@imitationlearn 🤔 what sort of training pipeline wouldn't have its own worse inductive biases?