A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in AI’s ability to understand and reason about the physical world through video. The research introduces a framework where an AI model learns by watching videos of objects in motion, such as a stack of blocks falling or a rolling ball. This allows the system to develop an intuitive sense of physics, predicting how objects will interact without being explicitly programmed with physical laws. The model can then apply this learned ‘common sense’ to perform complex tasks in simulated environments, like manipulating objects with a robotic arm. This approach moves beyond static image recognition toward AI that can anticipate outcomes in dynamic, real-world scenarios. The work highlights progress in building more general and adaptable AI systems that learn from observation. For the full details, read the complete article at https://www.technologyreview.com/2023/10/26/1082391/ai-learns-physics-by-watching-videos-that-dont-exist/