From virtual reality to providing internet access to remote parts of the globe, search-engine giant Google has funneled significant funding into a variety of projects. Because of this, it is easy to forget the company’s somewhat less publicized endeavors into artificial intelligence. While Elon Musk’s fear-mongering over the potential threat of rampant AI or Facebook’s efforts to employ artificial intelligence to moderate content might be more attention-grabbing, Google recently had their own AI victory.
First founded in 2010 and acquired by Google in 2014, DeepMind is the California-based tech company’s primary effort into researching artificial intelligence. DeepMind’s past successes have largely centered around the system’s ability to learn complex strategy games and subsequently trounce human opponents. The AI system’s latest success is within this vein, but in reality is about so much more than that. Google’s eventual goal is to create a sort of generic AI that can respond to and master a broader range of challenges. The next generation of Google’s AI software, dubbed AlphaZero, may do just that.
In a paper released this week, the company revealed that AlphaZero, like its predecessor, has mastered the board game Go to a level beyond the skill of human players. The AI system, after managing to master Go in just eight hours, unbelievably accomplished the same for chess with just four hours of practice. With these newly acquired skills, AlphaZero successfully beat Stockfish, the world’s reigning champion for chess playing programs. To further punctuate this victory, AlphaZero was then taught to play shogi, a Japanese variant of chess which uses a larger game board. After just two hours of experimenting with the game, AlphaZero beat Elmo, one of the world’s top AI shogi players.
While these feats are impressive in their own right, the true shock is that, unlike DeepMind, AlphaZero was not designed with the ability to master these and similar games in mind. The new AI’s programming did not include data for tactics or strategy specific to these games; rather, it was provided with some general rules and instructions for the game, no different than what a human might be provided their first time attempting chess or Go.
Based on this, it would seem that AlphaZero demonstrates the sort of “generic” thinking ability that Google is attempting to cultivate in their AI systems. That said, Demis Hassabis, CEO of DeepMind, has clarified that while mastering these games is an achievement, they differ significantly from the sorts of scientific problems they one day hope to use AlphaZero to solve. Though many hurdles remain, Google’s paper clearly indicates AlphaZero is a massive step in the right direction.