No Such Thing as a General Learner: Language models and their dual optimization