In 2016, the news was that AI beat humans at Go. Fast forward seven years, and the news is that humans beat AI at Go. But it’s not like we got much smarter between tries — we simply learned to exploit its bugs.
Go is so mind-bendingly complex that it makes chess seem like tic tac toe. Go is played on a 19 by 19 board (compared to just 8 by 8 for chess), and a typical game of around 150 moves has around 10360 possible moves, or 1 followed by 360 zeroes — a number that’s simply unfathomable. For comparison, it’s estimated that there are some 1082 atoms in the universe.
Calculating everything in the game of Go is simply not possible, so players often rely on their intuition and pattern recognition skills, which is why Go was thought to be unconquerable by AIs. But in 2016, DeepMind’s AlphaGo turned all that on its head. Despite staunch resistance from mankind’s champion, AI triumphed and got more and more ahead of humanity.
The best player of Go is currently KataGo, a machine-learning algorithm that taught itself how to play, surpassing even previous AI iterations.
KataGo is a monster, it just wipes the floor with all opponents. But researchers have been looking for potential flaws or weaknesses in KataGo. Recently, a team of researchers published a preprint of their research in which they describe how they train their own AI opponents, specifically aimed at KataGo. They don’t want to become better players, they just want to trick the AI.
“Notably, our adversaries do not win by learning to play Go better than KataGo – in fact, our adversaries are easily beaten by human amateurs,” the team wrote in their paper. “Instead, our adversaries win by tricking KataGo into making serious blunders.”
This is where Kellin Pelrine steps in. Pelrine is a good player, but an amateur. Specifically, he’s one level below the top amateur ranking. He’s also one of the study authors, so he was well aware of the vulnerabilities of KataGo, so he thought why not try his own hand?
Apparently, it was surprisingly easy to find a way to defeat AI by exploiting its weakness. Pelrine managed to beat KataGo 14 out of 15 times. For comparison, KataGo beat AlphaGo 100 times out of 100, and AlphaGo beat mankind’s best player 4-1.
But as is so often the case, this isn’t about the game itself, it’s about what this means for the future of artificial intelligence. The main takeaway is that performance doesn’t always translate into robustness. This failure of the Go-playing algorithm is a bit like a self-driving car crashing into a tree because the bark had a specific color. In other words, even when something seems to be performing extremely well, there could be fringe situations where it behaves badly. This is less of a problem in Go, and more of a problem when AI steps into the real world, so this is an important cautionary tale.
Crucially, Pelrine’s tactic would have been quite easily spotted by a human. He simply created a loop of stones to encircle the opponent’s stones, but then started making moves in the corners of the board to distract the AI. It’s not completely trivial, says Pelrine, but not very difficult.
Artificial systems, however, don’t have the ability to react to situations they’re not prepared for. They don’t have “common sense”. In fact, this is why game-playing AIs are so important: they teach us about how these algorithms behave — not just in terms of opportunities and performance, but also in terms of what can go wrong.
It’s common to find flaws and exploits in AI systems. Ironically, this is also done with the aid of computers, but this field is extremely important and often overlooked. More and more, we’re seeing AIs being deployed into the world with little verification. Maybe, just maybe, we should learn from this type of event and pay more attention to how we deploy such systems in real life.