Games●●●●●Difficulty 5 of 5

Did AlphaZero really beat the top chess engine fairly?

AlphaZero taught itself chess in hours and beat Stockfish, then grandmasters said the hardware wasn't comparable, and a rematch had to answer them.

▶ Start the story

Mostly yes, but only after a second try. The first match, announced in 2017, had AlphaZero beating Stockfish 8 over 100 games, with 28 wins, 0 losses and 72 draws. Critics attacked the set-up. Stockfish's own Tord Romstad criticized its settings as suboptimal and pointed out the version was a year old. Grandmaster Hikaru Nakamura said AlphaZero used a Google supercomputer while Stockfish was basically running on what would be his laptop.

28-0-72

wins, losses and draws in the 2017 match against Stockfish 8

DeepMind answered in the 2018 Science paper. It clarified that AlphaZero played on four TPUs and a 44-core CPU, not a supercomputer, and re-ran the match with Stockfish 9 dev under the same conditions as in the TCEC superfinal, with 3 hours plus 15 seconds per move. AlphaZero won a 1000-game match with 155 wins, 6 losses and 839 draws, and Stockfish needed 10-to-1 time odds to match it.

The argument about hardware did not fully go away: AlphaZero still ran on a much more powerful machine, with four TPUs in addition to 44 CPU cores. What is remarkable is how it played. It searched just 80,000 positions per second against 70 million for Stockfish, leaning on a neural network that was trained only by playing itself, with no opening books or endgame tables.

Two very different ways to play chess

AlphaZero

  • About 80,000 positions per second
  • Trained only by self-play
  • Neural network picks the promising lines

Stockfish

  • About 70 million positions per second
  • Played the 2018 match with endgame tablebases

Quiz me

0/3

  1. 1.How did AlphaZero make up for searching far fewer positions than Stockfish?
  2. 2.What did DeepMind change in the 2018 paper to answer the criticisms of the 2017 match?
  3. 3.What objection about the comparison remained even after the 2018 results?

Recap

AlphaZero beat Stockfish twice, but whether the hardware was comparable stayed debated.

💡 A trick to remember it · Fewer looks, wiser looks: a self-taught network looked at 1,000 times fewer positions and still won.

Surprising fact · It searched about 80,000 positions per second against Stockfish's 70 million.

Sources (1)

No source, no claim. Every fact in this lesson (20 claims) cites at least one of these.

  1. [1]AlphaZero · Wikipedia
More lessons in 🎲 Games (3) See all games lessons →

One more light on your map.

Get one lesson like this every day, about the things you love. Free, in two or five minutes.

Get the share card for this lesson ↗