ptiChess

AlphaZero v Stockfish, 2017: the knight left on h6

In December 2017 Google's DeepMind announced AlphaZero, a program that had learned chess in about four hours by playing against itself, with no opening book and no human games. In a closed match against Stockfish 8, then one of the strongest engines in the world, it scored 28 wins and 72 draws without a loss, and DeepMind published ten of the games. This is one of the most admired: AlphaZero gives up a knight for no immediate return and wins over 56 moves.

a8b7c6d5e4f3g2h1
White to move after 18...g5. AlphaZero leaves its knight on h6 to be taken.

Find the move. Find the move AlphaZero played at move 39.

Find the move

The whole game

a8b7c6d5e4f3g2h1
AlphaZero v Stockfish 8, 2017. Step through all 56 moves; Black resigned.

The opening was a Queen's Indian. AlphaZero's 7.d5 and 8.Nh4 offered a pawn to open lines toward Black's king, and its knight went to f5 and then h6, where it stayed until move 19.

What Stockfish makes of it

We had today's Stockfish, far stronger than the Stockfish 8 that lost this game, check every move. Its verdict is that AlphaZero's play was sound and that Stockfish 8 lost without one big blunder: the position slid away a little at a time.

15.Nh6+ to 19.Re1: the knight is given up

a8b7c6d5e4f3g2h1
15.Nh6+ Kg7 16.bxc3 Bc8 17.Qf4 Qd6 18.Qa4 g5 19.Re1 Kxh6: the knight is taken, and White plays on a piece down.

After 19...Kxh6 White is a knight and a pawn down. Even so, Stockfish rates White about half a pawn better: Black's king has been pulled into the open, and the bishop on e7 and the pieces around it are tied up. 20.h4 was the only good follow-up; 20.Rxe7, taking material back at once, would have left Black better.

Moves 24 to 31: Black has to keep finding only moves

a8b7c6d5e4f3g2h1
24.hxg5+ fxg5 25.Qh4+ Kg6 26.Qh1 Kg7 27.Be4 Bg6 28.Bxg6 hxg6 29.Qh3 Bf6 30.Kg2 Qxa2 31.Rh1 Qg8: White's queen and rook line up on the h-file.

From here Stockfish 8 had to find one move at a time. 24...fxg5, 25...Kg6, 28...hxg6, 29...Bf6 and 31...Qg8 were each the only move that kept Black's position in one piece; anything else would have left White three pawns or more better. Stockfish 8 found them all: today's Stockfish rates White about one pawn better after 26.Qh1, and still about one after 31...Qg8. 26.Qh1, the queen tucked into the corner behind the h-file, is the move people remember.

34 to 39: the material comes back

a8b7c6d5e4f3g2h1
34.Rxd4 Rd8 35.Rxd8 Qxd8 36.Qe6 Nd7 37.Rd1 Nc5 38.Rxd8 Nxe6 39.Rxa8: the queens come off, and White has the better ending.

In the exchanges from move 34, AlphaZero again had to be exact: 34.Rxd4, 36.Qe6, 37.Rd1, 38.Rxd8 and 39.Rxa8 were each the only move that kept White's advantage, and several alternatives would have thrown it away completely. After 39.Rxa8 White has a rook against a knight and two pawns, and Stockfish rates White about three and a half pawns better: Black's g-pawns are doubled and the queenside pawns are a target for the rook. The rest was technique; Stockfish 8 resigned after 56.Kd4.

Was it a fair fight?

The match drew criticism at the time. Hikaru Nakamura called it "dishonest", pointing out that Stockfish played without its opening book under conditions DeepMind chose; Larry Kaufman also questioned whether the result would hold on ordinary computers. Few doubted AlphaZero was very strong, but the margin was disputed. This game doesn't settle the question either way: the opening was ordinary, and Stockfish 8's mistakes, as today's Stockfish sees them, came in the long middlegame.

How we checked

Try it. Real positions, set at your level, every answer checked against the board. No account needed to start.

Start training

More famous games

Ways to train