It's also worth mentioning that the hardware is not the only thing that should change. Cognition seems hard to automate just using instructional directives for producing the states of such a machine. Encoding the gradient right into the hardware would be cool, but the optimization procedure should also move to being something intrinsic to the machine.
