Weak-to-strong generalization

【RSS 原文 · 未经 AI 筛选或翻译】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?