Yann LeCun: World Models and Foundation Models Are Distinct Concepts
Meta's chief AI scientist clarifies that world models and foundation models are orthogonal — not synonyms, though they can overlap.
In a reply to researcher @anshulkundaje on X, Meta chief AI scientist Yann LeCun drew a clear line between world models and foundation models, calling them "orthogonal concepts." LeCun explained that a world model trained on sufficiently diverse data can become a foundation model — one whose encoders transfer across a wide range of downstream tasks — but the two terms address different things and should not be treated as interchangeable. He also began unpacking what the word "world" means in this context, identifying one sense as the capacity to predict how a complex system or environment evolves over time. For practitioners designing or evaluating architectures, the distinction is load-bearing: whether a system is optimized for world-modeling, broad transfer, or both shapes how its capabilities and limitations should be interpreted — and which benchmarks are actually relevant.