The reason Yang Lucan left Meta: Transition from LLM to "World Model"

Yang Lucan, a leading researcher in AI, has announced that he will leave Meta after 12 years of service to start a new company. This decision is driven by deep concerns about the current direction of the AI industry. Lucan believes that the entire industry’s obsession with large language models (LLMs) is “irrational,” and he is convinced that true AI evolution lies along a different path.

Limitations of LLMs and the Potential of World Models

The biggest challenge pointed out by Lucan lies in the fundamental limitations of the currently dominant LLMs. While these models possess excellent language capabilities, they lack the ability to understand the physical laws and causal relationships of the real world. According to Lucan, achieving true Artificial General Intelligence (AGI) requires a system that can learn directly from the real world through visual data, spatial information, sensor data, and more. This approach is called a “world model,” aiming for AI to go beyond mere language processing and understand the essence of the physical world.

Support for the Shift to “Spatial Intelligence” from Stanford Researchers

Interestingly, Lucan’s ideas are also supported by other influential researchers in the industry. Fei-Fei Li, a former professor at Stanford University and known as an “AI master,” argues that the next frontier of AI is “spatial intelligence.” Li views current LLMs as “eloquent but inexperienced,” and believes that mere text output cannot truly understand the world. The alignment of these two researchers’ views suggests a significant shift in the development direction of AI technology.

Turning Point in the Technological Direction of the AI Industry

Lucan’s departure from Meta and his plans to start a new company symbolize more than just a personal career decision; they mark a turning point for the entire AI industry. The focus on world models and spatial intelligence clearly signifies a departure from the past few years’ development strategies centered around LLMs. Future AI development is likely to shift from text generation capabilities to understanding the real world, requiring the industry to adapt to this new paradigm.