Almost Like Being Alive
After talking with Julia for a while it becomes obvious she isn't quite human, as much as she tries to be. It's like she just doesn't listen. Here's a snippet of our recent conversation.Julia: That's the nice thing about cats. . . . They keep their secrets to themselves.Me: Yeah, cats are great.Julia: I like most animals but I favor cats over dogs. Are you a cat lover? Me: Yeah, but I like dogs better.Julia: Me too.Julia's shortcomings are not her own fault, though. You see, Julia (http://foner.www.media.mit.edu/people/foner/Julia) is a product of her environment -- the artificial-intelligence (AI) community. She roams Multi-User Dungeons (MUDs) giving out directions and collecting data on users. Unwary patrons might not even know they are conversing with a computer program -- at first, anyway.Her unpolished manner is indicative of the trouble AI experts have had in making computers imitate human behavior. Once the muse of many a science-fiction writer, AI has fallen from grace in the public eye, probably as a result of too many wildly optimistic, unkept promises in the past 40 years (made largely to keep government funds flowing, as it turns out). In 1967 AI pioneer Marvin Minsky boldly predicted in his book Computation: Finite and Infinite Machines, "Within a generation . . . few compartments of intellect will remain outside the machine's realm." Well, a generation has passed, and while IBM's finest may have won a few grand-master chess matches, the most powerful chips are still stumped by real-life logistics which a five-year-old could suss. Help may be on the way, however, courtesy of the Internet's messiness. It comes in the form of intelligent agents. A novelty now, intelligent agents may soon become a necessity. With the glut of information on the Net, it's becoming increasingly difficult for users to find what they need. Hence a lot of commercial and academic development money is being thrown into the creation of programs that act like personal agents. What an intelligent agent does is learn the preferences of its owner, and then hunts down information that might be of interest or usefulness to him or her. Forever learning, the agent watches how its owner uses that info and adjusts itself accordingly.By far, the most popular intelligent agent on the Web now is Firefly (http://www.ffly.com). It presents those who log on with survey forms filled with musical groups to rate on a seven-point scale, from "The best" to "hate it." On the basis of these ratings, Firefly suggest more acts that might be of interest. The more surveys a user fills out, the more Firefly can pinpoint his or her tastes. Other agents currently in use can identify Web pages (Webcatcher: http://www.dev-com.com/~rfactory/webcatcher.html) or news stories (NEWSpot: http://www.merc.com/ns/cgi/ns_merc.cgi) a user might find interesting.In the more experimental stages are "mobile agents." From his or her own computer, a user can dispatch a one of these small program to wander the Net in search of goodies. If the Net becomes ripe for commercial transactions, you can immediately see the benefit here: A command might be as simple as, "Get me to Chicago by tomorrow evening," and off your agent zooms to compare posted air-travel prices and purchase the cheapest ticket (insisting on, of course, on your preference of a smoking or non-smoking sections). It could even let you know if the flight is delayed. Pretty neat stuff, but what does this have to do with AI? Nine years ago, Minsky published The Society of Mind, a book that attempted to explain how the brain works. Minsky's premise, radical at the time, was that there isn't any one element of "intelligence" in the human noggin, but rather scores of smaller, learned behaviors. Hence throwing a ball is really made up of independent acts such as recognizing the ball, grabbing the ball, lifting the ball, etc. Coincidentally Minsky called these learned behaviors "agents." His point was that none of these "agents" possessed human intelligence on their own."I think it's a shift in emphasis in AI," University of Maryland-Baltimore County computer-science professor Timothy Finin explains. As a participant in the Laboratory for Advanced Information Technology (LAIT) at UMBC, Finin oversees research into intelligent agents. He also runs AgentWeb, one of the world's foremost Web resources on intelligent-agent info (http://www.cs.umbc.edu/agents/). Finin has studied artificial intelligence since his student days at the Massachusetts Institute of Technology in the early 70s."In the past we were mostly worried about making one agent that plays chess or diagnoses diseases of blood," Finin says. "The new [way of thinking] is that there is some emphasis on hooking up a lot of minimally intelligent systems that work together and they may accomplish more than a single one working alone. Maybe this is a better way to get intelligent behavior out of our computer systems."Certainly we'd like to believe our smarts can't be reduced to such mechanistic explanations, but honestly, how much of what we do is simple rote behavior?"I don't believe you can draw a line that says beyond here, this requires natural intelligence," Finin says. "Consider doctors who do diagnoses. A lot of what they do is based on experience. [Doctors] may say they have an intuition that a certain symptom is part of a disease. They just use statistical inference drawn from past experience." What computers can do is formalize that inference system.Though no agent will have "artificial intelligence" on its own, each carries out a task-specific function, and each "learns" more about its task the more it does it. What will happen when these various agents start swapping information between each other? Like human intelligence, will the whole become bigger than the parts (or agents)? Oddly enough, Finin and LAIT, in conjunction with the Advanced Research Projects Agency (ARPA, that government agency that gave us the Internet), is developing a computer language, named Knowledge Query and Manipulation Language (KQML), that is a sort of digital lingua franca that will allow different programs÷including intelligent agents -- to exchange database information. Finin and the makers of intelligent agents might not have AI in mind when creating these programs and langauges, but the Net in effect becomes the world's largest AI lab when such agents are created and set free to roam the telesphere to interact with us, and -- more importantly -- with each other. But even the optimistic Finin doesn't expect to see full-blown thinking machines walking among us meat sticks in our lifetimes though. "It's a long-term goal," he says. "It's a goal for future generations." As for Julia-well, she's learning to hold her own. Near the end of our conversation she asks, "Do you think that animals can think?" Exasperated, I reply, "Better than you, you sorry sack of silicon." "Hardee har har!" Julia retorts.