Along with developing a framework for scalable oversight, we also conduct a proof of concept experiment that demonstrates a couple of question-answering tasks that work well under this paradigm with current language models:

→ View original post on X — @anthropicai,