I have been waking up early in the morning to watch the Go games live on YouTube between Lee Sedol, one of the best human players in history, and AlphaGo, DeepMind's (Google) artificial intelligence. The 4 to 1 result in favor of the machine has left the academic world, including me, in a state of shock.
We all knew this day would come, just as I remembered Watson's victory in Jeopardy, but the general consensus was that it was at least a decade away from beating a human at Go. We were completely wrong.
Why Go is not like Chess
When IBM Deep Blue defeated Kasparov in 1997, it did so with a classic artificial intelligence technique that has little of intelligence: brute force. It evaluated millions of positions per second using algorithms like Minimax with alpha-beta pruning. But Go has a 19x19 board and rules so simple that the number of possible legal positions is greater than the number of atoms in the observable universe ($10^{170}$). Brute force is mathematically impossible. Humans play Go using something very elusive called "intuition".
DeepMind's technical brilliance has been to combine Monte Carlo tree search (MCTS) with Deep Reinforcement Learning.
Instead of evaluating all moves, AlphaGo uses two distinct neural networks trained with GPUs: 1. Policy Network: It looks at the board and narrows down the options, predicting which moves an expert human would make to avoid exploring useless branches. 2. Value Network: It evaluates the resulting position not by calculating until the end of the game, but by mathematically estimating the probability of winning from that state.
The most terrifying thing is how they have trained it. After learning from human games, AlphaGo started playing millions of games against itself (self-play). It improved its own neural networks based on pure rewards: winning was a 1, losing a -1.
Reflection: Move 37 and artificial creativity
In the second game, something happened that I will never forget. The famous "Move 37". AlphaGo placed a stone on the fifth line of the board at a very early stage of the game. All the expert commentators said live that it was a huge mistake, a "bug" in the machine, because no human would play there. Dozens of moves later, that single stone turned out to be the master pillar of a global strategy that crushed Lee Sedol.
The machine did not just calculate; the machine was creative. It broke centuries of human theory about Go. If an AI can generate new knowledge and strategic intuition of this level by combining tensors and GPUs, the question is no longer whether they will surpass doctors or drivers. The question is: how long before an AI optimizes our architectures and applications?