Yann LeCun on What Comes After LLMs

Large Language Models (LLMs) excel at language manipulation but fail to achieve human-level intelligence because they lack the ability to predict the physical consequences of their actions. True intelligence requires "world models"—architectures like Joint Embedding Predictive Architectures (JEPA)—that enable agents to plan through search and optimization rather than autoregressive token prediction. While LLMs are useful for specific tasks, they are intrinsically unsafe for real-world applications due to hallucinations and an inability to adhere to hardwired constraints. The future of AI lies in data-efficient, objective-driven systems capable of understanding high-dimensional, continuous environments. Beyond technical architectures, the industry requires open, sovereign platforms like Tapestry to prevent centralized control over global knowledge and cultural values, ensuring that AI development remains diverse and accessible rather than restricted to a few dominant corporate entities.