Man vs. machine

I recently had time to catch up on the incredibly interesting battle between Google’s AlphaGo and Lee Sedol. AlphaGo, developed by DeepMind which was acquired by Google, developed a hyper sophisticated AI that specializes in the Chinese game Go. The technology behind AlphaGo is fascinating and anyone who is interested in AI or machine learning absolutely must read-up on the technologies and engineering that’s powering AlphaGo.

Back to the game however, if you haven’t seen it then spoiler alert: Lee Sedol has, as of midnight in Denmark, lost two to zero — that is, an AI for the first time has beaten not only a human in a complex game, but one of the best humans in the world at a incredibly complex game, twice in a row. They will play three more matches (no matter what) for a best three out of five.

I’m no Go player so at times it was difficult to follow or understand the significance of the plays during the game. However, what was significant to me was not the actual plays, but the commentary. Several times during the game the commentators characterized some of AlphaGo’s moves as everything from “interesting” to “very unexpected” to “what?”. Yet it was apparent although not always immediately that AlphaGo knew exactly what it was doing. Sedol would later say that “. . . there was not a moment in time where I felt that I was leading the game.” DeepMind CEO Demis Hassabis added to this saying “AlphaGo played some quite surprising and quite beautiful moves . . . which was pretty amazing to see.” The commentary revealed to me two primary advantages that AlphaGo has that are both a true testament to the power of AI, and downright scary.

Games are largely made up of two components: computational and emotional. A good Go player not only has to have the mental capacity to calculate a multitude of scenarios, but they must also be able to play the mind game with their opponent. That is, certain moves can have psychological effects on your opponent that can give you an edge. Likewise, a player can build some form of intuition and prediction based on the signals from the opposing player. AlphaGo has an absolute advantage in computational power, that much is obvious. But it also has an emotional advantage as well — that is, without emotion, the AI is unaffected by things such as time-pressure or signals, it just coldly calculates optimal plays.

Another key and incredibly intriguing advantage is how the AI simply perceives the board. Sedol and every Go player on Earth learned to play by studying the humans and the human developed strategies for Go. Every player has their own twist, or improvises of course, but these core learnings all fit within frameworks or theories that are the equivalent of the Standard Model for Go. The comments from Sedol and the commentators indicate that AlphaGo’s moves were often inconsistent with what humans know is the “right” way of playing — those moves wouldn’t be played because it’s not how humans perceive the game. This means that AlphaGo must have an entirely new, or radically different way of seeing the board and the interactions between the pieces.

When you put these two things together it makes me feel like we have possibly witnessed the closest thing to what would occur if a human played Go with a highly intelligent alien species — one that was not bound by the preconceptions and tainted worldview of humans.

It’s going to be interesting to see how the third match will play out, and more importantly, see if human weaknesses, like being repeatedly beaten, begin to show their wear on Sedol.