Weak-to-strong generalization

2 years ago 9
We present a new research direction for superalignment, together with promising initial results: can we leverage the generalization properties of deep learning to control strong models with weak supervisors?
Read Entire Article