Computers versus Common Sense An Engineering Approach to AI (04 Jun 2003)
Computers today are idiot-savants. They may manage bits flawlessly and furiously, but they have no understanding of what those bits signify. And they have poor models of themselves and of the human beings they serve and represent. To break that "brittleness bottleneck," we need a new software layer that contains the millions of things the average person knows about the world. Some of this is factual, such as how often a U.S. Presidential election is held, or even ephemeral, such as the name of the current President; but most of the needed content is more like rules of thumb, such as why one should carry a glass of water open-end up. In terms of a newspaper or book, we are talking about codifying the white space - the things the authors don't need to bother saying (e.g., the White House is in Washington, D.C.; tables have flat horizontal tops; appliances stop working when turned off.) Since 1984, my team has spent the seven person-centuries necessary to build that artifact. In this talk, I'll describe what we did, and why, and some of the lessons we learned about representing commonsense knowledge, and doing reasoning in huge knowledge-based systems. In particular, I'll explain why we took an empirical, engineering approach to the problem, rather than a theoretical, scientific approach. I'll also discuss some current and future commercial applications of our technology (CYC).
Article URL: http://murl.microsoft.com/LectureDetails.asp?1032
(Doug Lenat (Cycorp, Inc.))
Read 123 more articles from Stanford HCI Seminars sorted by
Next Article: Designing for visitors who dont enter through the home page