swirl swirl

Thoughts on AlphaGo: We've Lost Something

This week, Lee Sedol, the famous 9 dan professional Go player, was beaten three games to zero by AlphaGo the Go-playing computer. This surprised everyone, because computers kind of sucked at Go until about 3 months ago, and now suddenly it seems we have a computer that's nigh-unbeatable.

I'm a programmer, and I also play Go (badly). One would think that my feelings would be mixed, something like "Yay programmers! Too bad, Go players!" But in fact I find that my feelings are all sadness. I feel like we've lost something.

Of course humanity has been in the business of obsoleting itself for a long time. We used to have items made by master craftsmen, and now we mass-produce those things and the craftsmen are gone, or hobbyists. Machines extend all of our senses: telescopes, microscopes, machines to lift heavy objects, machines to move fast. Computers can do more arithmetic problems in a second than I've probably done in my lifetime.

Somehow none of that bothers me too much. Maybe I'm just used to it. I don't feel particularly bad about the shoe-smith that might've lived down the street being replaced by a shoe factory. But maybe I would feel bad if I'd lived in the time of shoe-smithery, and witnessed the time someone put into training to make shoes, the care and love they put into creating a pair. I have lived in a time when Go something that humans were specially good at. We've all lived in that world for the past 2,000+ years. And suddenly in the blink of an eye, we're no longer nearly as special.

The loss of Go as a human domain feels especially bad because it's not just a tool we make or a problem we solve. It's a game, and we care about it because it's fun and it evokes strong feelings in us. Moves are described as feeling good or being "vulgar". You make shapes on the board, and those shapes are beautiful or ugly depending on how they might affect the outcome of the game. There's a list of Go proverbs a mile long that've been repeated since long before I was born, and many of them feel like they apply to life as much as to the game. Go players' personalities shine through their playstyles. You can be aggressive or patient, risky or solid. Just listen to this quote by Tim Brent on the Sensei's Library Wiki:

The aesthetics of Go are almost as important as the play in some ways.

The harmonisation of the board and pieces, the materials used, etc., can be for some almost as important as the game (i.e. never use a yuki stone set on a cheap wooden board or don't use thin plastic stones on a table kaya board, etc.).

Also, as many a beginner's book points out, there is aesthetic harmony in the contrasts of wood and stones, line and circle, black and white.

How true!

But a computer surely doesn't appreciate any of this. It has no personality to put into its games. Its Go is reduced to a maximization problem. And I start to get an inkling that maybe the emotional content that humans put into the game is actually a weakness rather than a strength. The computer doesn't care how ugly its moves look, it only cares that in 100 moves it'll be slightly more likely to be ahead by a few more points.

The computer also didn't work its whole life to become good. Some very smart people came up with an algorithm, dumped data into a machine, and let it run for a while. You can probably make it better just by adding more GPUs or more training time. It feels like cheating! Up to now, people literally trained their whole lives in Go schools to become 9 dan, and someday soon you'll probably be able to download a 9 dan and run it on your cell phone.

I don't hate computers. I'm a programmer; I find good code to be very pleasing. The way AlphaGo works, at least as much as I understand of it, is pretty freaking awesome. But I don't know if I want computers stepping over a certain line. I want books to be written by people, not auto-generated by a server farm. I want paintings that an artist uses to convey an idea from his brain to mine; this is not the same!

Regardless of what anyone might want though, the ship has sailed. There's no arguing with reality. It doesn't matter how much time you spent getting good at sword-fighting, how cool it looks, how much history there is in the martial art, how much life philosophy is wound up in it... you'll still lose in a fight against a gun.

In this New Yorker article, Ben Lockhart says "A dolphin swims faster than Michael Phelps, but we still want to see how fast he can go", and I agree with that. Go isn't ruined just because a machine can do it better. Go is still a very hard thing to become good at, and people who were masters of it last week are still masters of it today. Lee Sedol sure impressed me by playing for so many hours against a building full of supercomputers and almost holding his own, all things considered.

Many pros even think that AlphaGo will improve the game of Go overall, by teaching people how to get better. Maybe there are things we've been doing wrong for hundreds of years. The computer will help people improve by giving us a great and (hopefully, eventually) cheap study tool, just like what happened for chess after the advent of chess computers. So there's that.

I still think Go is a cool game, and I'll probably still enjoy reading reviews of famous games, and playing it (badly) myself. I'll still get a kick out of a clever move or pretty shape. But I can't help but feel a little sad.

One comment

  1. From a programmers perspective I am definitely excited by the outcome of these matches but I don't think it means a lot less for the future of AI than some people think. AI will always be a tricky problem, and Neural Networks have been a popular solution method to the problems involved with it since it became a big research topic again.

    As you mentioned it's often a process of throwing more resources at the problem and that's exactly what they did with AlphaGo. Certainly they had some very clever people refining the heuristic, but at the end of the day the output of this program will look like total gibberish to an observer and it will be difficult to change or expand upon it.

    I expect that we will hit another ceiling in AI progress as we look into fields that to me are the most interesting and complex in AI such as Natural Language Processing and Developmental Robotics.

    On your note of painting though, I think this recent research is far more bizarre to see about the implications of AI in human domains.
    Deep Learning Painting

Leave a comment — be nice !

All fields optional (except Comment, duh). Markdown allowed.