“The goal of OpenAI's Superalignment project is not to train a general-purpose ML researcher, but to develop the specific alignment techniques needed to trust an AI that can perform ML research.”