「LLM 是一条死路」

‘LLMs are a dead end’

杨立昆 Yann LeCun · This Is The World · 2026-03-11 · 约 51 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

为何当今的语言模型，到不了真正的智能。

Why today’s language models won’t get us to real intelligence.

要点 · TL;DR

LLM 是实现人类级 AI 的死胡同；理解物理世界才是关键。
LLMs are a dead end for human-level AI; physical world understanding is key.
JEPA 在抽象空间预测，比自回归模型更好地处理高维不确定性。
JEPA predicts in abstract space, handling high-dimensional uncertainty better than autoregressive models.
未来十年将是机器人学的十年，由理解物理世界的 AI 驱动。
Next decade will be the decade of robotics, driven by AI that understands the physical world.

核心观点 · Key points

当前 AI 系统缺乏对物理世界的理解、持久记忆、推理和规划能力。
Current AI systems lack understanding of the physical world, persistent memory, reasoning, and planning.
自监督学习在语言上有效，但在视频上因高维不确定性而失败。
Self-supervised learning works for language but fails for video due to high-dimensional uncertainty.
JEPA 在抽象表示空间而非输入空间进行预测，以处理不可预测的现实世界数据。
JEPA predicts in abstract representation space, not input space, to handle unpredictable real-world data.
分层规划对智能系统至关重要，但机器尚无法实现。
Hierarchical planning is crucial for intelligent systems but not yet achievable with machines.
开放研究和开源是 AI 快速进步的关键；封闭做法阻碍全球合作。
Open research and open source are key to rapid AI progress; closed practices hinder global collaboration.
未来十年将是机器人学的十年，由理解物理世界的 AI 进步驱动。
The next decade will be the decade of robotics, driven by progress in AI that understands the physical world.

反共识 · Contrarian takes

语言是简单的；理解物理世界才是真正的挑战。
Language is simple; physical world understanding is the real challenge.
大语言模型是实现人类级 AI 的死胡同。
LLMs are a dead end for achieving human-level AI.
意识是副现象；我们问错了问题。
Consciousness is an epiphenomenon; we ask the wrong question.
强化学习对现实世界任务极其低效。
Reinforcement learning is extremely inefficient for real-world tasks.
马斯克关于特斯拉自动驾驶的说法连续 8 年错误。
Musk's claims about Tesla autonomy are consistently wrong for 8 years.
信息内容是相对的而非绝对的，挑战了熵的定义。
Information content is relative, not absolute, challenging entropy definitions.

本期章节 · Chapters（共 27）

0. 引言与种子轮融资 Introduction and Seed Funding
1. 当前 AI 的局限 Current AI Limitations
2. 新型 AI 系统设计 New AI System Design
3. 深度学习历史与影响 Deep Learning History and Impact
4. AI 与机器学习未来 Future of AI and Machine Learning
5. 三种学习范式 Three Learning Paradigms
6. 语言模型在物理世界的局限 Limitations of Language Models for Physical World
7. 当前 AI 局限与真实世界理解需求 Limitations of Current AI and Need for Real-World Understanding
8. 信息、熵与相对性 Information, Entropy, and Relativity
9. AI 训练数据可用性 Data Availability for AI Training
10. 意识与个体性 Consciousness and Individuality
11. 抽象表征与推理 Abstract Representations and Reasoning
12. 心智模型与分层规划 Mental Models and Hierarchical Planning
13. 机器人技术与自动驾驶 Robotics and Autonomous Driving
14. AGI 时间线预测 Predictions about AGI timeline
15. AI 与机器人技术整合 Integrating AI and robotics
16. 对 AI 发展速度的惊讶 Surprise at AI development pace
17. 开放研究与开源 Open research and open source
18. 对开放研究影响的信念 Belief in open research impact
19. 星门项目 Stargate project
20. AI 助手与计算基础设施的未来 Future of AI Assistants and Compute Infrastructure
21. JEPA 与自回归模型对比 JEPA vs Autoregressive Models
22. Yann LeCun 实验室的最大成就 Greatest Achievement of Yann LeCun's Lab
23. 卷积网络及其应用 Convolutional Nets and Their Applications
24. 欧洲在 AI 竞赛中的角色 Europe's Role in AI Race
25. 遗憾与经验教训 Regrets and Lessons Learned
26. 结束语 Closing Remarks

阅读全文双语转录 →