The difficulty of finding the right reward function in reinforcement learning and how defining an objective to avoid "weird stuff" can mitigate this challenge.