Tim Hua comments on Understanding and Controlling LLM Generalization