Time-Digital Magazine, October 1998 -- Reprinted with permission

Get Smart: How intelligent can artificial intelligence be?

Lev Grossman talks to Doug Lenat to find out

Hal, it wasn't. Still, plenty of people had Space Odyssey-like expectations for a computer program that was pushing the limits of the hot field of artificial intelligence (AI). The program was written by the Ford Motor Co., no less, and was designed to be "smart" enough to actually process a car-loan application. Imagine the possibilities, the end of so much drudge work, the money saved. But would it work?

One day in the mid-1980s, the program was finally crash-tested. In went a fatally flawed application from a 20-year-old---who cited an employment history that went back 10 years. The program didn't blink.

Loan approved.

Oops.

The program didn't know enough to question whether a 20-year-old could have worked half his life. Its failure was one of many that was experienced at research facilities worldwide. After years of pouring millions into the once promising field of artificial intelligence, the half-billion-dollar-a-year industry went into a steep decline. AI software companies were blowing up, forced into bankruptcy. The period that scientists call "AI winter" had begun.

Enter Douglas Lenat, then 34, a computer-science professor at Stanford. In 1984 he proposed a plan for an AI program that could not only tell the difference between a child and an adult but would also know that most countries have child labor laws. For that matter, it could probably try a case itself. By his estimate, his AI would take 20 years to build and cost more than $25 million. He called it Cyc (rhymes with bike), short for "encyclopedia."

Lenat knew that Cyc was too expensive for a university to fund, so he sought funds from the Microelectronics and Computer Technology Corp., a research consortium then backed by, among many others, Microsoft, Apple and the U.S. Department of Defense. In 1995 he founded Cycorp to develop and market commercial versions of Cyc. The company, with headquarters in Austin, Texas, has brought Cyc a long way. If it works as its master thinks it will, AI spring could be coming soon.

From the start, Lenat assembled a unique group of talents to build the program. members of his core team of 30---who like to call themselves "Cyclists"---include philosophers, anthropologists and linguists, as well as programmers. The extraordinary task they faced was nothing less than dividing up and articulating the most basic structures that make up the universe as humans know it.

To Lenat, the problem is simple: computers don't know enough about the real world to solve real-world problems, so you have to explain to them what it's like to be us, living in our messy landscape of grass, dirt, love, hate and shopping carts with wobbly wheels.

But how do you teach a computer about the real world?

Lenat puts the problem thus: "Suppose you're an alien who has come to Earth. You have to figure out, What are the categories? What are the concepts? What are the relationships? If anything, it's more like anthropology than anything else. We're studying the world and how humans see the world, as viewed by the outsider---Cyc is the outsider."

Lenat, now 48, is a cheerful man, round rather than fat, with intensely twinkly eyes and brush of brown hair that swoops straight up. He is widely considered one of the greatest computer scientists of his generation, and he has that eternal boyishness that's weirdly common among the software elite. As he talks, he pushes himself energetically around his cluttered office in a rolling desk chair. "We took enormous sheets of white paper," Lenat remembers, "and drew wall-size diagrams of the topmost few thousand concepts that would need to be present."

Among the categories that emerged was things, followed soon after by subordinate concepts such as tangible objects and intangible objects. The group went on to sketch out events, processes and animals (which are both tangible and intangible at the same time). Soon it came up with stuff, a subcategory covering things like peanut butter, which continuous and homogeneous, as opposed to things like marbles, which are discrete objects.

John McCarthy occasionally advised Lenat on the project. McCarthy is a professor of computer science at Stanford and a founding father of the field. He is famous for, among many accomplishments, coining the term artificial intelligence. He helped focus the group's efforts by formulating a pair of test sentences: "Napoleon died on St. Helena. Wellington was greatly saddened."

To understand what they meant, Cyc would have to grasp such concepts as time, space, geopolitics, history, mortality and the complexities of human motivation. Why would Wellington, Napoleon's enemy, mourn the passing of his old foe? Cyc flunked the test first time out, but its failures acted as a guide. When it ran into something that its knowledge couldn't account for, it asked questions. Lenat answered them.

Cyc's mind is like an exhaustively cross-referenced encyclopedia. Listed under bird, for example, are the rules "Every bird is an animal" and "Every bird is an air-breathing vertebrate." A long list of statements follows, sounding a little like something by Gertrude Stein:    Every robin is a bird.    Every thrush is a bird.    Every hummingbird is a bird.    Every raven is a bird.    Every ibis is a bird.

Further on, the rules become increasingly precise. Under the subheading HumanSocialLife, we find: "If ?X is a bird and ?X feels somewhat level of pain, then ?X evokes positive sympathy." In other words, if a person sees a bird that is suffering, that person will usually feel sorry for the bird.

The sum of these facts and rules---about a million of them---is the knowledge base. Some have a weirdly insightful quality. There is everything from practical etiquette ("While a person is driving a car, eye contact is not socially required during conversations") to the romantic ("You can usually see people's noses, but not their hearts") to the theological ("When people die, they stay dead").

At this stage, Cyc can answer only specific kinds of questions, although it answers them quickly and accurately---sometimes with surprising intuition. Given a database of sample phrases and a vague query like "Show me happy people," Cyc selected the phrase, "A man watching his daughter learn to walk."

Specialized versions of Cyc are already on the market. Cycorp's clients, among them Glaxo Wellcome and Kanisa, use Cyc to perform tasks that are repetitive and boring but require common sense, like translating the chatter between two databases. One product in development would employ Cyc to "scrub" its patient database for errors: if twins are listed with different ages, for example, Cyc is smart enough to know human error is afoot.

But the Cyclists still have enormous obstacles to overcome. Cyc completely lacks a sense of humor, and it doesn't yet thoroughly understand everyday English, which is rife with the kinds of petty ambiguities that paralyze computers. Indeed, McCarthy is skeptical of Cyc's potential: "I agree with Lenat's approach, but I think he is expecting too much from the methods he currently has available. Without a conceptual breakthrough, Cyc won't reach the level of HAL."

But when Cyc does learn English, Lenat says, it will start reading books and crawling the Web on its own. Already Cyc can read and understand texts, such as almanacs, that are mostly hard data. In a recent test conducted by the Defense Advanced Research Projects Agency, Cyc read a set of foreign policy documents, including the CIA World Fact Book, and then fielded questions on how it would handle such hypothetical crises as the closing of the Suez Canal by terrorists. The Cyclists see themselves as merely jump-starting a knowledge-acquisition process that will eventually run by itself.

None of this is to say that there aren't other AIs out there. Limited forms of AI are common, in fact. The handwriting-recognition software used in handheld computers, for example, is smart enough to read your scrawl. The Navy trains AIs to look at infrared images and recognize potential targets.

There are projects in the pipeline too. A group called the Guardian Angel Consortium, which includes doctors and scientists at the Massachusetts Institute of Technology and Tufts University, is working on a software agent that's designed to stay with a person for a lifetime. It will learn every aspect of your medical history so that when you're in Emergency having a heart attack, it will make sure that everybody in the room knows you are allergic to nitroglycerin. M.I.T.'s AI lab is at work on a Star Trek--style intelligent room that watches people inside it and responds to their spoken commands.

But most of these AIs are what scientists call "expert systems": they have one very narrow, very limited field about which they know everything there is to know---but that's all they know. Cyc's area of specialty is the entire scope of human knowledge. Listening to Lenat, you begin to realize just how ambitious his plan is. "I really see Cyc as underlying most of the software in the 21st century," he says matter-of-factly.

Cyc is remarkably small, hardly more than 100 MB. You don't need a Cray supercomputer to run it: even a current low-end desktop machine will do. Lenat envisions computers stamped CYC INSIDE the same way they're stamped INTEL INSIDE now---and not just computers. He sees a future in which Cyc will inhabit our phones, our TVs, our cars, our bedrooms. We will spend every minute of the day with some form of Cyc.

"When you make a gesture using a mouse or something," Lenat explains, "instead of your having to make the whole gesture, it'll understand enough about who you are and what you're trying to do that just your starting to make the gesture will be enough, the same way an old couple can finish each other's sentences."

Cyc could mediate every contact we have with machines, and potentially with other humans. It could learn to speak every human language, translate conversations as they're spoken and perform vast amounts of mental labor perfectly--and for pennies.

In the meantime, Cyc has to work for a living, while growing up in our real world. Who knows? One day we might find Cyc teaching us instead of the other way around.

Once a group of Cyclists gave Cyc a database of photographs to play with. When they asked Cyc to show pictures of grandchildren, it responded by showing all pictures in the database that had people in them. It took the Cyclists a few seconds to get Cyc's point: everybody is somebody's creation.