Castling in the cloudIn the middle of last week Google DeepMind's newly minted chess-playing AI,
AlphaZero, wiped the floor with the former champion superhuman chess engine,
Stockfish.
Or, at least one partially under-resourced instance of Stockfish.
www.chess.com/news/view/google-s-alphazero-destroys-stockfish-in-100-game-matchHo-hum, right? They built a stronger computer, as expected.
No, not only. They applied a naive but powerful neural net with an incomparably
stronger generalized learning algorithm.
Nobody told AlphaZero about the history of chess, or gave it any advice,
or 100,000 championship games to remember, or anything. Just the setup,
the object, and the rulebook.
Then they left it alone to play by itself for a reported 24 hours.
It was also working on Shogi and Go during the same training period,
and those are harder for computers to play, so I'm not quite clear about
how much time the neural net ultimately spent with each.
But it was like, hello little AI, your dad and I are going to see a show in
the city, and we'll be back tomorrow for lunch. Here are some games.
All your rook are belong to us — Do you think she'll be OK alone?
— Of course, she's very resourceful. You worry too much.
— But what if she decides to wipe out humanity?
— Well, that's why I gave her chess to learn, isn't it?
To be fair, on Google's hardware baby A.Z could experiment for umpty-eleven
lifetimes in an afternoon. And, at the rate of 1,000 games per year it did
teach itself for 700 years, but by the 300th year it was looking like much the
strongest player ever seen.
Against Stockfish, observers said it was developing games that its opponent
did not seem to understand until too late. As usual in chess tournaments,
Stockfish frequently fought to a draw. But it won zero games out of 100,
and got surprised a lot. Cruising, cruising, cruising, cru... holy shit.
A.Z didn't win by overwhelming force. In tournament play, for one minute it
would evaluate 80,000 options per second (against Stockfish's 70 million/sec)
and then move. By a couple of orders of magnitude, it's more selective about
which alternatives to consider.
The implication is that it's deciding in a more human-like way. Maybe something
like, "There are 24 things I could do here, and 9 of them aren't so stupid, and 5
of those seem to have more potential than the rest, so 30 clicks down the road...
which path has the best chance of getting me home?
In a rematch, the human Stockfish team would need to retool its prodigy.
Tweak the strategy, give it new reference material, analyze the opponent,
massively upgrade the hardware, and equip Stockfish to play harder.
No one has said if AlphaZero actually *can* be hyped by such techniques,
but it has already had a chance to learn how to deal with Stockfish by playing
against it. That's what makes the two systems so different.
This is Gary Kasparov:
"…obviously the implications are wonderful far beyond chess and other games.
The ability of a machine to replicate and surpass centuries of human knowledge
in complex closed systems is a world-changing tool."Hey, buddy. I appreciate being called a tool about as much as you do, okay?
So... the discovery of the meaning of life is not that far off, although we'll
still be too dumb to understand it when we're told. Some things you can't fix.
Regardless, at least Google will know.