💭 alphazero specifically could get more training data by using in search tree states as rows so long as they're thoroughly enough explored. Could actually weight how strong the training data is as evidence though I've never heard of that being done