Scaling Up: Computers vs. Common Sense

Abstract

Over the last 21 years, we’ve spent almost a person-millenium producing Cyc, an axiomatization of general human knowledge. Though still far from complete, Cyc contains over three million axioms. The need to express the range of things a person knows has led us to ever more expressive representation languages – currently we use an nth order predicate calculus with an overlay of contexts which are themselves first class objects. These pressures and others (e.g., elaboration tolerance) have driven us against numerous sorts of ”scaling up” problems. In this talk I will briefly describe Cyc, the processes by which new axioms are added and deleted, applications of it, etc., but I will focus on some of these scaling up issues and approaches we have taken – and plan to take – to keep inference fast enough and to keep contradictions from being more than a nuisance.