“The "Weak to Strong" generalization method involves using a more capable model, like GPT-4, to teach a less capable model, like GPT-2, to be smarter than itself.”