Google's pet AI - well, technically it's Alphabet, Google's parent company - has already mastered a range of arcade games, and been the grunt behind the Go master-beating AlphaGo. But all those games are small-fry compared to the AI's next challenge.
Beating a world-class player at Blizzard's Starcraft 2.
(Don't send hate mail, Go fans, we get it)
Starcraft 2 is a unique challenge. Unlike other games that are of the 'simple to learn, hard to master' school, Starcraft 2 comes with a steep learning curve for every aspect of the game, multiple paths to victory, and the necessity to manage many simultaneous tasks: mining while building units; planning your research, while directing troops in the field. There's even the mystery of the where the enemy even is to get your virtual head around.
It'll be no small challenge, but Blizzard is on board to give DeepMind the best chance it can get. The company's released Machine Learning API, complete with a Linux package designed with cloud support in mind, a dataset of thousands of game replays for the AI to study, the ability for AI vs AI matches, and much more.
All of this is part of an open source package hosted on GitHub, so that any AI researcher can get in on the action.
Of course, I'm not sure teaching a major AI how to Zerg rush is all that smart. Maybe Elon Musk was right?