Two models of AI oversight – and how things could go deeply wronggarymarcus.substack.com·1 pts·son_of_gloin·0