Data is the linchpin for advancements in artificial intelligence, but robotics faces a unique challenge: collecting enough diverse and high-quality data is expensive and time-consuming. While large language models (LLMs) and vision models have access to vast datasets from human-generated text and images, robotics is lagging behind. Current methods rely heavily on teleoperation, which is limited by the number of robots and skilled operators available. The article argues that the key to overcoming this data bottleneck lies in leveraging human experience. With 8 billion people each generating around 16 hours of sensorimotor experience daily, the potential for data collection is immense. Recent developments in human egocentric video datasets highlight a promising approach. The author envisions achieving 100 million hours of such data—equivalent to 150 human lifetimes of experience. However, raw video alone isn't sufficient for teaching robots; it must be paired with descriptions of actions. The proposed solution is to create a world model that predicts how the physical world evolves based on this video data. This model should not only understand the dynamics of physical interactions but also be capable of predicting outcomes based on human intentions, which can be expressed through natural language. The article introduces DreamZero, a model trained to predict future frames and actions from video data. This approach shows that a well-designed world model can generalize across tasks, even those not present in its training data. For example, it can enable a robot to perform actions like untying shoelaces or shaking hands. The next step is to transfer this understanding to physical robots, which requires a body capable of executing tasks based on the knowledge gleaned from human experiences. The integration of language annotations with video data is crucial for teaching robots to interpret and act on human intentions effectively.