Intelligent computers have long been a staple of science fiction, usually with dire consequences for their creators. HAL 9000, the remarkable computer that ran amok in Arthur C. Clarke's "2001: A Space Odyssey," was not designed to be malicious. But HAL had been programmed not to lie, and later given instructions to do just that.

HAL was an example of artificial intelligence, a field of computer science that has trouble dealing with common sense. Human beings make numerous assumptions about the world before they proceed to abstract thought, whereas computers can't assume anything. When you try to make a computer intelligent like HAL or the Enterprise's computers in "Star Trek," you have to program in every premise that a human being would already understand.

And that's precisely what a computer scientist named Douglas Lenat is trying to do. For the past 17 years, Lenat has been building a database of assumptions for a project named Cyc (pronounced "psych"). First at an Austin, Texas, consortium (the Microelectronics and Computer Technology Corporation), and now at spinoff Cycorp, Lenat has chosen to build artificial intelligence one proposition at a time. His work is a labor-intensive task that could well consume his lifetime.

Cyc isn't a machine, although a film crew once mistook Cycorp's air conditioner controller for it and ran an amusingly mistaken photograph. It's simply software, but software with an extraordinary difference: Lenat is trying to build a knowledge base so vast that it would cover all the assumptions we now include as human common sense. 500 person-years have gone into this project, as well as a total of about $50 million from the Defense Department, Microsoft co-founder Paul Allen and pharmaceuticals giant GlaxoSmithKline.

To understand how ambitious Lenat's goal is, consider the checkered history of artificial intelligence. To create a computer that can learn by itself -- and ultimately attain something we might call "consciousness"-- scientists labored through the 1960s and later, only to find their hopes dashed. Human learning itself is barely understood; creating machinery that can mimic it proved to be a task too daunting for the most powerful computers.

The result was that artificial intelligence began to focus tightly on specific tasks. So-called "expert systems" were developed that could do only one thing, but do it well. IBM's Deep Blue knows how to play chess well enough to beat world champions such as Garry Kasparov. But don't ask it how birds fly, or whether "Hurricane Floyd" refers to a weather system or a hockey player. Other expert systems are similar; you might be able to use one to diagnose an illness, but only if it were built around a highly specific set of data for that particular task.

What bothers Lenat is that this tight focus makes machines that are susceptible to failure. Make a single typing mistake and you freeze the system because the computer can't figure out what you really meant. On a basic level, this is why my spell checker (itself an expert system of sorts) can tell me when I've written a word that's not in its vocabulary, but not when I've written a legitimate word but am using it in the wrong place.

So how do you teach a computer to be as bright as a 5-year-old kid? At Cycorp, the answer is that you start inputting facts about the world, setting up categories and creating software that allows the system to make inferences. After 17 years, Cyc now contains 1.4 million assertions, which include elementary facts about the world. Cyc needs to know that rain makes things wet. It has to be taught that people have a mother and a father.

Next page >

Get $150+ in coupons in every Sunday N&O. Click here for convenient home delivery.