Yann LeCun, Chief AI Scientist at Meta, in his first public interview after leaving the company, openly stated that the huge investments by large tech companies in large language models (LLMs) are a "strategic mistake" and do not lead to true computer intelligence. He pointed out that LLMs are only good at "statistical text completion," lacking understanding of the physical world, long-term reasoning, and planning capabilities, and therefore "are destined to not produce human-level intelligence."

Metaverse, Sci-fi, Cyberpunk, Painting (3) Large Models

Image source note: The image is AI-generated, and the image licensing service provider is Midjourney

LeCun believes that the next breakthrough should come from "World Models." These frameworks can integrate vision, hearing, and environmental states, simulate real-world dynamics, allowing machines to gain spatial, temporal, and causal concepts similar to those of human infants. He revealed that Meta has already launched a world model project trained on 1 billion hours of video, with a prototype expected to be released in 2026. The goal is to perform multi-step tasks in virtual environments and deeply integrate with AR/VR hardware.

This statement is seen as an open correction of Meta's AI roadmap. A company spokesperson responded that LLMs remain the core of short-term products, but the company has also simultaneously increased investment in world models and embodied intelligence research. Industry analysts point out that if world models make progress, autonomous driving, robotics, and AIGC will have a new technological foundation, but data collection, computing power requirements, and interpretability remain obstacles to implementation.