Yann LeCun bets on world models to drive next phase of physical AI

AI pioneer Yann LeCun, founder of AMI Labs, asserts that the next breakthrough in artificial intelligence will stem from ‘world models’ capable of understanding and predicting the physical world, rather than large language models (LLMs) alone. During a seminar in Seoul, LeCun highlighted that while LLMs excel at predictable sequences like code and math, they struggle with the complexities of physical reality, such as robotics and autonomous driving. AMI Labs is utilizing Joint Embedding Predictive Architecture (JEPA) to develop systems that can autonomously plan and interact with environments. The company recently secured $1.03 billion in seed funding, with participation from investors including Bezos Expeditions, Nvidia, and Samsung.

In the robotics sector, industry leaders are recalibrating their expectations. Wang Xingxing, founder of Unitree Robotics, admitted he previously underestimated the speed of AI advancement, noting that the arrival of embodied intelligence could occur within one to two years. Simultaneously, DeepSeek is reportedly exploring entry into the embodied intelligence market. The industry is shifting focus toward how general-purpose models, such as GPT-6, can assist in planning and error correction for specialized robotic hardware, potentially accelerating the integration of AI into physical tasks.

At the Hong Kong Autumn Electronics Fair, various companies showcased advancements in consumer and industrial AI. Beijing-based Acceleration Evolution presented its 1.4-meter T4 humanoid robot for research and education, while SenseRobot introduced AI-integrated chess hardware aimed at global markets.