thaumasiotes 6 days ago

> Here's one way to test whether it really understands chess. Make it play the next move in 1000 random legal positions

Suppose it tries to capture en passant. How do you know whether that's legal?

1
BalinKing 6 days ago

I feel like you could add “do not capture en passant unless it is the only possible move” to the test without changing what it’s trying to prove—if anything, some small permutation like this might even make it a stronger test of “reasoning capability.” (Personally I’m unconvinced of the utility of this test in the first place, but I think it can be reasonably steelmanned.)