Google's AlphaGo AI beat the world's best human players earlier this year in repeated contests that saw it facing near insurmountable odds to your average human. AlphaGo took it in stride though and even though games were often close, it never missed a step. Although Google promised to retire its AlphaGo AI after such a victory, it has been using it to help train another and the new one has blown passed all expectations.
The traditional method for training an AI involved thousands of hours of programming individual functions, but as any gamer knows, AIs created that way rarely pose much of a challenge. They're too cumbersome, can't fluidly react to nuance and are easily defeated by the powers of the human mind. What AlphaGo did so successfully, was to build upon initial training from humans and figure out its own methods to victory by playing itself.
The new version of Google's AI takes that a step further and has only ever played itself. While that might sound counterintuitive, it turned out to be a far better method of teaching. The AI has now beaten AlphaGo in over 100 games, quite conclusively and it's only been training itself for 40 days. How much better can it get in the weeks and months to come?
Of course the new "AlphaGo Zero" AI does work at a different pace to human minds. It's played millions of games against itself and has refined the process of play, even for something as complicated as Go, down to an exact science. It requires very little power compares to its predecessors, making it much more efficient in every way.
AIs like AlphaGo Zero could in the future, help us tackle ever more complicated problems, simply by tackling them over and over and analyzing the results.
Image source: Fcb981 at English Wikipedia
