Weak-to-strong generalization

We current a brand new analysis course for superalignment, along with promising preliminary outcomes: can we leverage the generalization properties of deep studying to manage robust fashions with weak supervisors?