研究者将 LeCun 的 JEPA 扩展为跨物理到生物的通用世界模型
原标题：Researchers stretch LeCun's JEPA AI into a universal world model that works from physics to biology
AI 摘要
PhAI Labs 联合斯坦福、牛津、普林斯顿等机构提出 JEPA-Anything，将 Yann LeCun 的 JEPA 架构扩展为跨物理、机器人、医学等七个领域的通用世界模型。该方法把未来状态拆成多个部分预测而非单一预测，在动力学任务上误差最多降低约 35%，并在肝癌研究中提出 IL-18 联合 CD73 阻断的候选疗法，实验室和小鼠实验显示联合效果优于单用。作者提醒，学到的部分预测分离并不等于因果关系，模型何时可靠到能指导实验设计仍是开放问题。
正文节选
Researchers stretch LeCun's JEPA AI into a universal world model that works from physics to biology Key Points - The JEPA-Anything AI model extends the architecture pioneered by Yann LeCun into a world model that works across fields like physics, robotics, and medicine. - The method breaks future states into multiple partial predictions instead of funneling everything into one, letting it pick up patterns that standard models miss. - The model also identified a liver cancer treatment candidate t