Researchers at OpenAI have created Random Network Distillation (RND), a new approach to reinforcement learning that encourages exploration by predicting what will happen next. This method has led to an AI agent surpassing average human performance on the challenging game Montezuma's Revenge for the first time. The RND technique uses a secondary network to predict the output of the main network, which helps guide the agent's actions. --- Why it matters: This breakthrough matters because it shows that reinforcement learning can be improved by incorporating prediction-based rewards, potentially leading to more efficient exploration and better performance in complex tasks. Source: https://openai.com/index/reinforcement-learning-with-prediction-based-rewards