I didn't get much from the linked blurb, but the title made me think: if you sample training data exhaustively and then run a learning process, could you then take the training data maximizing algorithm performance to then better train humans at the same task?
Just realized that feeding video from 2600 games into an AI is a form of cheat. All sprites are going to be fed perfectly into the AI as compared to using a video camera looking at at TV where pixels will not be perfect. Still a nice achievement, but one very complex aspect of "learn by watching" has been eliminated.