Recently, an amateur chess player from the United States named Kellin Pelrin defeated the top open source Go AI developed by Harvard University: KataGo.
In 2016, Google’s AlphaGo defeated the world’s top chess player Lee Sedol 4:1, leading to a cruel consensus in the Go community that humans cannot defeat AI.
But now, this consensus has been broken by an amateur chess player.
Faced with such a huge disparity, Kellin Pelrin adopted an unexpected method and won with a score of 14:1.
Before the official game, Pelrin used a software from FAR AI to play millions of games with KataGo and successfully found KataGo’s weaknesses.
In the actual game, Pelrin successfully distracted the AI by placing pieces in other corners of the board from time to time, causing it to ignore the increasingly dangerous situation and ultimately lose the game.
In fact, in 2016, Lee Sedol won the only game of AlphaGo, which revealed the weaknesses of AI chess players.
In that game, Lee Sedol made a move beyond AlphaGo’s calculation on the 78th move. This seemed to cause a bug in the AI, causing frequent mistakes and ultimately losing the game.
Judging from subsequent research by AI researchers, finding the blind spots of AI chess players and defeating them through targeted tactics is indeed an effective and reproducible strategy.
But ironically, this tactic can beat the top Go AI, but it cannot beat any experienced amateur player.