In this talk I argue that instead of trying to construct generally intelligent systems, we instead should focus on making systems intelligent in many highly specific domains. It is widely accepted that context influences people's perception of the outside world. What a person perceives is some combination of the signal from the world and her expectations given the current situation. If we want computers to behave as humans behave, we must provide them with similar interpretation capabilities. It is not enough to simply build pattern recognizers for individual objects when computers currently lack the highly specific, contextual knowledge that is essential to providing the expectation component of perception. Therefore, we should focus on incorporating into computer systems these highly specific worlds of knowledge--these million million expert systems. I will first give examples of the importance of contextual knowledge from cognitive psychology and from my own work in sketch understanding. I will then discuss how these bodies of knowledge can be represented, interconnected and acquired.