Man beat AI in a game of go… with a computer

The human player manages to defeat one of the best AI in the game thanks to a vulnerability discovered by another AI.

An amateur but very experienced Go player has beaten a very powerful artificial intelligence system. This feat was made possible by a vulnerability discovered by another computer, according to The Verge. Taking advantage of this disadvantage, American player Kellin Pelrin managed to defeat the KataGo system, winning 14 out of 15 games without further help from his computer ally. This is one of the very few human wins since AlphaGo in 2016 that helped pave the way for the current AI frenzy. It also shows that even the most advanced AI algorithms can have blind spots.

Human player manages to beat one of the best AI games ever

Kellin Pelrin’s victory was made possible by specialized research company FAR AI, which developed a program to identify flaws in KataGo. After playing a million games, this software was able to find a flaw that a reasonably skilled human player could exploit. It’s “not exactly trivial, but it’s also not very difficult”to master, said Kellin Pelrin. The man used the same method to defeat Leela Zero, another highly advanced Go-AI.

Here’s how it works: The goal is to create a large “loop”of rocks to surround a group of opponents and then distract the computer by making moves to other areas of the board. Even when his group was almost surrounded, the AI ​​did not notice this strategy. “It would be pretty easy for a human to spot,” Kellin Pelrin said when the surrounding rocks are clearly visible on the board.

thanks to a vulnerability discovered by another artificial intelligence

This flaw demonstrates that AI systems cannot “think”beyond their learning. Then they often do things that people think are completely stupid. We’ve already seen this with chatbots like Microsoft on their Bing search engine. While the system is fairly good at simple repetitive tasks like plotting a route, it spits out false information, scolds users for wasting time, and even exhibits “unbalanced”behavior – most likely due to the template it was trained on.

Lightvector, the developer of KataGo, is likely aware of this vulnerability, which players have been exploiting for months. In a GitHub post, the company explains that it is working on fixing several types of attacks of this kind.