Attendance: 1

Frisco Del Rosario writes about chess960, women's basketball, minor league baseball, unsupported collectible card games, lettering in comic books, and Golden Age movies

AlphaGo makes me think I went in the wrong direction 10/21/2017

I like to think I was born at the perfect time. I was 9 in 1972, when Fischer won the world chess championship, giving American chess a jolt. I was 21 when Apple’s ‘1984’ commercial aired, and 34 when Deep Blue beat Kasparov in an exhibition match. That is, I’m old enough to remember Fischer as a great chessplayer — whereas younger folk might tend to remember Fischer as a crackpot — and young enough to relate to computers better than my parents did.

But when I read about AlphaGo Zero — Google’s latest iteration of a Go-playing program that taught itself to crush the world champion — I think: damn, I picked the wrong game.

Most people who excel at a game of skill have tried the others, and spent enough time to get good enough to understand that they’re in the right place. Some get really good at more than one — in my circle of friends, there’s Elliott Winslow, an international chess master and a world champion backgammon player. Jennifer Shahade won the US women’s chess championship twice, but now she’s living by poker.

I’ve never met Irina Levitina, who’s won national titles at bridge and chess. That’s where I wanted to be when I was in my 20s, but didn’t get so good at either — my friends might say it’s because I spent a year mastering the arcade game Centipede, and a couple of years in the late ’90s at the card game NetRunner, which was a beautiful game that didn’t appeal to enough people to last.

I gave Go a try around the turn of the century. It was fascinating in that wow-I-will-never-get-the-hang-of-this way. It’s so difficult — I can’t imagine how more than a few masters actually enjoy Go. This is one of the things about chess — if you don’t have a master-level understanding, you can’t appreciate games played at a high level. At Go, that seems far out of reach, even for the best players at the local Go club.

AlphaGo makes moves that world-class players describe as “alien”, beyond our human understanding. To me, this sounds like a level beyond the best computer chess programs, which play moves that world-class players describe as “human”. The best chess computers perform at 3300 strength, when the world champion is just above 2800, and at that level, 200 points is greatly significant.

If we applied an Elo-like rating formula to AlphaGo, where is it? The best Go player in the world spends all his time these days analyzing AlphaGo games. This is another indication that Go is so much different from chess — looking at games played between the best chess computers is dreadfully dull, because there’s no scope for creative play in computer chess; the kind of risky, sacrificial plays that make human chess interesting doesn’t happen in top-flight computer chess. For the best Go player to find AlphaGo games interesting enough for immersion, that’s truly something extraordinary.

Now there’s AlphaGo Zero, which has removed the human element. It learned Go from scratch, and continued learning from games played against itself.

There are probably people who are genuinely afraid of this development, as if it’s the first real step on the way to SkyNet from the Terminator movies, the artificially-intelligent agents that determined it didn’t need humans at all, for any reason.

That AlphaGo has come so far makes me think I should’ve determined to reach some level of competence at Go so that I could teach it through storytelling. Go teachers and chess teachers operate in the same way: they describe patterns in memorable ways — lions and dragons take shape on both boards — and describe the game’s underlying logic.

But the impression I get from reading about the Go community is that when Go teachers “tell stories”, the students, as a whole, listen. Chessplayers suffer from weird self-delusion: They believe they’re so much better than they are that they seek teaching that exceeds their grasp. Study of the opening at chess is for masters, because the early game matters in master play, where neither side is doing stupid things immediately. Average chessplayers want to believe they’ve reached this level, and want to study openings.

The great teacher Mark Dvoretsky was famous for teaching not-yet-grandmasters to get to the next level. It’s shocking to me how many chessplayers recommend Dvoretsky to other players, as though they believe they learned enough from Dvoretsky to be qualified to share his teachings with others. Average chessplayers should be studying Cecil Purdy, who said average players should be trying to reduce the number of stupid things they do.

The impression I get from Go players is that they better appreciate how incredibly difficult their game is, and pay an appropriate amount of respect to it by striving to learn at the right level. I’ve almost always had just the opposite feeling about chessplayers.

The AI That Has Nothing to Learn From Humans

Categories chess

Leave a Reply

Your email address will not be published. Required fields are marked *