弱到强泛化：用弱监督控制强模型的新方向

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?