We Used AI to Analyze Sai's Games: How Does Shusaku Compare with Today's Pros? | Hikaru no Go

In Hikaru no Go, a mystery player called "sai" appears online, beats strong players from all over the world, and everyone asks, "Who is sai?" Today, anyone would say, "That's AI, right?" But back then nobody did: computers were terrible at Go. So flip it around: if Sai showed up in online Go today, could he beat AI? Sai's games are Honinbo Shusaku's, so we had a modern Go AI, KataGo, score them! We're Miko, a dog-eared apprentice shrine maiden, and Tsuno, a dragon-girl researcher of otherworld science! From Deep Blue to Monte Carlo tree search, AlphaGo and AlphaGo Zero: how computers learned Go, one algorithm at a time. Deep Blue beat Kasparov at chess. Why not Go? Minimax, alpha-beta pruning, and more positions than atoms in the universe "It may take a hundred years": a 1997 prediction "Just play random moves to the end?" Monte Carlo tree search AlphaGo's two neural networks: how do you build an "eye" for good moves? Lee Sedol, game 4, move 78: was the "Divine Move" really right? "Master", the mystery account of 2016 AlphaGo Zero and the loop of reinforcement learning Shusaku vs. pros then and now (40 games each), scored by KataGo Could Sai beat today's AI? Watch to find out! Is there a great game from Hikaru no Go you'd like an AI to look at? Tell us in the comments! [Chapters] 00:00 Opening: today, sai would be accused of using AI!? 01:05 Chapter 1: Chess Fell. What About Go? (more positions than atoms in the universe) 03:11 Chapter 2: Playing Go with Dice (we played 10,000 random games) 04:36 Chapter 3: A Machine with Intuition (was the Divine Move real?) 07:21 Chapter 4: A Mystery Player Online (60 straight wins over top pros. Who?) 10:14 Chapter 5: Could Sai Win? (Shusaku vs. pros, then and now) [How We Tested] KataGo v1.18.1, network kata1-b18c384nbt, on one PC. Per move: KataGo's first choice or not, and points lost. 100 visits per position; decided positions (10+ point lead) left out. 40 even games each, at random: Honinbo Shusaku (1846–1862) and Japanese Kisei, Meijin and Honinbo title matches, 1998–2003 and 2020–2026. No komi in Shusaku's day: points lost don't depend on komi, win rates do, so we compare points. First 60 moves, points lost (median per player per game): 7.0 (1998–2003) vs 4.1 (today); from move 61, 0.33 vs 0.49 a move. The 70-year study: Shin et al., PNAS 2023. The ear-reddening move (vs Inoue Gen'an Inseki, 1846, move 127): 20,000 visits. Move 78: later reviews by pros and AIs (L10 refutes it; An Younggil 8p and others, GoGameGuru), checked with our KataGo. "Hand of God": Gu Li 9p; "1 in 10,000": DeepMind. Monte Carlo demo: our own program, one 9x9 position picked for its clear answer, 10,000 random games per move. [References] Silver et al., Nature 529, 484 (2016); Nature 550, 354 (2017) D. J. Wu, arXiv:1902.10565 (2019) (KataGo) R. Coulom, Computers and Games (2006) (Monte Carlo tree search) Campbell, Hoane, Hsu, "Deep Blue", Artificial Intelligence 134, 57 (2002) J. Tromp, "The number of legal Go positions" (2016) G. Johnson, The New York Times, July 29, 1997 (Piet Hut) M. Müller, "Two Representative Computer Go Games" (1998) GoGameGuru, AlphaGo vs Lee Sedol game 4 review (2016) Shin et al., PNAS 120, e2214840120 (2023) Game records: Andries Brouwer's collection [Channel] Miko & Tsuno's Otherworld Science (@MikoTsunoLabEN): anime, manga and game worlds, and everyday mysteries, tested seriously (and for fun) with real science and engineering. Subscribe to catch our next experiment, and like and comment if you enjoyed it! 🇯🇵 The original Japanese version: @MikoTsunoLab [Credits] Go AI: KataGo (David J. Wu) / Characters: channel originals (AI-generated illustrations) / Voices (AI): Google Gemini 3.8 Neural TTS / Music: なぐもりずの音楽室 (東方の薬師, Strategy Meeting, Dokubutsu Mania, Jardin Botanique, シュットフェルト地下遺構, ご隠居魔女の憂鬱, Butterkartoffeln, 木漏れ日のロンド, Strahlburg) / Sound effects: 効果音ラボ [Disclaimer] A fan-made science explainer about Hikaru no Go (Yumi Hotta, Takeshi Obata, supervised by Yukari Umezawa / Shueisha), not affiliated with the authors or publisher. No images or footage from the original are used, and the characters' words are retold in our own, not quoted. All rights to the original work belong to its authors and the companies involved. KataGo's scores are estimates under this video's conditions, not a ranking of any player's ability. #HikaruNoGo #Go #Baduk #GoAI #KataGo #AlphaGo #ArtificialIntelligence #Shusaku #Science