Since 2016, when AlphaGo beat Lee Sedol, in all probability the world’s strongest participant, people have primarily not had an opportunity towards a well-trained AI.

Now, a paper “Adversarial Policies Beat Professional-Level Go AIs” has been uploaded to the ArXiv. Ars Technica gives a nontechnical account: “New Go-playing trick defeats world-class Go AI—however loses to human amateurs”. If I perceive accurately, the concept is that the adversary takes a small territory within the nook, leaving the bigger half to the sufferer AI, however the adversary performs lifeless stones within the bigger remainder of the board. The sufferer AI, on the one hand, sees these stones as lifeless, and subsequently doesn’t make investments strikes in killing them. But throughout scoring, the sufferer AI apparently does not see them as lifeless and subsequently doesn’t depend the big territory as territory.

Here is an instance from the Ars Technica article:

Note on the appropriate that the AI believes that black wins by 49.5 factors. There are 6.5 factors komi, so black ought to win by 56 factors on the board. For the lifetime of me, I can not determine this out – the highest proper nook is 52 factors massive and there are not any captures. And extra importantly, any participant of 20-kyu power or stronger instantly sees that the black stones scattered within the enormous white territory are merely, unquestionably, lifeless – so previous to counting, they need to be eliminated and depend as captures for white below any cheap algorithm.

Thus, it appears to me like this doesn’t exploit a weak point in KataGo’s go taking part in means, however a bona fide bug in its game-end territory counting algorithm. It’s relatively clearly a bug, as a result of the foundations are clear on what ought to occur on this state of affairs after we depend the rating, and KataGo clearly doesn’t observe the foundations.

It doesn’t make sense to me to publish this below a headline of “beating” KataGo, relatively as “we discovered a bug within the game-end scoring of KataGo”.

What am I lacking?