One day, you might look back at the moment when AlphaGo trounced the world's best Go player and feel an acute sense of guilt. Guilt that you failed to see the warnings signs and stop the robot uprising that’s soon to come.
Now, a new AI by Google's DeepMind research group has taught itself, using only the basic rules of the game, to beat AlphaGo at its own game. While we wouldn’t go as far as calling this a “god-like” AI, we would say that it’s a seriously impressive, and perhaps daunting, milestone.
AlphaGo learned how to play the 3,000-year-old board game by analyzing matches played between humans in the past. The latest incarnation is essentially self-taught, and no human data or input was necessary after its initial construction. That’s why it’s been dubbed “AlphaGo Zero” – it uses no human knowledge at all.
Yes, AlphaGo was impressive. AlphaGo Zero, however, spent just three days training itself to become a world-beater; it then pitted itself against its predecessor and won 100 games to nil. In just 40 days, AlphaGo Zero surpasses all other versions of the AI, and arguably becomes the best Go player in history. It took this AI just 72 hours to transcend three millennia of human knowledge in this respect.
This may now make it sound like humans are obsolete in this regard, but in a blog post, and in an accompanying Nature article, the designers explain why this was never just about playing an ancient game.
This technique, of teaching itself how to handle a new situation, is for some purposes better than anything any of the previous incarnations managed, because “it is no longer constrained by the limits of human knowledge.”
“If similar techniques can be applied to other structured problems, such as protein folding, reducing energy consumption or searching for revolutionary new materials, the resulting breakthroughs have the potential to positively impact society,” they add.
This shines a light on what most are considering to be the near-future fate of AI; programs that augment and amplify human roles, rather than replace them, for the most part. From helping to confirm diagnoses of diseases to corroborating or disagreeing with the latest climate model, this type of pattern-recognition AI is going to be the next big thing.
If you’re a pessimist, you’ll see this latest development as another step towards our machine overlords. If you’re an optimist, this marks an exhilarating leap forward for humankind.
1
