Broca's Brain (40 page)

No device of this sophistication has yet been built, although I am not sure how many humans would pass Turing’s human test. But the amount of effort and money put into artificial intelligence has been quite limited, and there are only about a half-dozen major centers of such activity in the world. One of the more striking results obtained in a very limited universe of discourse—that of children’s blocks—has come from the work of Terry Winograd, then at the Massachusetts Institute of Technology. Here is a dialogue between man and machine, the machine having been programmed by Winograd:

The machine does. It reaches out its mechanical hand, moves the cubes and builds the structure that has just been described.

As another example, a machine psychiatrist has been developed by Joseph Weizenbaum, also at MIT. This is a much simpler program than Winograd’s, since it involves nondirective psychotherapy, which, generally speaking, requires extremely simple-minded computer programs. Here is a striking dialogue between a human being and a computer psychotherapist (There is certainly a selection effect here. Many other such machine/human psychiatric dialogues are not nearly so striking. But I suppose most human/human psychiatric dialogues are not very striking either.) In this interaction, in order to spell out clearly the respective roles of therapist and patient, the computer therapist types upper case while the patient types lower case:

This astonishing—one is very tempted to say “perceptive”—response from the computer is, of course, preprogrammed. But, then, so are the responses of human psychotherapists. In a time when more and more people in our society seem to be in need of psychiatric counseling, and when time-sharing of computers is widespread, I can even imagine the development of a network of computer psychotherapeutic terminals, something like arrays of large telephone booths, in which, for a few dollars a session, we are able to talk to an attentive, tested and largely nondirective psychotherapist. Ensuring the confidentiality of the psychiatric dialogue is one of several important steps still to be worked out.

ANOTHER SIGN
of the intellectual accomplishments of machines is in games. Even exceptionally simple computers—those that can be wired by a bright ten-year-old—can be programmed to play perfect tic-tac-toe.
Some computers can play world-class checkers. Chess is of course a much more complicated game than tic-tac-toe or checkers. Here programming a machine to win is more difficult, and novel strategies have been used, including several rather successful attempts to have a computer learn from its own experience in playing previous chess games. Computers can learn, for example, empirically the rule that it is better in the beginning game to control the center of the chessboard than the periphery. The ten best chess players in the world still have nothing to fear from any present computer. But the situation is changing. Recently a computer for the first time did well enough to enter the Minnesota State Chess Open. This may be the first time that a non-human has entered a major sporting event on the planet Earth (and I cannot help but wonder if robot golfers and designated hitters may be attempted sometime in the next decade, to say nothing of dolphins in free-style competition). The computer did not win the Chess Open, but this is the first time one has done well enough to enter such a competition. Chess-playing computers are improving extremely rapidly.

I have heard machines demeaned (often with a just audible sigh of relief) for the fact that chess is an area where human beings are still superior. This reminds me very much of the old joke in which a stranger remarks with wonder on the accomplishments of a checker-playing dog. The dog’s owner replies, “Oh, it’s not all that remarkable. He loses two games out of three.” A machine that plays chess in the middle range of human expertise is a very capable machine; even if there are thousands of better human chess players, there are millions who are worse. To play chess requires strategy, foresight, analytical powers, and the ability to cross-correlate large numbers of variables and to learn from experience. These are excellent qualities in those whose job it is to discover and explore, as well as those who watch the baby and walk the dog.

With this as a more or less representative set of examples of the state of development of machine intelligence, I think it is clear that a major effort over the
next decade could produce much more sophisticated examples. This is also the opinion of most of the workers in machine intelligence.

In thinking about this next generation of machine intelligence, it is important to distinguish between self-controlled and remotely controlled robots. A self-controlled robot has its intelligence within it; a remotely controlled robot has its intelligence at some other place, and its successful operation depends upon close communication between its central computer and itself. There are, of course, intermediate cases where the machine may be partly self-activated and partly remotely controlled. It is this mix of remote and
in situ
control that seems to offer the highest efficiency for the near future.

For example, we can imagine a machine designed for the mining of the ocean floor. There are enormous quantities of manganese nodules littering the abyssal depths. They were once thought to have been produced by meteorite infall on Earth, but are now believed to be formed occasionally in vast manganese fountains produced by the internal tectonic activity of the Earth. Many other scarce and industrially valuable minerals are likewise to be found on the deep ocean bottom. We have the capability today to design devices that systematically swim over or crawl upon the ocean floor; that are able to perform spectrometric and other chemical examinations of the surface material; that can automatically radio back to ship or land all findings; and that can mark the locales of especially valuable deposits—for example, by low-frequency radio-homing devices. The radio beacon will then direct great mining machines to the appropriate locales. The present state of the art in deep-sea submersibles and in spacecraft environmental sensors is clearly compatible with the development of such devices. Similar remarks can be made for off-shore oil drilling, for coal and other subterranean mineral mining, and so on. The likely economic returns from such devices would pay not only for their development, but for the entire space program many times over.

When the machines are faced with particularly difficult situations, they can be programmed to recognize that the situations are beyond their abilities and to inquire of human operators—working in safe and pleasant environments—what to do next. The examples just given are of devices that are largely self-controlled. The reverse also is possible, and a great deal of very preliminary work along these lines has been performed in the remote handling of highly radioactive materials in laboratories of the U.S. Department of Energy. Here I imagine a human being who is connected by radio link with a mobile machine. The operator is in Manila, say; the machine in the Mindanao Deep. The operator is attached to an array of electronic relays, which transmits and amplifies his movements to the machine and which can, conversely, carry what the machine finds back to his senses. So when the operator turns his head to the left, the television cameras on the machine turn left, and the operator sees on a great hemispherical television screen around him the scene the machine’s searchlights and cameras have revealed. When the operator in Manila takes a few strides forward in his wired suit, the machine in the abyssal depths ambles a few feet forward. When the operator reaches out his hand, the mechanical arm of the machine likewise extends itself; and the precision of the man/machine interaction is such that precise manipulation of material at the ocean bottom by the machine’s fingers is possible. With such devices, human beings can enter environments otherwise closed to them forever.

In the exploration of Mars, unmanned vehicles have already soft-landed, and only a little further in the future they will roam about the surface of the Red Planet, as some now do on the Moon. We are not ready for a manned mission to Mars. Some of us are concerned about such missions because of the dangers of carrying terrestrial microbes to Mars, and Martian microbes, if they exist, to Earth, but also because of their enormous expense. The Viking landers deposited on Mars in the summer of 1976 have a very interesting
array of sensors and scientific instruments, which are the extension of human senses to an alien environment.

The obvious post-Viking device for Martian exploration, one which takes advantage of the Viking technology, is a Viking Rover in which the equivalent of an entire Viking spacecraft, but with considerably improved science, is put on wheels or tractor treads and permitted to rove slowly over the Martian landscape. But now we come to a new problem, one that is never encountered in machine operation on the Earth’s surface. Although Mars is the second closest planet, it is so far from the Earth that the light travel time becomes significant. At a typical relative position of Mars and the Earth, the planet is 20 light-minutes away. Thus, if the spacecraft were confronted with a steep incline, it might send a message of inquiry back to Earth. Forty minutes later the response would arrive saying something like “For heaven’s sake, stand dead still.” But by then, of course, an unsophisticated machine would have tumbled into the gully. Consequently, any Martian Rover requires slope and roughness sensors. Fortunately, these are readily available and are even seen in some children’s toys. When confronted with a precipitous slope or large boulder, the spacecraft would either stop until receiving instructions from the Earth in response to its query (and televised picture of the terrain), or back off and start in another and safer direction.

Much more elaborate contingency decision networks can be built into the onboard computers of spacecraft of the 1980s. For more remote objectives, to be explored further in the future, we can imagine human controllers in orbit around the target planet, or on one of its moons. In the exploration of Jupiter, for example, I can imagine the operators on a small moon outside the fierce Jovian radiation belts, controlling with only a few seconds’ delay the responses of a spacecraft floating in the dense Jovian clouds.