“The transition from GPT-1 to GPT-2 demonstrated that models can achieve generalization by training on a wide variety of data, rather than being explicitly taught every possible skill.”