When will machines outthink us?
msnbc.msn.com
msnbc.msn.com
The traditional approaches to build a strong AI all involved forms of explicit knowledge representation: where facts about the world would be stored in memory and then pieced together to come to new (hopefully intelligent) conclusions and plan actions. Most people quickly realized that this approach doesn't scale at all: most knowledge that humans have about the world is very context-dependent and fuzzy. The new approaches around machine learning and probabilistic modeling are more promising but almost all of them involve building mathematical models around a specific problem rather than building a general learning machine.
Most leading AI researchers in academia aren't even trying to build a general-purpose intelligence anymore, instead focusing on solving domain-specific problems, and the field of AI has been much more successful over the last fifteen years or so as a result.