It’s probably been a while since you picked up an Atari controller, but even if you played for decades, you probably couldn’t beat Google’s DeepMind artificial intelligence program.
While a program that plays video games might seem gimmicky, artificial intelligence researchers (AI) have told Tech Insider over and over again that it’s one of the most impressive technology demonstrations they have ever seen.
DeepMind made big news back in February 2015, when researchers announced it could learn to play and win games on the Atari 2600 — a simple console that was popular in the 1980s — without any instructions or prior knowledge of how to play video games.
The program watched its own gameplay and learned how to win all by itself. It also had a “reward” system so it knew when it was improving at gameplay.
The computer beat all human players in 29 Atari games, and performed better than any other known computer algorithm in 43 games.
Its one weakness, according to the MIT Technology Review, was in games like Ms. Pac Man, where it had to plan ahead to clear the last dots from the maze.
Still, the program came up over and over again in the last few months, while Tech Insider reporter Guia Marie Del Prado interviewed dozens of AI researchers. Del Prado asked them what AI they have seen that really blew them away, and the Atari-playing DeepMind program was the most-mentioned.
“The DeepMind system starts completely from scratch, so it is essentially just waking up, seeing the screen of a video game and then it works out how to play the video game to a superhuman level, and it does that for about 30 different video games,” Stuart Russell, a computer scientist at the University of California at Berkeley, told Tech Insider.
You can see how bad it is during the first ten minutes of learning the game Breakout here:
“The DeepMind results on learning to play Atari games while only having access to raw pixels and the game score have been very inspiring,” Pieter Abbeel, another computer scientist at UC Berkeley, told Tech Insider. “I have been very excited about our own recent results on the same benchmark, as well as learning to walk in simulation — with a single algorithm able to learn those two very different types of tasks.”
It takes about 120 minutes for the program to reach expert level at Atari’s Breakout:
Michael Littman, a computer scientist at Brown University, told Tech Insider: “I think that’s really neat because it starts to point the way towards systems that aren’t just really really clever pieces of programming but are actually taking their experience and turning it into intelligent behaviour.”
But Russell notes the program is so good it’s almost frightening. “That’s both impressive and scary in the sense that if a human baby was born and by the evening of its first day was already beating human beings at video games, you’d be terrified,” he said.
Check out DeepMind in action playing Breakout in this video from Karoly Zsolnai-Feher on YouTube. After a few hours of existing, the AI learns the easiest way to beat the game is to tunnel through the bricks (bet your newborn can’t do that):
Business Insider Emails & Alerts
Site highlights each day to your inbox.