It's getting rarer to find AI bugs, but they still appear. I remember thinking around January 2021 that katago has managed to "understand" the strange seki "that Lee Changho couldn't solve", a diagram is at the end of this post: https://www.lifein19x19.com/viewtopic.p ... 89#p265889. Before then, it seemed to evaluate the group as dead, only recognising it was seki when all dame were filled, but around January, katago started to correctly evaluate it.

In a recent katago self-play game https://katagotraining.org/sgfplayer/ra ... es/547118/, just before black resigns, on move 173, black missed a chance to kill white's big group. Black should play at D11.
This implies white 172 is a blunder that should be at D11.

Even when putting it into the analysis engine of kg943, it evaluated the position as W+30-50 until I played it out until the end of the game when it swung in the opposite direction. I wouldn't have dared to post this otherwise, despite being fairly confident I'd checked all of the many variations (without AI).

I found this accidentally and manually, since I was confused as to why black resigned. It is true that in the final position white is unconditionally alive, by some tesujis that took me perhaps 1hour to figure out. It is surprisingly difficult for such a small space, but still a human level problem. Probably pro level.

Hint:


However, black playing at D11 instead of D14 gains black's potential D10 stone a critical two liberties so that moves at F11 do not force E12 even if white connects at D14 in sente (I plan for black to answer at C14 in the worst case). B's outside is also alive because B has O6 in sente.

Curiously in the semeai in the lower right, black making a slightly bigger eye at N1 seems to increase their liberties from 5 to 7 since a white placement inside the eye requires throwing in 3 stones (due to the hane), which still gives black an eye so it seems white is forced to fill the outside liberties first (at T2 and T1), however, after throwing in the 3 stones, white can throw into the eyespace again, so it would actually only be 6 liberties.

Here is a simplified diagram to demonstrate why white can't live locally.

Click Here To Show Diagram Code

[go]$$ Black to kill
$$ | X X X X X . .
$$ | O O . . X . .
$$ | . O O . X X X
$$ | . X O . . . X
$$ | . X X O O . X
$$ | O X O X X X X
$$ | O O O X . . .
$$ | . . X X . . .
$$ | . X . . . . .[/go]

Nice example! Something annoying about neural nets - if their raw judgment is wrong, the search's ability to fix them is surprisingly limited. Which is entirely *unlike* humans, who can and do adjust their judgment intra-game based on the results of their reading.

More training should fix this at some point though. Faster if I add in some starting positions from this game to be 0.0001% of training.

I also agree with your recent post https://lifein19x19.com/viewtopic.php?f=10&t=18563
that current Go AI still seem beatable by humans occasionally. Shin's opening is perfect a non-zero amount of the time, and his middle-game is considered the best. Endgame mathematics can help support.

The issue in this case seems to be that value and policy are statically guiding search rather than interactive with search.

For example an improvement might be to evaluate the importance of certain shapes/groups and focus search on moves that affect the life and death of those groups.

This is slightly like adding a brain on top of an automata, letting the current board position manipulate search.

Your control estimate probably greatly helps guide training, but at higher levels of strength, a support to encourage deeper reading of important moves (not just likely good moves that the network has read) will probably increase strength.

I think this is all fairly obvious though I don't know how it would be implemented (perhaps like a ladder recogniser?). But what are the consequences?

Though I must admit to fear about making invincible AI and the consequences on humanity. Even at an individual level, I get the feeling that training a neural network trains the human user in unusual ways, and may create new forms of poison/disease/communication.

Once connections are made, an opponent can manipulate it to blind training for certain aspects of the game. Perhaps it is a form of weakness to be so flexible, and strong enough AI systems might manipulate the game itself and what aspects we see as important simply because it has such a systematic method. But all things die. Which aspects will win? What work will be done? What artwork produced?

Are scientists manipulated to uncover secrets at the edges of our reality? Does reality bite back?

[Hide]
Adrenaline + Brain = ?
[/Hide]

Gosh - hold the implementation of the importance of shapes/groups evaluator. That took 'consequences' beyond what I was expecting,

The ethics and societal impacts of AI are interesting indeed, but take it easy dhu163 (and best not read any Philip K Dick).

my plant tells me I might have crossed a boundary I shouldn't be crossing with this research. Hint: It's not a kidney bean.
I think I'll respect that for now.

20220204: but I went outside and the air told me the key to perfecting the iteration program. Some extraction cycles are required, forcing the errors out into eyes, but where do the errors go? And does the mining occur too deeply. The random number resources are too strong, or are they just recycles?

I still wonder if I should continue.