I’d like to open this post by admitting I am not a Go master. I’ve played a few times, watch Hikaru no GO when nothing else was on. But that’s about it. However, I don’t need to be an expert at the game to point out the flaw in some of the press coverage. I suspect actual AI researchers already know what I mean.
The first thing to remember is that AlphaGo is a deep-learning program built on a neural network. What that means is that rather than an artificial intelligence program, AlphaGo is an artificial learning program. Public perception of AI is still focused on artificial intelligence, but the field has now expanded to cover many related or tangential or component areas of study. AlphaGo also has some form of reasoning ability. But this ability is solely related to Go. You cannot generalize it’s algorithms to other tasks. In fact, DeepMind even admits there are better programs out there to play Chess. Chess and Go are both “perfect information”(PI) games. You can if you so choose know everything about a given game of Chess or Go by looking at the board. You know all the rules and the position of all the pieces. PI games are a very popular area of AI research, because programs can do a lot with them. The information can be reduced to a very small set of states and rules, which is ideal for computers to excel at. The trick of course is to teach the computer the best set of tactics for taking those rules and the initial state of the game, and trading states with another player to get to the win state. And yet, even in two PI games, the best AI solution to a player capable of competing with the best of humans is different for each game.
I like to call this specific intelligence, although the more popular terms are weak AI or narrow AI, a kind of non-sentient intelligence focused on solving one task or a narrow range of tasks. But even that is a bit of a misnomer. After all, the machines aren’t truly smart, just impressively programmed dumb machines.
However, a learning program like AlphaGo comes a bit closer to true intelligence(though not sentience) by being able to take the initially programmed rules and knowledge and extrapolate from them on its own to do things it wasn’t explicitly hard-coded to do by the programmers. It’s incredibly impressive. But it’s not “AI” in the way most layfolk think of it. It’s not general intelligence, even a crude version. It’s a very sophisticated piece of specific intelligence.
But there’s a second flaw in the coverage. Besides the great deal of mystique that’s built up around Go, which isn’t an issue of AI, although some of it is misplaced–for example, another lifeform does not “almost certainly play Go” whereas Chess is too human specific–there’s the issue that even as a powerful example of narrow AI, AlphaGo does not–as stated by some professional players–“play go just like a human but better”. There has been much talk of its unorthodox tactics, or its algorithm’s focus on win-rate over all else. Some have even said it made moves “only God could have made”–a common expression of a perfect move.
But the real truth is this: much like how genetic code, a style of coding in which a computer is given basic building blocks of code and tasked to mix them up until it finds a closer to optimal solution, AlphaGo has no idea it is playing go. As far as AlphaGo knows, it’s just trading ones and zeroes around until it finds the desired sequence. The ways in which a human player attempts to reach the winning board position are inherently different than the way a computer does, because they aren’t really pursuing the same goal.
We’re not particularly closer to strong or general AI than we were before. Go isn’t truly so different from any other PI game. AlphaGO has not learned intuition. It’s merely played millions of games of Go subtly adjusting the value it places on a given set of stone positions on the board as it goes until more and more the win-rate increases to the point it wins the game. Although the process is superficially similar to the way a human learns the game, the lack of framing devices such as vision used by humans has taught it to value entirely different things, and unlike a human, a computer has a perfect memory to go with the perfect information, and it is incapable of making an error.
After that, we can consider the psychological warfare aspect of multi-player games. AlphaGo may be able to beat anyone Lee Se-dol could, but it cannot judge its opponents experience and thus alter its strategy to beat that player faster or more elegantly. Instead, it will always play the same way every time, and react no differently to a master making three opening moves than to a novice making the same. But where a human might see those moves and be able to make a variety of plays depending on their intuition of the players skill or likely next move, AlphaGo will continue to inexorably play exactly the move that will have the highest chance of victory against any and all players, rather than the one with the highest chance of victory against a specific individual.
Atsiko's Chimney 