"I became operational at the HAL Plant in Urbana, Ill., on January 12, 1 1997," the computer HAL tells his interlocutors in Arthur C. Clarke's 1968 novel, 2001: A Space Odyssey. That day is upon us, but nothing resembling an intelligent computer is ready to be switched on in Urbana or anywhere else. Clarke was in many respects an acute visionary: he predicted the existence of communications satellites-and their effect as cultural cement mixers-so accurately that life may well have imitated art. But artificial intelligence (Al) remains the same "four to 400 years" away that the field's namer, John McCarthy, estimated some 30 years ago.

Many computer scientists have given up entirely on Al, and HAl's Legacy-
both in its text and as an object lesson- may help lay readers understand why. Machine-intelligence researcher David G. Stork has enlisted a dozen computer scientists (plus a mathematician and a philosopher) to consider, at heart, a question he was asked at a dinner party: "How realistic was HAL?" The authors tackle a variety of topics in hardware, software and cognitive science with a firm conviction that they are discussing the building blocks of an intelligent machine like HAL, but, tellingly, no underlying coherence emerges. Much as the expert systems of the 1980s were full of "brittle" knowledge that proved useless outside narrowly specialized applications, so human experts in supercomputer design, fault tolerance or computer chess seem unable to step outside the boundaries of their disciplines.

A discussion of the subnanosecond "clock speeds" (the time necessary to carry out each operation) that are potentially possible using gallium arsenide integrated circuits offers a prime example. Ultrafast circuits may be useful for the kind of supercomputing that tackles complex physical simulations, such as predicting weather or modeling the interior of a hydrogen bomb, but it is not at all clear that this kind of computing has anything to do with intelligence and self-awareness as cognitive scientists are beginning to understand it. Similarly, the state of the art of fault tolerance for computer hardware and software is only minimally relevant to the story of HAL's "mental" breakdown in 2001; Clarke attributed HAL's troubles to a fundamental and quite emotional contradiction in its duties to its crew and to its mission.

Even those contributors who recognize that standard microchips and operating systems are unlikely to yield intelligence can succumb to technological tunnel vision. Inventor Raymond Kurzweil makes the remarkable assertion that increases in the resolution of brain-imaging technology will shortly enable researchers to map human neurons into silicon. That bold leap is a little like imagining that a simultaneous readout of the speedometers in all the cars on Boston's streets would let you predict the results of its next mayoral election.

So many aspects of everyday life as depicted in 2001 have receded over the technological horizon that it should come as no surprise that even the most basic of HAL's abilities-carrying out simple conversation-is beyond modern computers. They cannot reliably convert sounds to an internal representation of meaning; they cannot even generate naturally inflected speech. Machines still lack the enormous, implicit base of knowledge about the world and the intuitive understandings of emotion or belief most people take for granted.

Joseph P. Olive of Bell Laboratories does a good job of explaining the complexities underlying effective speech synthesis. Not only must a computer know what it is saying to produce the proper inflections, but it must also mimic the vagaries of a human vocal tract with surprising precision. Visual cues., too, are important in conveying meaning. Olive and his colleagues have found that displaying an animated, expressive face synchronized with the voice can make the result significantly more intelligible. "If HAL had had a real face, rather than one large eye," he asks, "would it have been so easy to kill him?"

There are always pitfalls in reading too much into the technology or psychology of a fictional entity. Arguments over what kind of being HAL was are ultimately even less resolvable than questions of the true intentions of Clarke or Stanley Kubrick (who collaborated on the screenplay in addition to directing the movie version of 2001). Murray S. Campbell of the IBM Thomas J. Watson Research Center delivers an interesting discussion on the way that human chess-playing styles differ from computer ones. He goes astray, however, by indulging in a tediously detailed examination of the 1913 chess match between two undistinguished German tournament players from which the closing moves shown in the film were taken.

Perhaps a dark side of HAL's legacy is to have fixed an anthropomorphic view of artificial intelligence so firmly in the minds of a generation of researchers that one of them would take such a throw-away detail so seriously. During the 1980s, those anthropomorphic visions found some kind of fruition in computer programs that demonstrated near-human and occasionally superhuman autonomous abilities in a range of fields from medical diagnosis to ore prospecting or financial analysis.

But these idiot savants did not show even the slightest signs of achieving general competence. In the subsequent AI winter-brought on by the end of a military research spree as well as the inevitable collision between venture capital and reality-only the mechanical cockroaches survived. Researchers scaled back their ambitions and aimed at achieving the cognitive and survival skills of a lobster or a cricket rather than a
virtuoso surgeon or an ace fighter pilot. if mechanical evolution proceeds a million times faster than its natural predecessor, we might expect the emergence of a digital dog in a century or two. Some stalwarts of the Al establishment, however, are calling for a reevaluation of the essential goal of artificial-intelligence research. They contend that trying to create a thinking machine-for the time being, at least-is like asking the Wright brothers for an artificial bird, complete with feathers and flapping wings. Patrick Hayes of the Beckman Institute (who did not contribute to HAL's Legacy) has coined the term "cognitive prosthesis" to embrace a range of software tools, including automated memory aids and job-scheduling systems, that help people think more effectively, much as cars help them to move from place to place or hydraulic presses help them to bend and form metal.

No one expects to get into a car and sleep at the wheel while being conveyed automatically to the correct destination, and perhaps no one should expect a computer program to diagnose patients infallibly or to command a major battle. But even now software written using the techniques developed by Al researchers reminds doctors of possibilities they might have missed or schedules the transport aircraft that deliver supplies to combatants in far-off lands. The programs are nothing like HAL, but without people working toward the same vision expressed by Clarke and Kubrick, even these limited intellectual tools would not exist.