In many areas of artificial intelligence (natural language processing, computer vision, speech recognition, automated planning, robotics, etc.), researchers used to spend thousands of person-hours hand-coding rules of behavior to solve narrow problems.  In the past ten to twenty years (depending on the subfield), we've moved away from directly writing code to solve these problems and are instead moving to a framework in which humans demonstrate the desired behavior and the machine learns to solve the problem directly.  I'll discuss (a) the major successes of the machine learning paradigm in some of these areas, (b) some core ideas on which they are built and (c) why we still have to write some code by ourselves.