Grounding LLMs with JEPA-based world models trained in simulation — has this been tried? [D]

Explain Like I'm Normal

A proposal to solve the 'hallucinating physics' problem by grounding LLMs in Joint-Embedding Predictive Architectures (JEPA) trained in simulators. Instead of guessing the next word, the model learns to predict future physical states in a latent space where gravity isn't optional. This addresses the gap between knowing the word for 'falling' and understanding the momentum of a drop.