Toy models of AI control for concentrated catastrophe prevention — LessWrong