Move 78

Move 78 is the 78th move of the fourth game between the South Korean Go professional Lee Sedol and DeepMind’s AlphaGo, played in Seoul on March 13, 2016. Lee had already lost the first three games and with them the match. Playing white, he placed a stone in the center of the board that split AlphaGo’s position – a “wedge” that commentators called brilliant. AlphaGo answered badly, its game fell apart, and it resigned. It was the only game Lee won, and Move 78 became the human counterpart to AlphaGo’s Move 37.

God’s touch

Lee thought long before playing it. The Chinese 9-dan professional Gu Li called it the “hand of god”; Wired reported that Go players dubbed it “God’s touch.” The commentator Michael Redmond said it had taken him by surprise and had probably surprised AlphaGo as well. DeepMind’s chief executive Demis Hassabis later said that AlphaGo had put the chance of a human playing the move at about one in ten thousand – the same odds it had assigned to its own Move 37 – and so had barely considered it. AlphaGo’s answer at 79 was the decisive error; according to Hassabis, its evaluation still showed about a 70 percent chance of winning, and the program noticed its mistake only about ten moves later. It then made several plays that professionals saw as obvious errors. Lee said afterward: “It is the single priceless win that I will not exchange for anything.”

Blind spots

Move 78 showed that even a superhuman system can fail in situations it rarely met in training, and that its failures can look strange to human eyes. Studying such failures, known as adversarial robustness, is an active field of research. In 2022 and 2023 researchers at the nonprofit FAR AI, MIT and the University of California, Berkeley, trained adversary programs against KataGo, an open-source Go engine stronger than AlphaGo was in 2016, and found that it misjudged large ring-shaped groups of stones. In early 2023 the amateur player Kellin Pelrine, a co-author, used the trick to win 14 of 15 games against a top program without computer help. Adversarial training has reduced such weaknesses but not removed them. The same pattern appears in language models, where unusual inputs can make a capable system behave badly (see Jailbreak) – one reason developers pay testers to look for exactly such moves (see Red Teaming).

Afterward

AlphaGo’s successors went further. In 2017 a stronger version beat the world’s top-ranked player, Ke Jie, 3–0, and DeepMind retired AlphaGo from competition. Lee Sedol retired from professional play in November 2019, saying that AI had become “an entity that cannot be defeated.” Move 78 remains one of the best-known human victories over a superhuman game AI.

Why it matters

Move 78 carries two lessons. Human creativity did not become worthless when machines surpassed people at Go; players learned from the programs and the programs’ rare failures in turn taught researchers about generalization. And robust AI systems have to be tested where they are least prepared – the cases nobody thought to include in training. Both lessons are central to AI safety today.

Written with Claude, made by Anthropic, which competes with Google DeepMind, the developer of AlphaGo; Google is an investor in Anthropic. Anthropic uses red teaming and adversarial testing on Claude.

Related

Move 37 · Red Teaming · Jailbreak · Generalization · AI Safety · Centaur · Deep Blue

Sources: Cade Metz: In Two Moves, AlphaGo and Lee Sedol Redefined the Future (Wired) (March 16, 2016) · Google DeepMind: AlphaGo · Cade Metz: Go Grandmaster Lee Sedol Grabs Consolation Win Against Google’s AI (Wired) (March 13, 2016) · David Ormerod: Lee Sedol defeats AlphaGo in masterful comeback – Game 4 (Go Game Guru, archived) (March 13, 2016) · Yoon Sung-won: Lee Se-dol shows AlphaGo beatable (The Korea Times) (March 14, 2016) · Tony Wang et al.: Adversarial Policies Beat Superhuman Go AIs (arXiv) (2022, revised 2023) · FAR AI: Even Superhuman Go AIs Have Surprising Failure Modes (July 14, 2023) · Tom Tseng et al.: Can Go AIs be adversarially robust? (arXiv) (2024) · Dan Robinson: Humans strike back at Go-playing AI systems (The Register) (February 20, 2023) · Demis Hassabis, David Silver: AlphaGo’s next move (Google DeepMind) (May 27, 2017) · James Vincent: Former Go champion beaten by DeepMind retires after declaring AI invincible (The Verge) (November 27, 2019)

Leave a Reply

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