 We really don't know what's happening inside these models because they're black box models. They're neural networks that have many, many hidden layers. So the relationships that they're developing between words, between concepts, the patterns that they're latching onto are completely opaque even to the people designing them. So it's a kind of alchemy.