Two models of AI oversight – and how things could go deeply wrong

garymarcus.substack.com