Models & research
OpenAI
Weak-to-strong generalization
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?
Excerpt supplied by the publisher’s feed