Wow, with such overlap it looks like optimally-played chess has a much smaller state space than what you'd expect given possible moves each turn. I don't think

Wow, with such overlap it looks like optimally-played chess has a much smaller state space than what you'd expect given possible moves each turn. I don't think this applies to language or image models. But that would also allow models to foil detection by "playing" far enough from optimal.