Yann LeCun's $1B Bet Against LLMs [Part 2]

Yann LeCun's JEPA (Joint Embedding Predictive Architecture) framework is examined as an alternative to the dominant VLA (Vision-Language-Action) approach for AI and robotics. The video covers VJEPA 2, a Meta model trained on 1 million hours of video without language supervision that achieves state-of-the-art video understanding benchmarks when paired with a language model. It also explores VLJA, a JEPA-based vision-language model that outperforms 7B parameter models using only 1.6B parameters by predicting text embeddings rather than raw tokens. LeCun's two main critiques of VLA models are detailed: (1) behavioral cloning is brittle and unscalable, and (2) VLAs lack explicit planning and world models. The JEPA alternative learns an action-conditioned world model that enables explicit planning via methods like the cross-entropy method, demonstrated on the Push-T task. Hierarchical world models are proposed as LeCun's solution to long-horizon planning. While JEPA-based approaches show theoretical advantages and promising early results, their demonstrated performance on robotics tasks remains significantly behind current VLA systems like Physical Intelligence's PI07.