A revolution by Yann LeCun: AI finally gains an understanding of the physical world

The revolution in artificial intelligence takes a new turn with the introduction of the V-JEPA 2 model by Meta, led by Yann LeCun. This new AI model stands out for its ability to understand and anticipate actions in the physical world. By integrating an understanding of physical laws into its algorithms, V-JEPA 2 promises to transform the way robots interact with their environment. The ability to perform tasks in unfamiliar environments opens up vast possibilities for practical applications, ranging from domestic robots to assistive technologies. Artificial intelligence is making great strides, but one notable gap remained: understanding the physical world. This has often posed a problem for AI models, which struggled to simulate realistic actions. Today, thanks to a significant effort by Yann LeCun and his team at Meta, the V-JEPA 2 AI model provides an innovative solution. By combining extensive pre-training with minimal data, this model is poised to transform tasks such as robotics and assistive technologies, enabling machines to navigate unfamiliar environments with ease. The Current Limits of AI in the Face of the Physical World Artificial intelligence has evolved considerably, but it has so far encountered a major limitation: understanding the physical world. The performance of video generators, such as OpenAI’s Sora or Google’s Veo 3, while remarkable, often reveals artificial movements that betray a limited understanding of physical laws.V-JEPA 2: The Innovative World Model Meta, led by Yann LeCun, unveiled the V-JEPA 2 model , which stands out for its ability to understand and anticipate actions in the physical world. It is a world model capable of visually interpreting a scene and predicting the reactions of objects. For example, a ball hitting an obstacle bounces back, illustrating a possible prediction for a model such as V-JEPA 2.The Pre-Training Phase and Its Data RequirementsTo provide this advanced understanding, V-JEPA 2 relies on enormous amounts of data during its pre-training phase. This phase required over one million hours of video and one million images, providing the model with the necessary foundations before its specialization with only 62 hours of data from robots performing specific tasks.