Yann LeCun:LLM不是通往人类智能的道路Yann LeCun: LLMs Are Not the Path to Human-Level Intelligence
核心要点:LLM在语言任务上表现出色,但缺乏处理物理世界真正智能的架构;基于JEPA的世界模型才是前进方向。Yann LeCun作为Meta前首席AI科学家和图灵奖得主,分享了他离开Meta创立新公司专注于现实世界AI的原因。他认为现实过于复杂、连续且嘈杂,无法用自回归token预测处理。能够预测动作后果并通过搜索进行规划的世界模型至关重要。“LLM本质上是不安全的……它们无法预测自己动作的后果。”他对比了JEPA的非生成式方法,该方法学习抽象表示。他预计这种范式将在机器人、工业控制和医疗保健领域实现五年内主导世界。
The Takeaway: LLMs excel at language tasks but lack the architecture for true intelligence in the physical world; world models based on JEPA are the way forward. Yann LeCun, Meta's former Chief AI Scientist and Turing Award winner, shares why he left to found AI for the real world at his new company. He argues reality is too complex, continuous, and noisy for autoregressive token prediction. World models that predict action consequences and enable planning through search are essential. "LLMs are intrinsically unsafe... they cannot predict the consequences of their actions." He contrasts this with JEPA's non-generative approach that learns abstract representations. LeCun sees five-year world domination for this paradigm in robotics, industrial control, and healthcare.
查看原文 →