or several months now, Microsoft has been blitzing the airwaves with commercials for "enterprise software," a mysterious commodity which is apparently supposed to amaze us with its ability not to hold grudges,
be surprised or get bored or lonely. I find this highly amusing, since all it really says is that the world's favorite monopoly is still operating in the mechanistic 20th century, whereas in both tradition and fact, 2001 is the year of artificial intelligence.
2001 occurred for the first time in 1968, when writer Arthur C. Clarke and director Stanley Kubrick introduced the world to HAL, an intelligent computer with human-like emotions--including murderous anger and fear. This week it occurs again, with a blockbuster collaboration between Kubrick (now deceased), SF author Brian W. Aldiss and writer/director Steven Spielberg called, appropriately enough, Artificial Intelligence. And next month, 2001 will occur yet again with the commercial launch of Cyc, an honest-to-God A.I. which could, among other things, revolutionize the business of Internet search engines.
"Artificial intelligence" is really an umbrella term which refers, vaguely, to a whole host of different technologies, and like many such terms, it's frequently misapplied for purposes of hype. In science fiction and the popular imagination, it refers specifically to conscious, personable entities like HAL and David and Star Trek's Commander Data. In industry, it means any system capable of replacing human judgment for a particular (usually very limited) application. Systems which don't require judgment, or whose judgment is trivial and not based on any sort of memory or learning, are occasionally called "smart" (e.g., smart tires for cars, and smart focus for cameras) but shouldn't be thought of as intelligent.
The power of speech is associated with intelligence in the animal kingdom (we'll be talking more about this in next month's "Planet of the Hominids"), and it certainly goes a long way toward personifying computers and the software which runs on them. Consequently, speech synthesis is sometimes considered an A.I. issue, particularly in the text-to-speech arena, because the sounds of letters depend so heavily on what they're being used for. This is a marginal example, though, and the vast majority of speech systems, while complex, are as "dumb" as, forgive me, Microsoft's enterprise software.
Genuine A.I. applications include speech recognition, which translates imprecise sound waves into exact streams of dictionary words and phrases, and natural language processing, which translates ambiguous words and phrases into compact nuggets of logical meaning. Both problems are extremely difficult, and have been the subject of ongoing research since the 1960s. Both are rapidly yielding, though, and exist today in crude but workable form in various desktop products.
Another A.I. challenge is the art of conversation. The ability to speak, and to understand when spoken to, do not by themselves imply any ability to converse intelligently. Probably you have co-workers who demonstrate this, and if not, well, there are always politicians. Still, the "Turing Test," proposed by British mathematician Alan Turing in 1950 as a marker of machine intelligence, requires that a computer successfully pass itself off as human. In the 1990s, programs became available which could, in limited settings such as online chat rooms with narrow subject matter, pass the test. In fact, Cambridge University sponsors an annual Loebner Prize for the software that does it best, although the $100K Grand Prize has yet to be awarded. These programs aren't conscious, self-aware or even remotely intelligent in the human sense--they're a bit of a parlor trick, really--but they do hold great commercial promise, especially when linked up with speech synthesizers and recognizers. Imagine an answering machine that could act as your personal secretary, or even a digital stand-in for yourself.
Behold the Power of Brain
Vision is another hotbed of A.I. research. The traditional application for this is the spotting of particular objects, such as parts on a factory assembly line. Technology continues to advance in this area, but the fundamentals were really laid down in the 1970s, and have changed little since that time. (I've even dabbled in this field myself.) A relatively new area, though, is face recognition--the rapid analysis of those minute differences which make each human face unique. There are already commercial systems that can do this, for a limited database of perhaps 100 faces. Rockefeller University has one which someday soon will be searching for known terrorists in our international airports. In similar research, terrain analysis is the subject of the PerceptOR program at the Defense Advanced Research Projects Agency; humans can tell at a glance which portions of a scene represent grass, rocks, bushes, pavement, etc., but these judgments are still quite difficult for machines. At least until PerceptOR succeeds.
Robotics, which links computer software to machine parts capable of manipulating the real world, has its own set of A.I. challenges, including path planning, obstacle avoidance and hand-eye coordination. The first two were well understood by the late 1980s, and the third was taking solid shape by the late '90s, with Hans Moravec's "Cog" at MIT, which can perform such delicate actions as playing with a Slinky, or Omnitech Robotics' automated forklift, which can identify, maneuver around and manipulate cargo pallets in its visual field. Bipedal locomotion, a necessity for android robots such as David and Data, was first demonstrated by Honda's P2 walking robot in 1997. It is definitely a neat trick, but falls into the category of "smart" rather than "intelligent." The reading and generating of facial expressions, though, is definitely an A.I. problem, also well in hand in places such as MIT's Media Lab. (This technology has a lot of desktop potential as well--imagine a computer which could see and respond to expressions of tiredness or frustration.)
The collective value of all this should not be underestimated. It's a testament, really, to the amazing power and diversity of the human brain, which performs all of these functions continuously, with the input of little more than oxygen and sugar. But the human brain has an additional ingredient to tie it all together: cognition. Without that, these "intelligent" traits are simply inputs and outputs of a mechanistic judgment system that has more in common with insects than with human beings.
Robot Friends Are Around the Corner
The cheesy way to supply cognition is to fake it, by running brute-force recipes to analyze inputs and optimize potential responses. With enough computing power behind them, these lowly methods can do a good job of mimicking human intelligence, or even brilliance. Just ask Gary Kasparov, the human world chess champion, who was defeated in a 1997 exhibition series by IBM's computational juggernaut, Deep Blue. A closer and perhaps easier simulation of cognition comes when a system is provided with memory (perhaps in the form of an expandable database), and an "inference engine" which matches new inputs up against stored patterns, searches for correlations and derives new information (e.g., more rules or facts for the database) from the results.
This is where Cyc comes in. The brainchild of computer scientist Douglas B. Lenat, Cyc was born in 1984 as a small database of "assertions" (e.g., the statement "people are made of meat") and a primitive inference engine for analyzing new ones. Since that time, with occasional infusions of cash totaling $50 million from the Defense Department, pharmaceutical companies and yes, Microsoft, Cyc has matured. Its inference engine has become a sophisticated tool for so-called "data mining" or "knowledge mining," and its knowledge base has mushroomed to more than 1.4 million assertions.
Cyc (short for "Encyclopedia") knows not only that people are made of meat, but that People magazine isn't, and that a toxin which affects human beings does not affect People. Cyc not only converses in (typed) natural language, it also exhibits a human-like quality of common sense, enabling it not only to accept new information, but to understand it and even derive theories and conclusions from it which no human has yet thought of. Cyc is expert at searching large databases (e.g., the World Wide Web) for highly relevant information based on concepts rather than keywords. In a recent demonstration, Cyc was asked to find a picture of people relaxing, and presented one of men carrying surfboards. Cyc is also good at identifying inconsistencies and contradictions in stored records.
And this trick is much more than a laboratory curiosity; next month, Cycorp will be releasing OpenCyc, a 70,000-assertion subset of the full knowledge base. This tool will be accessible via free Web pages, and its source code will be available free to developers, for inclusion in various types of software. The full knowledge base will be licensed for a fee, and will also serve as the core of a network security product and an interactive encyclopedia which answers questions posed in natural language.
Cyc probably isn't conscious (although creator Lenat sometimes has his doubts), but it certainly is the most intelligent piece of software ever written, and will serve as the first component of a "semantic web" that will gradually replace the smart Internet of today with the intelligent one of tomorrow. So it really is 2001, and the age of A.I. is finally upon us. As we build on technologies like these in the coming century, it will almost certainly be possible to create computers and/or robots with not only humanlike behaviors, but a humanlike ability to learn and reason and grow. And yes, perhaps even feel. And once this is possible, it will almost certainly be done, because as SETI and spiritualism and dolphin language studies amply demonstrate, we humans are desperate for someone else to talk to. In the future, we may all have robot friends.
Still, as my agent once remarked, two people can already produce a fully functional human being with materials they have lying around the house. What we really need, for companionship and utility alike, are entities that think and behave in radically different ways. Indeed, the really killer applications for A.I. will come from systems as different from the human brain as a space shuttle is from a nightingale. Neither design is "better" in the abstract, and people definitely keep more songbirds around than spaceships. But even a million nightingales lassoed together could not lift an astronaut into orbit. Similarly, the intractable problems of human society may wind up being painfully easy to solve, if we're brave enough to build--and listen to--an intelligence greater than our own.
Wil McCarthy is a rocket guidance engineer, robot designer, science fiction author and occasional aquanaut. He has contributed to three interplanetary spacecraft, five communication and weather satellites, a line of landmine-clearing robots, and some other "really cool stuff" he can't tell us about. His short fiction has graced the pages of Analog, Asimov's, Science Fiction Age and other major publications, and his novel-length works include Aggressor Six, the New York Times notable Bloom, and The Collapsium.
