Move 37

Cristóbal Esteban
4 min readMar 14, 2016

As you probably already know, DeepMind’s A.I. named AlphaGo is now winning 3–1 against the top Go player of the last decade, Lee Sedol (update: the final result was 4–1). This is considered a great A.I. achievement because there are more legal Go positions than atoms in the universe. Therefore, a machine is not capable of computing all possible plays and pick the best one, but it has to develop some sort of intuition and strategical thinking in order to make the best move each time.

Lee Sedol said before the matches that either he would win 5–0 or 4–1. Imagine how it must feel to be a legend of Go and suddenly start playing against an opponent that beats you three times in a row. Moreover it plays moves that you cannot even understand.

In the second match between Lee Sedol and AlphaGo, we saw an historic moment when in turn number 37 the A.I. played a creative move that surprised all experts. Such a move was not about replicating other moves the machine had seen after watching many games. It was a move never played before by a strong player, and even knowing that, AlphaGo decided it was the best move to make.

Take a look at the next 30 seconds or so of this video:

Lee Sedol left the room for a couple of minutes after that move. The reaction of the commentators was also priceless: the move would have been classified as a mistake in any other situation, but this time no one dared to say if it was a mistake or if AlphaGo’s move was simply beyond the current human understanding of the game.

And that is precisely the issue when you play a strategy game against someone smarter than you are. The smarter player can predict the moves of the weaker player (e.g. “he is trying to reach this and that, and therefore he is going to play this move because he didn’t realize this thing; thus, I am going to beat him this way”). However, the only prediction the weaker player can make is that he is going to lose. He has no idea how it will happen, which move his or her opponent will make next or which strategy the other player is following. All he knows is that he is going to lose.

That is the reason why Elon Musk (and others) is concerned about the future role of A.I. and made his famous quote: “With artificial intelligence we’re summoning the demon.” Because when you play against a system that is smarter than you are, all you can predict is that you are going to lose.

However I think that, rather than an independent intelligent agent, our current implementations of A.I. are more like a way of extending our own intelligence, which seems a much more optimistic and interesting way of framing AlphaGo’s victory. AlphaGo doesn’t even have the will to play Go. Some humans have the will to win at Go and they used AlphaGo to extend their intelligence and beat the best player. Also, in order to travel to Seoul they used a plane. These are all tools that make us better.

Many people complain about the fact that we don’t even know why AlphaGo makes the moves it does. However our current understanding of how our own brains work is also very limited. And we make decisions every day with this “black box” we have on top of our shoulders, often without really knowing where our ideas came from. Indeed some neuroscientists claim that many times our brains make decisions unconsciously and then we make up a reason that explains why we made such decision, but that is for another post. My point is that having a smart artificial black box that extends our intelligence is a huge thing and not that different from using our own brains.

Michael Redmond, who is a professional Go player and commentator of the game, made a very interesting point when he said that it is very useful to have a new super strong player that brings new creative moves and strategies, because this will allow humans to study AlphaGo’s games and improve their own abilities! It reminds me of when DeepMind’s Atari player taught to its programmers this strategy to play the breakthrough game:

It has to be said that creating A.I. systems for gaming purposes is easier than creating them for other scenarios such as healthcare. The reason is that in order to train A.I. gaming systems you can generate your own data by making the machine play against itself or other machines. Indeed, one of the reasons why AlphaGo is much better than humans is that it has played way many more games than any human could play in his or her lifetime. In healthcare of course you cannot just make tons of decisions to see what happens and learn from such experiences, but you have to learn based solely on existing data.

However, many of us, including the DeepMind guys, are working on ways to bring A.I. to healthcare. And hopefully we will soon have “AlphaDoctors” in our laboratories, clinics and hospitals making awesome moves and showing us new ways of curing diseases.

Of course we will also use this intelligence in many other domains, as in marketing, where we will have “AlphaMarketers” helping us with the famous 4Ps of marketing, and “AlphaSalesmen”.

Fasten your seat belts and enjoy the flight.

— — — — — — — — — — — — — — — — — — — -