💭 going to try figuring out low hanging fruit for ml training efficiency to see if I can 'easily' make things adequate. Thinking of also using literal tictactoe with very few mcts sims as a way of testing the model works - at least on a toy problem. It leaves the scaling question for later.