“Yoshua Bengio's tentative solution to AI safety is to design AIs as non-agents that can only make predictions ("oracles") and cannot take actions.”