Computer boffins pop AI's $60m question

Publication Date: 
Sun, 06/09/2002
Media Coverage Category: 

June 09 2002 at 12:37PM
Austin, Texas - Day after day since 1984, teams of programmers, linguists, theologians, mathematicians and philosophers have plugged away at a $60-million (about R600-million) project they hope will transform human existence: teaching a computer common sense.

They have been feeding a database named Cyc over a million truths and generalities about daily life so it can automatically make assumptions humans make: Creatures that die stay dead. Dogs have spines. Scaling a cliff requires intense physical effort.

Though some critics question the potential of this painstaking effort, the inventors believe Cyc will form the brains of computers with supercharged reasoning abilities - which could help us work more efficiently, make us understand each other better and even help us predict the previously unforeseeable.

Cyc (pronounced "psych") has already helped Lycos generate more relevant results on its Internet search engine. The military, which has invested $25-million in Cyc, is testing it as an intelligence tool in the war against terrorism. Companies use Cyc to unify disparate databases and are examining a new application that warns when computer networks have vulnerabilities hackers can exploit.

This spring, the developers' company, Cycorp Inc, sent their 18-year-old creation off for some higher education, creating a Web link to let the public download Cyc's knowledge base and teach it things too.

Cycorp's founder and president, Doug Lenat, believes that if enough people log in to share more of the world's collective wisdom, Cyc quickly will become vastly more useful.

For now, Cyc is just a few hundred megabytes that can be stored on a single CD. Someday, Lenat envisions it becoming standard equipment in computers or being placed on a network server to fuel dozens of applications. It could annotate emails to put them in better context for their recipients, serve as an instant language translator, even offer humans advice from varying points of view.

"This is the most exciting time we've ever seen with the project. We stand on the threshold of success," Lenat, 51, said recently in Cycorp's offices in a quiet Austin complex. "What people are able to do on a day-by-day basis could be dramatically increased if we are successful."

Such hopes are not new in artificial intelligence, which has to date produced far more disappointment than marvel, with products of only limited use beyond specific tasks.

In 1983, when Lenat was a professor at Stanford University and a researcher for Atari, he decided artificial intelligence would go nowhere unless someone took the time to create a catalog of common sense that would let a computer recognize absurdities as well as humans can.

With colleagues at Microelectronics Computer Corp, a technology research consortium, Lenat began creating Cyc in 1984.

By typing messages in CycL a programming language created especially for Cyc, Lenat's team first taught it that there are things in the world, and that some are individual (such as the Parthenon) and others are collections (historic sites).

The programmers eventually took chunks of text and thought about every assumption the author knew readers would make. Upon reading something about how the Duke of Wellington was moved by Napoleon's death, the programmers decided to tell Cyc it could assume Wellington outlived Napoleon, knew him when he was alive, heard about his death - and so on.

The goal was not just to fill Cyc with straightforward facts but to "generalise as much as possible until further generalisation would be false," Lenat said.

The result is that if you ask Cyc whether Lassie has a nose, it would reason that Lassie is a collie, collies are dogs, dogs are macroscopic vertebrates and macroscopic vertebrates have noses, so yes.

The researchers also told Cyc to ask questions if it decides it needs more clarity about a concept.

In 1986 Cyc asked whether it was human. That same year it asked whether any other computers were engaged in such a project.

Lenat's team taught Cyc to make sure everything it was told conformed with everything it already knew - a protection that should keep Cyc from being filled with erroneous information during its public education, which for now is possible only on computers with the Linux operating system.

Cycorp was spun off in 1994 into a privately held company that Lenat says has been profitable from inception, funded by the government, private investors and side projects such as the Lycos search-engine deal, which ended last year.

Cyc's job at Lycos was to make sense of ambiguous search results. If a user entered "vets", Cyc would ask whether he meant veterinarians or veterans and then have appropriate follow-up questions.

Amusingly, the Lycos stint provided Cyc with an adolescence, because it learned about sex-related terms users typed into the search engine. Cyc's programmers taught it that certain things in the world are salacious and shouldn't be mentioned in everyday applications.

Some artificial intelligence experts question whether Cyc can be as revolutionary as Lenat predicts. They claim it is far more efficient to make computers search for and identify patterns than to have them follow predetermined sets of rules.

"There's a lot more to common sense than can be captured in a set of rules," said . Pandurang Nayak, chief architect for Stratify Inc, which uses pattern-matching technology to manage data in various kinds of files. - Sapa-AP

On the Net: http://www.iol.co.za/outgoing.php3?URL_to=http://www.cyc.com