Silly AI dumping all its skill points into actual Go play and not off-specing into rules pedantry. They would never make it here in the forums 
7 Likes
Maybe it’s annoyed that the adversary doesn’t resign, wants to show how often it can lose a few points and still win and then fails to win because it has no good understanding of the rules.
It wants to troll, but that backfires.
Now honestly, it’s an interesting question regarding AI research in general. Why does a seemingly superhuman AI fall for these silly things and how can we prevent that?
When it’s about Go, it’s not that bad to have such flaws. It has no impact on my use of KataGo, it’s even fun for those who can learn how to carry out the attacks. But what if it’s about something more serious (e.g. self-driving cars)?
1 Like
Ran a review on that second game.
I don’t understand this blue move. Is that the exploit?
Some bots do like to play way too moves at the end of the game, every atari, every tiny invasion to make sure of two eyes etc.
Is this a kind of weakness when they don’t put enough weight on passing again?
white group is dead
black for no reason plays M17
white captures
graph still shows black as far ahead
but black can’t recapture
so black plays instead:
score suddenly goes down
because white can create 2nd eye now and group became alive
here AI Kata score estimator don’t realize that white is already unbeatable
and 1 move later it realizes
again:
OGS AI review recommends move that is impossible to do
game over
5 Likes
if I capture it, ghost move is in impossible place
level 4 recommends impossible move too
@anoek , AI review and score estimator in game with rules with superko apparently ignores existence of superko
4 Likes
Yep on the todo list
4 Likes
They also tried to fix circle bug. But, its always possible to find new way to draw the circle. It doesn’t matter how many circle variants you fixed.
3 Likes
they even created Go bot with vision transformer architecture
here it wins vs 7-dan professional:
but, it has circle bug too
This rules out convolutional neural network backbone as the root cause of the cyclic vulnerability.
4 Likes
katago-micro had 5 stones handicap but still lost
how it looks like when no tricks are used