“Larger language models exhibit a different form of generalization than smaller models, such as being able to apply information learned in one language to another.”