Kellin Pelrine, an American gamer, managed to defeat an artificial intelligence on the game of Go. One of the AI’s greatest victories has lapsed.
The victory of artificial intelligence over the game of Go, one of the most complex strategy games in the world, caused a stir in 2016. Today, humanity has regained the ascendancy thanks to Kellin Pelrine, a player who is not even ranked. He defeated the artificial intelligence KataGo in 14 of 15 games that they played together.
But how ? The irony is that Kellin Pelrine used… a computer program for this. He is indeed part of the team of FAR AI, a Californian research center dedicated to artificial intelligence, which has designed a dedicated program only to the analysis of KataGo’s weaknesses. It is thanks to this that a great flaw in it was identified and used by the team, who were also able to achieve the exploit facing Leela Zeroanother AI specialized in Go.
AI fooled by lack of complexity
Decisions made by Kellin Pelrine during his games were not dictated by a computer. But the player was armed with crucial information: a confusion technique. The principle of the game of Go is to use its stones (black or white, depending on your side) to surround the opposing stones, capture them and take control of the field. It’s one of the oldest and most complex strategy games in the world, so much so that when world champion Lee Sedol was defeated by Google’s DeepMind in 2016, many people saw it as the victory of the machine against all humanity.
But these artificial intelligences can be fooled by a simple technique, discovered here. By placing his stones slowly and strategically to create a large capture circle while making moves on the other side of the game board, it is possible to deceive the vigilance of the AIs. A tactic that would, on the other hand, be very easy to spot for a human according to Pelrine. The team only tested their discovery on KataGo and Leela Zero, both publicly accessible and in their latest versions, but it’s possible the same technique could fool you. AlphaGo, whose access is private. Even when testing their discovery with a nine-stone handicap, which would give anyone playing in real life a 100% chance of winning, the Far AI team was able to turn the tide against the AI.
For Stuart Russell, an AI researcher at the University of California interviewed by the Financial Times, this discovery helps to highlight one of the biggest weaknesses of these systems. Since they can only “understand” specific situations to which they have been exposed through their learning models, artificial intelligences are not able to generalize just like a human does. For the researcher, it proves once again that we are in too much of a rush to deliver superhuman levels of intelligence to machines. »
Not being a ranked player has therefore been Kellin Pelrine’s greatest advantage in his games since his way of playing, however banal it may be, does not fit into the millions of models analyzed by KataGo and Leela Zero. Nothing says, however, that after studying this new data, these two artificial intelligences will continue to be fooled in the future. One lesson remains if science fiction ever becomes reality: typical human stupidity could be our greatest salvation.
FinancialTimes
Maxime “OtaXou” Lancelin-Golbery
