世界模型：下一场 AR 革命的推动力

World Models: Enabler for the Next AR Revolution

杨立昆 Yann LeCun · 苏黎世联邦理工学院计算机视觉与几何组 · 2026-06-09 · 约 59 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

探讨世界模型如何弥合机器学习与人类智能之间的差距，实现在复杂真实环境中的快速适应和零样本学习。

Exploring how world models can bridge the gap between machine learning and human-like intelligence, enabling rapid adaptation and zero-shot learning in complex real-world environments.

要点 · TL;DR

• 智能是快速学习新任务的能力，而非知识积累。
  Intelligence is quick task learning, not knowledge accumulation.
• 世界模型应在抽象表征空间训练，而非作为生成模型。
  World models should be trained in abstract representation space, not as generative models.
• LLM 无法实现类人智能，因为数据量远不足。
  LLMs cannot achieve human-like intelligence due to insufficient data volume.

核心观点 · Key points

• 智能是快速学习新任务的能力，而非知识或技能的积累。
  Intelligence is the ability to learn new tasks quickly, not accumulation of knowledge or skills.
• 世界模型应在抽象表征空间中训练，而非作为生成模型或视频生成器。
  World models should be trained in abstract representation space, not as generative models or video generators.
• 联合嵌入预测架构（JEPA）在学习视觉表征方面优于生成模型。
  Joint embedding predictive architectures (JEPA) outperform generative models for learning visual representations.
• 视觉自监督学习需要通过信息最大化或正则化防止坍缩。
  Self-supervised learning for vision requires preventing collapse via information maximization or regularization.
• 通过优化（能量最小化）进行推理比通过固定层前向传播更强大。
  Inference by optimization (energy minimization) is more powerful than forward propagation through fixed layers.

反共识 · Contrarian takes

• 大语言模型无法实现类人智能，因为文本数据量远不及视觉经验。
  LLMs cannot achieve human-like intelligence because text data volume is vastly insufficient compared to visual experience.
• AGI 是无意义的术语；人类智能是专门化且自适应的，而非通用的。
  AGI is a meaningless term; human intelligence is specialized and adaptive, not general.
• 学术界应放弃大语言模型研究，因为他们无法做出有意义的贡献。
  Academics should abandon LLM research as they cannot contribute meaningfully.
• 强化学习样本效率极低，应尽量减少使用，转而采用基于观察的学习。
  Reinforcement learning is horribly sample-inefficient and should be minimized in favor of observation-based learning.
• 视频生成系统不是世界模型；它们是根本不同的问题。
  Video generation systems are not world models; they are fundamentally different problems.

本期章节 · Chapters（共 28）

• 世界模型与莫拉维克悖论 World Models and the Moravec Paradox
• 皮亚杰与智能本质 Piaget and the Nature of Intelligence
• 人类如何学习：观察与直觉物理 How Humans Learn: Observation and Intuitive Physics
• 婴儿与机器对重力的学习 Learning about gravity in infants vs. machines
• 智能定义与 AGI 怀疑论 Definition of intelligence and AGI skepticism
• 扩展 LLM 不会产生类人智能 Scaling LLMs won't lead to human-like intelligence
• 智能系统的属性：推理模式 Properties of intelligent systems: inference mode
• LLM 推理与世界模型推理对比 Contrasting LLM inference with world model inference
• 能量最小化的推理与规划架构 Architecture for Reasoning and Planning via Energy Minimization
• 分层规划与训练世界模型 Hierarchical Planning and Training World Models
• 视频预测的挑战 Challenges of Video Prediction
• 联合嵌入预测架构（JEPA） Joint Embedding Predictive Architecture (JEPA)
• 测量信息内容的挑战 Challenges in Measuring Information Content
• 基于能量的模型框架 Energy-Based Models as a Framework
• 防止能量模型崩溃 Preventing Collapse in Energy-Based Models
• 信息最大化与表征学习 Information Maximization and Representation Learning
• 抽象表征的必要性 The Need for Abstract Representations
• 科学中的抽象 Abstractions in Science
• 世界模型与模拟器 World Models vs Simulators
• 复杂系统的世界模型 World Models for Complex Systems
• Sigreg：各向同性高斯正则化 Sigreg: Isotropic Gaussian Regularization
• 高斯化与世界模型 Gaussianization and World Models
• 自监督学习：C-Reg 与蒸馏方法 Self-Supervised Learning: C-Reg and Distillation Methods
• DINO 及其应用 DINO and Its Applications
• V-Jeppa 与常识 V-Jeppa and Common Sense
• 自监督学习与常识 Self-supervised learning and common sense
• 放弃生成模型与 LLM Abandon generative models and LLMs
• 问答：表征空间中的护栏 Q&A: Guardrails in representation space

阅读全文双语转录 →