Comment by tromp

I'm impressed and surprised that a relatively small model can learn so much from just the textual move records. Not even full algebraic notation 1.e2-e4 e7-e5 2.Ng1-f3 but just 1.e4 e5 2.Nf3 ... It has to figure out what o-o and o-o-o mean just from what King and Rook moves appear later. And the only way it could learn that the King cannot move through a checked square while castling is that such situations never appear in its training set.

Would love to see a similar experiment for 9x9 Go, where the model also needs to learn the concepts of connected group and its liberties.