Life In 19x19

It is well known that playing mirror go with tengen as black doesn't work well since tengen can't pay for komi if everything is symmetric.

Similarly playing mirror go as white is very difficult to refute. There are ladder examples but they tend to be very complicated. To the point where it can be disrespectful to mirror as white.

From what I've tried against katago, mirroring as white keeps the game even for a very long time but katago tends to strategically set up a 100 point semeai until it finally plays tengen which connects with several diagonal moves making a 1 eye kills no eye situation. It seems very difficult to mimic this as a human playing black since you have to play every move accurately while setting up the semeai.

Perhaps a much more advanced way to bully Akira.

20220521: corner fights are ok, but avoid fights that spiral into the centre. In fact with lots of weak groups fighting, this seems to be good for white if they can all settle just before the centre.

Note that you'll get different behavior depending on whether you are playing KataGo directly via a program that talks to it using GTP and has it play moves via normal GTP protocol, or instead playing using a program that asks it to "analyze" the board in real time via kata-analyze or lz-analyze (i.e. like Lizzie when in interactive mode), and then takes the top move from the analysis to play on the board.

KataGo has some logic to respond to mirror Go, that very gently nudges its normal behavior in the direction of doing the kind of thing you mentioned, and much more forcefully nudges it in harder cases (such as Black mirroring with reverse komi). But by default this logic is only enabled in play and not in analysis, because this nudging also biases the reported winrates in various messy ways. This logic is also a bit better in the latest release 1.10 than in prior versions.

Mostly programs use the correct protocol corresponding to what they want to do, but a few applications probably use (abuse?) the "analysis" for play as well. If you care to, you can also turn it on in analysis as well though, check out the options in the the gtp config.

I am happy enough to play ladder variations since they give an immediate advantage, although it wont be hige but goood enough

Can we determine how close an AI is to perfect play if we remove added anti-mirror go nudges and ladder recognition, and the better it responds to mirror baduk through nothing but just it's self-play training, the more inherently close to god it is as the sooner we can have AI baduk religion?