I wouldn’t call it a “leap” but a few years ago some researchers did find a weird byway in the then-current go-playing "AI"s (neural-network setups, as someone else pointed out). They had to cheat a bit by examining the NN’s “thinking” more directly, and found a cyclic strategy that allowed a human to beat the machine.
It’s interesting for more “AI”-type stuff generally. The point is not so much “us with our special brains will always find ways to defeat AI” but more “there are odd blind spots that you would not predict by just looking at the output/games, and these can be exploited with appropriate technology.”
I wouldn’t call it a “leap” but a few years ago some researchers did find a weird byway in the then-current go-playing "AI"s (neural-network setups, as someone else pointed out). They had to cheat a bit by examining the NN’s “thinking” more directly, and found a cyclic strategy that allowed a human to beat the machine.
https://www.far.ai/blog/even-superhuman-go-ais-have-surprising-failure-modes
It’s interesting for more “AI”-type stuff generally. The point is not so much “us with our special brains will always find ways to defeat AI” but more “there are odd blind spots that you would not predict by just looking at the output/games, and these can be exploited with appropriate technology.”
Edit: remove redundantly redundant redundancy
that is cool and not uncommon in algos in general
bad cost function, poor fit / overfit, local inflection etc