AlphaGo, the board-game-playing AI built by Google’s DeepMind subsidiary, just got a huge update, making it smarter — and potentially creepier — than ever before. In a paper published in Nature, the company reveals that the newest version of the AI, called AlphaGo Zero, requires no human training in order to make itself better, and it’s now so good that it can demolish even the past versions of itself.
The company approached AlphaGo Zero development much differently than that of its predecessors. Rather than making the AI smarter based on known, human-developed strategies, the team instead pitted the software against itself, with no prior advanced training. As the AI beat itself over and over, it continually updated its own knowledge of the game, it got better and better.
The AI taught itself to play the game at a higher level than any human player, running on marathon sessions where it played against itself for several days at a time. After three full days of constant playtime, Zero was capable of crushing the version of AlphaGo that already bested the human world champion. In fact, the battle wasn’t even close, as Zero took 100 wins in a row and didn’t lose a single match.
“By not using human data — by not using human expertise in any fashion — we’ve actually removed the constraints of human knowledge,” AlphaGo Zero’s lead programmer, David Silver, explained during a press conference. “It’s therefore able to create knowledge itself from first principles.”
In short, by not trying to mimic the best human Go players, AlphaGo Zero actually eliminated any of the biases or oversights Go players are hampered by, thereby creating a more pure strategy. In fact, now that Zero has shown just how good it is, it might be able to act as a guide for players who want to up their own game to the level of the computer.