Articles > Feature Articles

Living Up to the Nightmare

New Developments in Artificial Intelligence

By Eric Butterman and Travis C. Daub

Remember those horrible sci-fi movies you secretly watched while your parents thought you were asleep? Remember the nightmares they'd give you when you finally did go to sleep? Visions of computers outsmarting their human creators, becoming completely autonomous and wreaking havoc on lowly planet Earth were only bad dreams, as artificial intelligence has not lived up to the nightmare.

But if you're more disappointed than relieved by that, don't worry: they're working on it. Here is a list of the artificial intelligence efforts that are currently underway and how many years they are from creating a true artificial consciousness:

Neural Networks

Professor Leslie Smith, head of the computer science department at the University of Stirling, defines neural networks as a form of multiprocessor computer system with simple processing elements, a high degree of interconnection, simple scalar messages and adaptive interaction between elements.

Neural networks have been useful for only certain tasks in the last five to ten years, says Dr. Bart Selman, professor of computer science at Cornell. For example, reading handwritten text, translating handwriting into type script or understanding speech. Neural networks doing all human tasks looks beyond reach, but when it comes to limited domains for lots of data, they are excellent—those caught from credit card fraud may very well have neural networks to blame.


Selman states the agent challenge originally came from looking at robotics and then stretched to software. Those in the field realized the Internet was an interesting place to build software programs that could pretty much stand alone. A software agent is the kind of program that runs continuously performing tasks you give it, such as watching your calendar, shopping for you or even automatically searching the Web. It's almost like a personal assistant. That's where a lot of AI techniques are being applied, because an assistant needs to adapt and learn your needs in order to be successful.


According to an article by Steve Woodcock for Game Developer, finite state machines, the simplest of AI engines, are used in individual shooting games like the popular Doom. Neural networks can be an intricate part of heightening a game's degree of difficulty. By gauging a player's tactics, the network can quickly learn to counter, creating new challenges for the player. Improvements have included development in game engines and 3D graphics cards, which continue a steady tradition of upward graphics swing. In a survey Woodcock did at the 1999 Game Developer Conference, 60% of the attendees at his roundtable reported their projects included one or more dedicated AI programmers, up from 46% in 1998 and 24% in 1997. Pong would be proud.

Expert Systems

Simply stated, a computer program that contains knowledge used for a specified purpose. A breakthrough expert system was MYCIN in 1974, which diagnosed blood bacterial infections better than practicing doctors could, provided that its limitations were observed.

Natural Language Understanding

These programs cry out, "I hear what you're saying, but what do you mean?" Selman says search engines like Ask Jeeves are a major part of this growing sub-field, but their limitations are great. "Jeeves can do a reasonable job in responding to queries. On the other hand, there is a difference between being able to give a reasonable response on statistical properties and having a true understanding of natural language. We still can't say that computers get what we're telling them. We have a hard enough time understanding each other."

A continuation of the article, The Evolution of Artificial Intelligence

Eric Butterman is a New York City-based free-lance writer. Travis C. Daub is an Arlington, Va.-based free-lance writer and production manager of Foreign Policy Magazine.

artificial intelligence

Articles > Feature Articles