Video generation has the potential to serve as a new language for real-world decision-making, similar to how language models have transformed digital interactions. While language models have made significant strides in various tasks, video generation remains underexplored, despite its ability to capture rich physical world information that is difficult to express in text. This paper argues that video generation can be a unified interface for absorbing internet knowledge and representing diverse tasks, much like language models. By leveraging techniques such as in-context learning, planning, and reinforcement learning, video generation can function as planners, agents, compute engines, and environment simulators. The paper highlights opportunities in domains like robotics, self-driving, and science, supported by recent advancements in video generation capabilities. However, challenges such as dataset limitations, model heterogeneity, and hallucination must be addressed to fully realize the potential of video generation. The paper also discusses how video generation can be used for simulation, enabling the optimization of control inputs based on simulated results. Applications include generative game environments, robotics, self-driving, and scientific simulations. Despite these advancements, challenges remain in generalization, generation speed, and long-term consistency. The paper concludes that video generation has the potential to become an autonomous agent, planner, and environment simulator, capable of thinking and acting in the physical world.Video generation has the potential to serve as a new language for real-world decision-making, similar to how language models have transformed digital interactions. While language models have made significant strides in various tasks, video generation remains underexplored, despite its ability to capture rich physical world information that is difficult to express in text. This paper argues that video generation can be a unified interface for absorbing internet knowledge and representing diverse tasks, much like language models. By leveraging techniques such as in-context learning, planning, and reinforcement learning, video generation can function as planners, agents, compute engines, and environment simulators. The paper highlights opportunities in domains like robotics, self-driving, and science, supported by recent advancements in video generation capabilities. However, challenges such as dataset limitations, model heterogeneity, and hallucination must be addressed to fully realize the potential of video generation. The paper also discusses how video generation can be used for simulation, enabling the optimization of control inputs based on simulated results. Applications include generative game environments, robotics, self-driving, and scientific simulations. Despite these advancements, challenges remain in generalization, generation speed, and long-term consistency. The paper concludes that video generation has the potential to become an autonomous agent, planner, and environment simulator, capable of thinking and acting in the physical world.