PhAI Labs publishes JEPA-Anything, a JEPA-based universal world model across physics to biology

KEY POINTS

1. Researchers led by PhAI Labs (with collaborators at Stanford, Oxford and Princeton) introduced JEPA-Anything, a Joint-Embedding Predictive Architecture that predicts multiple partial future-state embeddings instead of a single summary.
2. The model outperforms a same-architecture JEPA baseline on diverse tasks (physics, robotics, weather, chemistry, single-cell and clinical data), discovered a near-exact numeric match to Kepler's third law in a trained run, and highlighted an IL-18 plus CD73-blockade combination that showed stronger anti-tumor activity than either component alone in organoids, patient tissue samples and mice; code and models are publicly available and the authors caution that causal interpretation and clinical utility remain unproven.
3. A single JEPA-based world model that generalizes across domains and can suggest experimentally testable biological hypotheses could reshape how AI aids scientific discovery, though practical reliability and causal validity remain open questions.