Google (GOOGL)’s artificial intelligence firm, DeepMind Technologies, has been making headlines with a computer that can program itself, a quest for “AI scientists” and software that can learn to play classic Atari games better than a human expert.
Today, a video showing their famed software learning to play Breakout and mastering nine other Atari games including Enduro, Boxing and Demon Attack has resurfaced online. Although it was uploaded to YouTube last April, it seems it’s finally receiving the attention it deserves.
In the beginning of the video—taken at the 2014 First Day of Tomorrow conference— DeepMind founder Demis Hassabis narrates as footage of the software training to play Breakout shows behind him. Note that “training” just refers to time spent playing; the software was merely presented the game with no prior knowledge as to how it works or how to go about learning to play it.
At first, it’s clearly not very good, but after an hour of training (about 200 games), the software is performing measurably better with a 30-40 percent success rate. With only one more hour of practice, the software never misses the ball and is playing better than any human could. After even more training, the software sort of outsmarts the game, playing more offensively than defensively by strategically hitting the ball in ways that will help it win more efficiently and with less risk of missing.
The video then goes on to show the software “ruthlessly” winning at some of the dozens of other classic Atari games it’s similarly mastered.
Prepare to be amazed.