AlphaGo's Designers Explore New AI After Winning Big in China

Google's DeepMind lab is retiring its Go-playing machine as researchers eye a much bigger future.
170523AlphaGONS00491.jpg
Noah Sheldon for WIRED

WUZHEN, CHINA --- After winning its three-game match against Chinese grandmaster Ke Jie, the world's top Go player, AlphaGo, is retiring.

Demis Hassabis, whose Google-owned artificial intelligence lab DeepMind built this historic machine, tells WIRED he will now move the machine's designers on to other projects. "These are some of the top people in the company," says Hassabis. "The idea is to really explore what we can do in other domains." Considering the world-shaking success of AlphaGo, those explorations promise some very powerful possibilities.

Last March in Seoul, South Korea, AlphaGo became the first machine to beat a top player at the ancient game of Go when it defeated Korean grandmaster Lee Sedol---a feat that most researchers didn't consider possible just a few years before. In the year since, Hassabis and his team significantly improved the machine. Today, in Wuzhen, China, AlphaGo won its third game against Ke Jie, and much as in the other two, the contest held little drama, even as the machine's peerless play sent the usual ripples across the worldwide Go community.

When DeepMind first unveiled AlphaGo, serious students of the game questioned whether it was skillful enough to challenge the world's best players. When it topped Lee Sedol, many lamented that a machine had eclipsed the powers of humanity, somehow mimicking the intuition required to play this enormously complex game. But as the DeepMind team continued to improve this surprisingly powerful system, the top Go players couldn't get enough of its unique and sometimes transcendent style of play. They've repeatedly called on DeepMind to release the countless games that AlphaGo has played behind closed doors, seeing the machine as a window into the future of Go.

"AlphaGo extends the horizon of Go games," Chinese grandmaster Lian Xiao said this week after playing a match alongside the machine. "It brings more imagination."

As AlphaGo completes its year-and-a-half arc from Go novice to the grandest of grandmasters, DeepMind is giving Go players at least some of what they want. Today, during the press conference following the game, Hassabis and DeepMind announced they will publicly release 50 games AlphaGo played against itself inside the vast data centers that underpin Google's online empire. "Some of the great games in history may be played on these servers," Hassabis says.

Quite literally, AlphaGo learns to master the mysteries of Go by playing game after game against itself. Though it typically does this under strict time limits---seconds or milliseconds for each move---DeepMind has released games that played out over several hours, much like professional matches. "These are beautiful games, with moves no one has seen," says Fan Hui, the European Go champion who helped train AlphaGo.

But even as DeepMind made the announcement, some Go players complained that those 50 games were a mere pittance, considering that the machine has played hundreds of thousands of games against itself over the past several months, if not millions. They believe they have so much more to learn from the machine they once viewed with so much skepticism and fear.

The evolution of AlphaGo is just one demonstration of the way artificial intelligence can not only replace human skills but advance them. So many of the top players have changed their play after watching AlphaGo, particularly since a new version of the machine won 60 online matches under an online pseudonym in January. This week in Wuzhen, Ke Jie opened the first game with the rather unusual strategy, "3-3 point," that AlphaGo used time and again during those 60 matches.

Hassabis and other researchers believe that AI will soon have a similar impact in the worlds of healthcare and scientific research, pushing humans to places they couldn't travel on their own. And that's no idle prediction. Underpinned by technologies that are already reinventing the world's most popular internet services while rapidly pushing into driverless cars and robotics, AlphaGo is very much a proxy for the near future of AI as a whole.

According to Hassabis and lead DeepMind researcher David Silver, the new incarnation of the machine is something the DeepMind team can apply to a myriad of tasks beyond the game of Go. After the match in China, DeepMind is disbanding the team that worked on the game, freeing top researchers like Silver and Thore Graepel to spend their time working on the rest of AI's future.