Probably drawing an analogy to how causal pretrained models go through stages of understanding language, words -> grammar -> meaning. Gwen mentions this experience when training character level RNNs. https://gwern.net/scaling-hypothesis#why-does-pretraining-wo...