This attach from underneath at the 2nd line tesuji came from a game between Onishi Ryuhei and Omote Yuto yesterday (Dec 21, 2025)
Not only was it the top candidate move from AI, but also a very surprising tesuji, and the aim is for white to have sente on the top side and expand the UL, or if black blocks from the outside, goes into the corner and takes out all the territory in the corner (connecting out from the right side, and in the real game progress Onishi indeed running out and take out the corner).
Onishi seemed to get stronger after his hiatus, more and more creative with his moves.
16 Likes
This is such a great move! I’d really like to see how the rest of the game went, but the link gives this result:
Is there anywhere else where we can find the sgf file?
2 Likes
I’ve uploaded the sgf to OGS, and change the link in the original post as well.
5 Likes
Thank you! I really appreciate the extra effort. I’ll look into the game shortly. 
1 Like
I find it fascinating how often bots find moves to be “intuitive” that look wild to humans.
Raw policy (i.e. 1 visit, no reading!) from one of katago’s not too old b28 nets:
Summary
For comparison, here’s the policy prior that katago’s human SL net gives when predicting what it thinks a human pro would play (although a human did find the move, the SL net arguably correctly predicts that humans would consider this move highly unusual) :
Summary
(Also of note here - the human SL net thinks that humans would have a decent chance to play S10, which does look pretty natural to me, but the full-strength b28 net considers S10 unlikely to be good. Cases where the b28 net confidently and correctly rejects a move without even needing reading but that rates high for humans are also potentially very interesting!).
I wonder if there’s some kind of new-move-learning / intuition-training tool that can be designed based on analyzing games for moves where the human SL net and the katago policy prior differ greatly.
Hard to develop though, because you might have to be at or not far from pro strength to benefit (us kyu players and middling amateur dans have far better things to focus on than try to understand all the esoteric moves bots come up with), because that suggests you might have to have a developer in a close feedback loop with a pro-level player who’d want to use it to be able to develop the tool to begin with.
6 Likes
Although this move isn’t in my bot-style policy network yet, it does fit with the general advice that to play more like a bot you should play more funky attachments with miai.
3 Likes
From what I heard from pros, some of them dive very deep into studying openings (not possible for all of the openings, but those they favored definitely). Considering Onishi liked unusual openings, it is entirely possible that he already studied some similar positions before, and found the tesuji already (whether he found the moves himself, or just happened to run into this shoulder hit and 4th line K10 response, with black taking the outside influence, it is not out of the question).
Also, I’ve heard rumors that some national pro teams have tools to specifically train players from specific styles or even targeted players (maybe something similar to this Katago feature that gives weights to avoid/encourage patterns from records, or even finetuned networks). Hence, it is not out of the question if you have specific penalty or reward functions to encourage some moves that differ from some specific positions. The problem is always how to define “unusual”, since players like Onishi do exist, and to some extent, like Iyama in the past with his “devil’s moves”. That is certainly shocking to most, but they find them regardless (some dance on the line of trick moves, which will be even more problematic to identify, since they will not be rated high in AI evaluation, but hard to refute regardless)
3 Likes
