Shin Jin-seo, a 9-dan and the world's No. 1 in Go, defeated KataGo, one of the strongest Go artificial intelligences (AI) in existence. He was rattled by KataGo's unfamiliar opening in the first game and lost after a come-from-behind defeat, but in the second match he reduced early variables with a large joseki devised by AI and maintained the edge in the center fight.
On the 19th, at The Korea Economic Broadcasting studio in Jung District, Seoul, Shin Jin-seo beat KataGo by 4.5 points with Black after about 4 hours and 50 minutes in Game 2 of the best-of-three "Ssen Math·Hankyung Gisinjeon." The match was played as a handicap game in which Shin placed two black stones first.
After resigning in 245 moves on the 17th in Game 1, Shin Jin-seo evened the score at 1-1 with the win. It was his first victory in an official game against KataGo. Shin had said before the showdown that in practice against a similar-strength AI, he had never won even once with a two-stone handicap.
The key to victory was an ultra-large joseki that arose in the lower-right corner early on. When KataGo, as in Game 1, took the upper-left star point, Shin answered at the lower-right star point. It was a different choice from Game 1, when he had chosen the small-knight move at the lower right.
When KataGo immediately invaded at 3-3 in the lower right, Shin steered into a large-scale joseki. In just about five minutes, the two players progressed to move 53, settling roughly a quarter of the board. Although they could have ended the joseki simply midway, Shin chose a complex variation.
This joseki is the so-called "AlphaGo joseki" that professional players have studied since AlphaGo appeared in 2016. By fixing a wide area early through a set sequence, it had the effect of reducing the number of variations KataGo could generate afterward. In effect, he used techniques AI left to the Go world to target AI.
With a two-stone handicap, broadcast AI analysis shows Shin starts with a 99% win rate, ahead by about 18.5 points. But in Game 1, Shin was flustered by KataGo's early high three-space approach, and the opening he had prepared unraveled.
At that time, after moves 70 and 76, when he tried to reduce the center, Shin's 99% win-rate line collapsed, and with move 90 in the lower right turning out to be a mistake, he allowed a chase. As the game tightened, he tried to surround the large white group in the center with move 102, but when KataGo escaped, the game flipped.
Game 2 was different. After narrowing the board early with a large joseki, Shin maintained the lead past move 150. In the center, he cut KataGo's stones and launched an attack. When KataGo attached near the connection area and counterattacked, Shin's win rate on the broadcast screen at one point fell to 89%, but he did not avoid the fight and answered through reading.
Unlike in Game 1, when he tried to surround the entire large white group in the center and got overturned, this time he cut White's consolidation and blocked KataGo's counterattacks in sequence, keeping the upper hand. KataGo threw a last-ditch tactic in the lower center, but Shin accurately sealed it off and preserved the point margin in the endgame.
After the game, Shin said, "I thought I had prepared very well, but KataGo attacked so hard I thought I would lose," and added, "I cooled down, calculated, and pulled myself together. I'm happy with a valuable first win."
KataGo is an open-source Go AI developed by U.S. computer scientist David Wu. It was developed by improving the self-play learning method of the AlphaZero family and can calculate not only the win rate but also the expected point margin. It is widely used for training by professional players and national team members and for broadcast analysis of Go.
In this match, Shin was given five hours and one 30-second byoyomi, while KataGo had to move within 20 seconds every turn. With the Game 2 win, Shin secured a 50 million won victory bonus. The final Game 3 will be held at 10 a.m. on the 21st at the same venue and will be broadcast live on Baduk TV. If Shin wins Game 3 to reach two wins, he will receive a Genesis G90 as a prize.