“Misha Laskin's "hot take" is that there is no such thing as true generalization in AI, only the process of bringing the test data distribution into the training data distribution.”