“It turns out that it achieves an average error of only 0.12 tiles.”
Othello-GPT only ever saw move sequences, never a board, and it plays illegal moves 0.01% of the time. Probes on its activations recover the board state anyway, and flipping a tile inside the model changes its next-move predictions to match the edited board. That is a real result about one 64-square game. Expect it to be quoted as proof that chatbots understand the world.