Curiosity gives us an easier way to teach agents to interact with any environment, rather than via an extensively engineered task-specific reward function that we hope corresponds to solving a task