Computers have been beating the best human Go players since 2016. The Go world champion retired in part because AI is “an entity that cannot be defeated.”
But a human just trounced one of the world’s best Go AIs 14 games to 1: https://arstechnica.com/information-technology/2023/02/man-beats-machine-at-go-in-human-victory-over-ai
I think this news story is more interesting than it might first appear (without knowing details, so grain of salt). It isn’t just a gaming curiosity; it points to a fundamental flaw with “deep learning” approaches in general.
1/
The Go AI was trained by feeding it a huge number of Go games played by humans. It built a model based on what humans do.
The human beat it by doing something so obvious a human would never do it — so it wasn’t in the training data, so the AI didn't counter it.
(Basically, the human forms a conspicuous giant capture ring while distracting the AI with tactical battles the AI knows how to counter.)
2/
@inthehands I don't about KataGo, but AlphaGo's stronger descendants were NOT trained on human games -- they were merely given the rules and told to play against themselves. I would be very curious to know if this exploit works against them.
