💭 now that I've solved self play data generation I get to discover that the training pipeline is too slow.