Flickr user LNG0004
A human player comprehensively beats a top-ranked AI system at the board game Go. This is a stunning reversal of his 2016 computer triumph, considered a milestone in the rise of artificial intelligence.
Kerrin Perrine, an American player one level below the top of the amateur rankings, took advantage of an unknown flaw identified by another computer to beat the machine. However, the head-to-head confrontation, in which he won 14 of his 15 matches, was played without direct computer support.
The previously unreported victory highlights weaknesses in the best Go computer programs shared by most of today’s widely used AI systems, including the ChatGPT chatbot created by San Francisco-based OpenAI. I made it
The tactic of returning humans to the top of the Go board was proposed by a computer program that surveyed AI systems looking for weaknesses. The proposed plan was relentlessly delivered by Perrine.
“The system was surprisingly easy to use,” says Adam Gleave, chief executive of FAR AI, the California research firm that designed the program. The software has played more than a million games against his KataGo, one of the top Go systems, to find “blind spots” available to human players, he said. Added.
Winning strategies revealed by software are “not entirely trivial, but not terribly difficult” for humans to learn. He also used this method to beat Leela Zero, another top Go system.
The decisive victory came when AI completely outperformed humans in what is often considered the most complex of all board games, albeit with the help of computer-suggested tactics. Brought in 7 years after what it looked like.
AlphaGo, a system devised by Google-owned research firm DeepMind, beat Go world champion Lee Sedol 4-1 in 2016. It cannot be defeated.” AlphaGo has not been published, but the systems Pelrine won are believed to be equivalent.
In Go, two players place alternating black and white stones on a board marked with a 19×19 grid, surrounding their opponent’s stones and trying to encircle the maximum space. The sheer number of combinations means it is impossible for a computer to evaluate all potential future moves.
The tactic used by Pelrine was to slowly string together large “loops” of stones to surround one of his opponent’s own groups, distracting the AI to move to the other corner of the board. Even when the siege was nearly complete, the Go bot was unaware of the vulnerability, Perrin said.
“As humans, it’s very easy to find,” he added.
Stuart Russell, a computer science professor at the University of California, Berkeley, said the discovery of weaknesses in some of the most advanced Go machines is a fundamental flaw in the deep learning systems that underpin today’s state-of-the-art AI. said to indicate that
The system can only “understand” specific situations it has been exposed to in the past, he added, and cannot be generalized in ways that humans can easily spot.
“It shows once again that blaming machines for superhuman levels of intelligence is far too hasty,” Russell said.
According to researchers, the exact cause of the Go system’s failure is a guessing problem. One possible reason is that the tactics exploited by Pelrine are rarely used. This means that the AI system wasn’t trained enough in similar games to realize it was vulnerable, Gleave said.
He added that it is common to find flaws in AI systems when subjected to “hostile attacks” such as those used against computers playing Go.Nevertheless, “we are very big [AI] Systems are deployed at scale with little validation. ”
Copyright The Financial Times Limited 2023 © 2023 The Financial Times Ltd. All rights reserved. Do not copy-paste FT articles and redistribute them via email or web postings.
