In this episode, we welcome Yann LeCun, a pioneer in AI, to discuss his journey, from early neural networks to world models.
要点 · TL;DR
杨立昆认为 AI 的未来在于非生成式世界模型（如 JEPA），而非大语言模型。 Yann LeCun argues AI's future lies in non-generative world models like JEPA, not LLMs.
推理需要在潜在空间中进行操作，而非自回归的逐词预测。 Reasoning requires latent space manipulation, not autoregressive token prediction.
开放、联邦式的 AI 对于防止文化同质化和权力集中至关重要。 Open, federated AI is crucial to prevent cultural homogenization and power concentration.
核心观点 · Key points
AI的未来在于世界模型，而非生成模型，以理解物理世界。 The future of AI lies in world models, not generative models, for understanding the physical world.
世界模型需要像JEPA这样的非生成架构来处理连续、嘈杂环境中的不确定性。 World models require non-generative architectures like JEPA to handle uncertainty in continuous, noisy environments.
推理不能通过自回归词元预测实现；它需要在潜在空间中进行操作。 Reasoning cannot be achieved through autoregressive token prediction; it requires latent space manipulation.
没有预测性世界模型的智能体系统存在根本缺陷，无法可靠工作。 Agentic systems without predictive world models are fundamentally flawed and cannot work reliably.
开放、联邦式的AI对于文化多样性和民主主权至关重要，避免权力集中。 Open, federated AI is essential for cultural diversity and democratic sovereignty, avoiding concentration of power.
反共识 · Contrarian takes
生成式AI应被放弃用于世界建模；它不适合现实世界的预测。 Generative AI should be abandoned for world modeling; it is unsuitable for real-world prediction.
LLM只是知识库，而非智能；它们缺乏发明新方法的能力。 LLMs are just knowledge repositories, not intelligent; they lack the ability to invent new recipes.
预测下一个词元不是推理；真正的推理发生在潜在空间，而非符号中。 Predicting the next token is not reasoning; true reasoning happens in latent space, not in symbols.
对比学习方法效率低下；蒸馏和正则化（如SIGReg）更优越。 Contrastive learning methods are ineffective; distillation and regularization like SIGReg are superior.
在视频游戏上训练世界模型是错误的；现实世界数据需要非生成方法。 Training world models on video games is misguided; real-world data requires non-generative approaches.
AI的主要危险不是失业，而是少数公司导致的文化同质化和偏见。 The main danger of AI is not unemployment but the homogenization of culture and bias from a few companies.
本期章节 · Chapters（共 39）
开场Introduction
猜嘉宾游戏Game to guess the guest
初次见面与编程语言First meeting and programming languages
Léon Bottou与LISP解释器Léon Bottou and the LISP interpreter
反向传播与论文Backpropagation and thesis
早期职业生涯与贝尔实验室Early Career and Bell Labs
学术界与工业界关系Academic-Industry Relationship
FAIR巴黎与法国生态FAIR Paris and French Ecosystem
早期工具与平台Early Tools and Platforms
可视化与沟通Visualization and Communication
贝尔实验室的支持Bell Labs Support
Sun 4的故事The Sun 4 Story
Meta时代与世界模型Meta Era and World Models
JEPA与世界模型简介Introduction to JEPA and World Models
直观物理与世界模型Intuitive Physics and World Models
知识与智能之别Knowledge vs. Intelligence
随机鹦鹉与潜在空间Stochastic Parrots and Latent Space
潜在空间中的推理Reasoning in Latent Space
LLM推理的未来Future of Reasoning in LLMs
潜在空间再注入与扩展Latent Space Re-injection and Scaling
世界模型与潜在空间投影World Models and Latent Space Projection
崩溃问题与早期工作The Collapse Problem and Early Work
对比方法与崩溃Contrastive Methods and Collapse
高斯正则化与信息最大化Gaussian regularization and information maximization
高斯投影与流形学习Gaussian projections and manifold learning
训练数据与世界模型Training data and world models
引言与现象学模型Introduction and Phenomenological Models
Amilab的全球布局与巴黎根基Amilab's Global Presence and Paris Roots
人才与法国的地缘政治原因Talent and Geopolitical Reasons for France
投资者多样性Investor Diversity
Ami Labs与发音Ami Labs and Pronunciation
Tapestry与联邦AITapestry and Federated AI
Tapestry的使命与正交性Tapestry's Mission and Orthogonality