Arguments about Highly Reliable Agent Designs as a Useful Path to Artificial Intelligence Safety — LessWrong