Teaching an AI to beat video games still takes human imagination

In 2013, researchers at the artificial intelligence research company DeepMind in London set out to create a system of AI networks that could master any Atari game. And they had excellent results, with their system outperforming skilled humans at exponential rates. However, one game with some novel gameplay characteristics, Montezuma’s Revenge (1984), left the system totally stumped, unable to score a single point. This delightfully retro animation explores how the DeepMind team was finally able to conquer the game by borrowing concepts from human psychology. Further, the video explores the ways in which AI development remains a deeply human enterprise that demands our creative guidance, even as AIs increasingly outperform us at certain tasks.