Until
next week,Dateline:03/02/97
YOU know how it is. You take the ol’ V3-engined Browser GT out for a Sunday afternoon spin down the Infobahn, tooling along at a modest 28.8 kilobits per second, stopping at a Yahoo here or an Altavista there (fast food for the Web traveler; not very filling—that’s why we have The Mining Company!) to ask for directions to nowhere in particular, when all of a sudden there it is, filling your windscreen.
The "it" in my case was a kindergarten for what I like to call Machina Sapiens, and others call electronic brain. Whatever you want to call it, we’re talking about (possibly, and arguably) the ultimate in artificial intelligence: a non-human entity that could pass the Turing Test. You can get more information about Alan Turing and his famous Test by clicking here, but here’s my simplified rendition of it:
Let’s suppose you are the person conducting the test. You sit at a computer terminal, which is connected to two other computer terminals, A and B, hidden from your view. At one of the hidden terminals is another human being. At the other? No-one. That terminal itself has intelligence programmed in. You don’t know which is the human—A or B—but your goal is find out by asking them anything you like, typing your questions and receiving their answers on your terminal. Turing believed that if you emerged from the test unable to say which was which, then the non-human program really is, to all intents and purposes, intelligent.
Some people don’t like the premises of the Turing Test, but it’ll do for our present purpose. Which brings us back to M. Sapiens’ kindergarten. It’s a Web page for a project with the acronym MISTIC devoted to building a database of general knowledge for consumption by an AI program, such as a neural network. (I’ll leave you to visit the site for the meaning of MISTIC, if you really want to know.) When the program (the infant machine) has absorbed all the knowledge, it will be ready for the Turing Test. Or so the MISTICs believe.
I think MISTIC is an intriguing idea. Or at least a valiant effort. I have to wonder, though, whether it is really necessary. The general knowledge being captured in tiny increments by the project is already available in mass in books, movies, Web pages, Usenet newsgroups, and on and on. We write down bits of general knowledge all the time. I might write in a newsgroup about Michigan: "The last of the Winter snows have usually melted by April each year." A neural net that could read would surely be able to deduce its own views of the world (such as "The daytime sky is usually blue") from the content of literature, so why bother to set down information that’s already set down?
MISTIC is not the only effort to create an artificial intelligence with a built-in knowledge of the world; indeed, two older, more ambitious, and more advanced projects already exist. They are Cyc, from Cycorp, Inc., a spinoff from AI pioneer Doug Lenat's work at MCC (Microelectronics and Computer Technology Corporation, part of the U.S. answer to Japan's Fifth Generation Project), and ThoughtTreasure, a project by MIT and UCLA-trained computer scientist, author, and AI programmer Erik Mueller.
"The Cyc system," Cycorp's Web site tells us, "comprises a very large, multi-contextual knowledge base, an inference engine, a set of interface tools, and a number of special-purpose application modules, running on a variety of platforms. The knowledge base is built upon a core of approximately 400,000 hand-entered assertions (or "rules") designed to capture a large portion of what we normally consider consensus knowledge about the world. (For example, Cyc knows that trees are usually outdoors, that once people die they stay dead, and that a glass filled with milk will be rightside-up, not upside-down.) This foundation enables Cyc to address effectively a broad range of otherwise intractable software problems."
The goal of Mueller's ThoughtTreasure project, which he says "will probably take many years to achieve," is "to develop a computer program which can understand, learn from, and communicate in natural language, using any and all available techniques." It's making pretty good progress; check it out.
In contrast to Cyc and ThoughtTreasure, MIMIC’s goal seems more modest. It is mainly to provide an initial training set for a neural net, and perhaps that's not a bad way to go about it. Start with a few (million, but that’s still a drop in a bucket that has no bottom) basic assertions and THEN let the neural net go out and read the books and the newsgroups for and by itself. How about starting it off with kindergarten primers even simpler than MISTIC? In other words, why not teach it the way we teach kindergartners? This is Spot [picture]. See Spot run [picture of Spot running]. Spot is a cat [picture of Spot among a group of cats]. Etc.
But I mustn’t ramble. if you are intrigued as I am by such things, why not pop over to MISTIC and put in your two cents’ worth of general knowledge? They’ve made it pretty simple. They are hoping to get one million "stimulus/response pairs" (pearls of wisdom) by July 1, 1997, and ten million pairs by July 1, 1998, from World Wide Webheads like you and me. As of last November, they had 246,788 items in the database.
I hate to end on a pessimistic note, but feel I should issue two cautions: First, the last entry on the MISTIC site was dated November, which leads me to wonder if the page—and the project—is alive and well. I think so, because I just received email thanking me for contributing to the knowledge base, but that may have been an email bot. Second, it’s not clear whether the project is being conducted under formal scientific auspices. But even if not, amateurs do occasionally beat the professionals to the punch, and I wish the organizers well.
Until
next week,
NEXT WEEK: Artificial Intelligence and God: An interesting question.