Yann LeCun pushes AI strategy beyond LLMs

Yann LeCun is challenging the dominant push toward ever-larger LLMs, arguing that systems built mainly to predict the next token do not understand the physical world. His critique centers on auto-regressive models, which he says accumulate small statistical errors that can lead to hallucinations and unreliable outputs.

LeCun’s proposed alternative, Joint Embedding Predictive Architecture, or JEPA, focuses on learning essential relationships in a representation space rather than reconstructing every detail. The approach aims to help AI build internal world models from video and images, enabling it to grasp causal relationships before generating language.

JEPA-style systems could reduce training costs by up to 90% compared with traditional Transformers, according to the report. The strategy also emphasizes local-device deployment, less reliance on large data centers, and stronger planning and real-time problem-solving capabilities.

Open-source development is a central part of LeCun’s position at Meta. By releasing research and models to the community, the strategy seeks to make JEPA principles a broader standard and challenge the dominance of proprietary AI systems from companies such as Microsoft and Google.