This paper drives a stake through the heart of claims that explicit symbolic reasoning is needed for intelligent systems. It builds extra-symbolic (deep learning) systems that substantially outperform symbolic reasoning on tests specifically designed to require symbolic reasoning.<p>Since today's programming is 1000% explicitly symbolic, the potential shift to better methods that are extra-symbolic will be pretty disruptive!
This paper is a great overview of the state of the art in building flexible generalizable deep learning systems. Not easy to read, but if you puzzle through it you'll understand where the field is and where it is going.