I stumbled on this interesting video the other day. DeepMind is a company founded to create general artificial intelligence (“general” meaning that the AI can understand all kinds of different things, like a human being can). Most AI is “narrow” – which means it it specifically created to solve one problem (like play chess very well).
The whole video is interesting, but you can jump ahead to 10:00 to see the AI learn to play 1980s video games. It does a surprisingly good job at playing Space Invaders and Breakout. It doesn’t do so well with more complicated games like PacMan, though. The amazing thing is that the only information they feed the AI is the screen image. The AI slowly learns to play the game by trying to maximize his score.