杨立昆讨论他的新创业公司 Advanced Machine Intelligence，强调开放研究和世界模型，并与大型 AI 实验室日益封闭的趋势形成对比。
Yann LeCun discusses his new startup Advanced Machine Intelligence, emphasizing open research and world models, and contrasts it with the growing secrecy in big AI labs.
要点 · TL;DR
世界模型必须在抽象表征空间预测，而非像素级。 World models must predict in abstract representation space, not pixel-level.
仅靠 LLM 无法实现人类级 AI，它们缺乏现实基础。 LLMs alone cannot achieve human-level AI; they lack grounding in reality.
开放研究对真正进步至关重要；保密导致自欺欺人。 Open research is essential for real progress; secrecy leads to self-deception.
核心观点 · Key points
世界模型必须在抽象表征空间中进行预测，而非像素级别，以处理高维噪声数据。 World models must predict in abstract representation space, not pixel-level, to handle high-dimensional noisy data.
仅靠 LLM 无法实现人类级 AI；它们缺乏现实基础，且无法处理连续数据。 LLMs alone cannot achieve human-level AI; they lack grounding in reality and fail with continuous data.
AI 安全可以通过设计带有内置约束的目标驱动架构来实现，而不仅仅是微调。 AI safety can be achieved by designing objective-driven architectures with built-in constraints, not just fine-tuning.
通用智能是神话；人类智能是专门化的，AI 将逐步进步。 General intelligence is a myth; human intelligence is specialized, and AI will progress gradually.
开放研究和发表对于真正进步至关重要；保密会导致自欺欺人。 Open research and publication are essential for real progress; secrecy leads to self-deception.
硅谷当前专注于 Scaling LLM 的 AI 单一文化是误导性的。 The current AI monoculture in Silicon Valley focusing on scaling LLMs is misguided.
反共识 · Contrarian takes
视频数据比文本更具冗余性，使得自监督学习对世界模型更高效。 Video data is more redundant than text, making self-supervised learning more efficient for world models.
世界模型不应是重现所有细节的模拟器；它们应使用抽象表征。 World models should not be simulators that reproduce all details; they should use abstract representations.
达到狗级智能比从狗到人类更难；语言只是一个小附加模块。 Getting to dog-level intelligence is harder than from dog to human; language is a small add-on.
对比学习方法（如 Barlow Twins 和 VICReg）在表征学习上优于像素级预测。 Contrastive learning methods like Barlow Twins and VICReg outperform pixel-level prediction for representation learning.
具有自主性和规划能力的 AI 系统可以通过约束优化实现内在安全，而不仅仅是护栏。 AI systems with agency and planning can be intrinsically safe through constraint optimization, not just guardrails.
目前最好的开源 AI 模型来自中国，挑战了美国在开放性上的主导地位。 The best open-source AI models are currently Chinese, challenging US dominance in openness.
本期章节 · Chapters（共 37）
引言与新公司祝贺Introduction and congratulations on new startup
开放研究理念与 AMI 计划Open research philosophy and AMI's plans
AI 进展与缺失环节Progress and missing parts in AI
避免像素预测与表征坍缩Avoiding pixel-level prediction and representation collapse
SimReg 与 JEPASimReg and JEPA
缺失环节与数据质量Missing parts and data quality
理想化世界模型Idealized world model
科学中的抽象层次Abstraction levels in science
LLM 的理解与复述Understanding vs. Regurgitation in LLMs
学习物体恒存与基础物理Learning Object Permanence and Basic Physics
从抽象环境与游戏中学习Learning from Abstract Environments and Games
人类不擅棋类游戏Humans are terrible at chess and Go
莫拉维克悖论与游戏 AI 停滞Moravec's paradox and game AI stagnation
AGI 时间线：乐观与悲观Timelines and optimism vs pessimism on AGI
人类级智能时间线Timeline to human-level intelligence
颠覆性影响与安全担忧Destabilizing impacts and safety concerns
AI 安全的喷气引擎类比Jet engine analogy for AI safety
约束 LLM 输出空间Constraining output space in LLMs
Meta 的 AI 组织与 Alex Wang 角色Meta's AI organization and Alex Wang's role
Meta 组织变革Meta's Organizational Changes
世界模型与其他公司World Models and Other Companies
硅谷从众心理与欧洲创业Silicon Valley Herd Mentality and European Startups
逃离硅谷单一文化Escaping the monoculture of Silicon Valley