Comparing to 5x5 Go, 9x9 Go feels much more like the real Go. One needs to balance territory and influence. It’s a lot more fun, but complexity is a lot higher.
To view the games, click the view-games link and choose the game from the top-left dropdown on the page.
I downloaded a 9x9 game recently from the current katago training run (early Mar 2022). The game was generated by one of the latest models (b60c320), which is likely super-human, and felt quite competitive.
Then I analyzed it with a b40c256 model (“strongest confidently rated” in katatrain). b40 disagreed with certain moves. Some analysis might be wrong, but it’s hard for me to tell. See commented game 1
Below is my own take on what happened. Disclaimer: I’m not nearly as close to any of these models (super-human). But that’s the way we humans look at a game.
The game turned into an epic fight around move #22 (C2), where white launches an attack on the main black chain, which seems to have played a heavy hand on the white corner in the bottom right. This is the only chance for white to win. The ensuing battle hinges on a couple considerations:
Black move #49 (F9) is an attempt to increase its territory, at the risk of losing the fight. Here the b40 model thinks the best move is E3, which prioritizes killing the white corner. If we follow this path, black wins the battle, but loses the game by maybe one point. Thus, F9 is a gamble to gain that one point, but risks losing the battle (and the war completely). No good options, unfortunately.
White move #52 (H7) is also interesting. Otherwise, black might be able to buy more time. But then again, I am not good at tight fighting, so you have to take my analysis with a grain of salt.
In the end black lost the battle. White took all. Bot knows this at move #55, as win rate dropped sharply and never recovered. When we examine the final board (#69), black is behind by only one step in the capture race. Also, if black wins the race, it wins by only 1 point. This brings us back to the motivation behind black move #49 (F9). F9 is purely for territory gain, which costed black the race in the end; but without which black will lose any ways.
All of this is highly unsatisfactory. We have several super-human bots, but still cannot agree on what the right moves are. At what point is black doomed? Is there a potential winning path that we didn’t know about?
Here is another self-play game in the recent run. Again, a complex fight, but this time it ends in eternal life. KataGo marked the result as void. Seems like the bot is pushing itself to the limit. See eternal life.sgf