on the same way.
It wasn't long before a horde of mathematicians and programmers were off and running with programs to imitate intelligence. They may or may not have been inspired by Turing; they certainly didn't need to be. After all, programmers like to fool around, and making a computer play chess is even more glorious than making it draw a nude on the printer or spell "MERRY CHRISTMAS" in the console lights.
Still, there is foolery and there is foolery. From the start, there was widespread awareness that some recreational uses of computers had far-reaching implications. Ere long, some courageous souls dared name their study by its staggering ambition (much as this raised academic eyebrows, and still does). Thus was born the research field of "artificial intelligence".
An account of some early classics in the field may be found in Computers and Thought, edited by Edward Feigenbaum and Julius Feldman (McGraw-Hill, 1963). Several principal types of effort were evident by 1961: there were programs that played games, programs that answered questions, programs that proved theorems, programs that modeled supposed human thought processes, and programs that recognized patterns. If you look up any current bulletin on artificial intelligence, you'll see that these are still among the dominant topics. (See, for example, the ACM SIGART Newsletter. ACM = Association for Computing Machinery, SIGART = Special Interest Group on Artificial Intelligence.)
This is no place for details. There are details enough to fill up your attic, and that would be just the place for most of them. What is of consequence is not how such-and-such a chess-playing program provided for queen sacrifices, but what sorts of problems are people working on, and whither does their labor tend. I could contribute nothing to a technical discussion that would be any news to computer professionals, and much that would put other people off. Meanwhile, the important offering would be missed: context. Direction. The sweep of history.
At the same time, this is no place for evasions. It is fair to pursue the matter further than, "What are people working on?" After fifteen years, where have they gotten with it? Lots of mobs go rushing off on lots of enterprises whose results would be one thing if they succeeded, but are another in the event. How has this one fared?
It depends what you are looking for. If you're going to demand the thing that was promised -- artificial intelligence --, then the handiwork of artificial intelligence enthusiasts has been disappointingly short of the mark. (A critic would say it's been a gigantic put-on.) There is now a five-foot shelf of literature about chess programs, mechanical proof procedures, and yes, even "shrink" simulators. You can read this stuff until your own pattern-recognizing ability is impaired by failing vision, and still you're left with a nagging question: where is the "intelligence" in any of these artifices?
Programmers are ingenious people, even after they become established and call themselves mathematicians. It is not their intelligence which is in doubt. The question is whether they have imparted any of it to the computer, in any more fundamental sense than Leonard Bernstein imparts musical mastery to a long-playing disk.
Take the chess-playing programs. The earliest examples lost to duffers or to amateurs. Now there are programs which lose to rated players. Tomorrow there will be programs which beat rated players, just as A. L. Samuel's checker-playing program long ago beat a checkers champion. But suppose one does. Will we then say that the computer is a chess master?
Or will we say, rather, that one-upmanship is getting fancier all the time? Whereas it used to be the last word to play without a board, now there are players who don't even stick around for the game. They just leave a list of instructions, telling you at every step what they would do next. The computer has nothing to do with it, it just decodes the instructions (which are sort of complicated and long-winded).
When matters are presented in this light, the answer seems obvious. And yet -- there is more to it than meets the eye. The instructions left by the absentee chess master can't be just any old set of instructions. They have to be in a tightly encoded form that unfolds in response to the opponent's play. With just a handful of instructions, comparatively speaking, the programmer must cope with board situations that would take until Doomsday just to count. He has to provide for configurations that he himself has never encountered, never will encounter, and hasn't given any specific thought to.
To call this a playback of the player's intentions is a true statement, but a misleading one. It glosses over the complexity of the encoding. It distracts attention from what may be the crucial thought: that enough versatility wrapped into few enough instructions may be just what constitutes "intelligence."
A chess-playing program, a question-answering program, or a theorem-proving program is a far cry from the marvel that is man -- so much so, that to present it as a study in "artificial intelligence" is to invite derision. At the same time, any of these is an order of magnitude more complex than, say, a payroll program. That is why such projects were chosen. Nobody ever did suppose that if a computer could do one or several of these things, that alone would qualify it as "intelligent" (unless, perhaps, the question-answering program could answer everything that a man could answer -- but that is no description of the programs we're talking about). Rather, the scheme was to extend the outer limits of programming this much, confidently expecting that they could then be extended some more, until the versatility of man himself was matched.
I think a sober judgment is the following. Artificial intelligence research to date has demonstrated what it didn't want to demonstrate: that we won't program our way to intelligence on the basis of present equipment and present encoding techniques. On that basis, a good chess program is still a taxing problem, let alone a good robot. On the other hand, nothing has happened to disprove the view that intelligence is indeed just a super-sophisticated program, representing information in ways we don't yet understand, in biochemical media we have barely begun to identify -- but a program for all that. To the contrary, discoveries in biology have been more and more confirming this surmise. When we hit upon computers and computer programming, we brought into focus for the first time the right questions to ask about the human mind.
In an era that looks forward to Bigger and Better as a matter of course, it is not too disappointing that we'll need great gobs more storage, and/or far cleverer representation of information. All that this means is that the schedule has slipped. What is perhaps more unsettling is that nobody can yet propose, even in principle, how to tie the pieces together.
Suppose that a variety of game-players, theorem-provers, question-answerer, pattern-recognizers, and whatnot have been put in working order. How do we integrate them into an all-round intelligence? How do we provide a sense of drive, of feeling, of purpose and emotion? Above all, how do we program self-awareness?
We'll come back to this last clump of questions in a later section. First we must digress from self-styled "artificial intelligence" efforts and look at some related things that are going on. What gives this dream its power is that all sorts of people, carrying on all sorts of activities for practical reasons that have nothing to do with building a robot or transplanting a man, are busily contributing to the result. Half a dozen scientific mainstreams are converging toward one grand concourse whose name is immortality.
|
|
Angels Unaware: The Business Machine as Engine of History. |
| Work in Progress: The Gathering Rainbow. |
 |
 |