Why LLMs aren't enough for AGI - JEPA is the future | Yann LeCun's 1 Billion $ Bet

In this video, we explore a groundbreaking shift in AI: moving beyond surface-level predictions to true physical understanding. Discover how new architectures like JEPA (Joint Embedding Predictive Architecture) enable robots to learn and act with remarkable efficiency, focusing on the meaning of the world rather than just replicating pixels.We delve into the limitations of current generative AI, such as hallucinating objects and struggling with physical reasoning, and highlight why high-bandwidth sensory experience and abstract world modeling are crucial for true intelligence. Recent breakthroughs, including Meta’s V-JEPA 2, showcase the stark difference when models prioritize logic over appearance.With billion-dollar investments and industry shifts at stake, this episode examines the debate between building world models versus language-centric visions. The future of AI hinges on architectures that grasp physical reality, not just generate visuals or text.Perfect for AI enthusiasts and visionaries, this video reveals why transitioning from pixel prediction to world understanding is essential for the AI of tomorrow. Join us to explore the next frontier in machine intelligence. Yann LeCUn take on world models. Introduction to AI and Robotics - 0:00 Success of a New AI Model - 0:27 Impact on AI Development - 0:53 Welcome and Introduction - 1:22 Generative AI vs. JEPA - 1:35 Understanding Language Models - 2:36 Challenges in AI Reasoning - 3:22 Real-World Intelligence - 4:19 Human Experience Bandwidth - 5:17 Limitations of Language-Based AI - 5:57 JEPA's Approach - 6:47 Efficiency of JEPA - 9:27 JEPA's Learning Mechanism - 9:51 Challenges in Abstract Prediction - 11:28 Real-World Application of JEPA - 13:48 The Billion Dollar Bet: The Rise of AMI Lab - 17:38 Why the "Right Question" is the Future of AI - 20:29