at Vice, Science writer Tatyana Woodall tells us that clever researchers have developed a rival opposing AI to make KataGo lose games:
Players have often used KataGo to test their skills, train for other games, and even analyze past games, but in a study recently published on preprint server arXiv, researchers report using an adversarial guideline – attack some kind of machine learning algorithm or learn weaknesses in other systems – they could beat KataGo by their own means 50 to 99 percent of the time, depending on how much “predictive” the AI is doing. Funnily enough, the new system doesn’t win by trumping KataGo by any means necessary, but rather by forcing KataGo into a corner and essentially making it offer to end the game at a convenient point for its opponent. “KataGo is able to see that passing would result in a forced win for our opponent, but given a tight tree-hunting budget, it doesn’t have the foresight to avoid it,” says co-author Tony Wang, Ph.D . Student at MIT said about the study on the website LessWrong, an online community dedicated to “causing safe and useful AI”.
Tatyana Woodall“Scientists Found a Way to Defeat an ‘Almost Superhuman’ Go-Playing AI” at Vice (November 10, 2022) The paper is freely accessible.
The researchers took advantage of KataGo’s well-known blind spots. Researchers hope to learn more about how powerful AIs tackle problems. According to one of the study authors, Adam Gleave, a powerful AI performs tasks very differently than a human. So when it fails, it will fail in “actually very surprising and alien ways.” Complete lack of foresight seems to be one of them in this case.
As philosopher of technology George Gilder points out, the reason why chess and go slots can usually win is because in these games the card is the territory. So dominance of the map usually means domination of the territory. Real life doesn’t work that way at all, which is why such machines don’t do SkyNet and take over the world.
And then come some people who study the specific weaknesses of AI and take advantage of them. That wasn’t on the map.
You might also like to read: Are computers that win at chess smarter than geniuses? No, and we need to look at why they can win at chess without showing the slightest bit of common sense. The AI thrives where the skill required to win is massive computation and the map IS the territory. In the real world alone, it is helpless. (Georg Gilder)
