Video is hilarious and worth watching the whole thing as others have mentioned, but there's some great stuff in the paper also. He mentions some really interesting future directions, e.g. obviating the need for the training data set of a human playing the game by having the computer initially try different sets of random inputs and continuously improving on those that do the "best" according to it's objective function. The bit about the computer trying to play "dinosaur coloring" is fantastic also.