Yann LeCun 讨论了专有 AI 集中的危险、自回归大语言模型的局限性，以及为什么开源 AI 对美好未来至关重要。
Yann LeCun discusses the dangers of proprietary AI concentration, the limitations of auto-regressive LLMs, and why open source AI is crucial for a good future.
要点 · TL;DR
自回归 LLM 缺乏世界理解和规划能力，仅限于系统 1。 Autoregressive LLMs lack world understanding and planning, limiting them to System 1.
联合嵌入架构（JEPA）是学习抽象世界模型的关键。 Joint embedding architectures (JEPA) are key to learning abstract world models.
开源 AI 防止权力集中，支持多样化应用。 Open source AI prevents power concentration and enables diverse applications.
核心观点 · Key points
自回归大语言模型无法达到人类水平的智能，因为缺乏对世界的理解、记忆、推理和规划能力。 Autoregressive LLMs cannot achieve human-level intelligence due to lack of world understanding, memory, reasoning, and planning.
联合嵌入预测架构（JEPA）对于从感官数据中学习抽象世界表征至关重要。 Joint embedding predictive architectures (JEPA) are essential for learning abstract world representations from sensory data.
开源 AI 平台对于多样性和避免权力集中在少数公司至关重要。 Open source AI platforms are crucial for diversity and avoiding concentration of power in a few companies.
未来的 AI 系统应通过在抽象表征空间中进行优化来规划答案，而非逐词生成。 Future AI systems should plan answers via optimization in abstract representation space, not token-by-token generation.
AGI 将是一个渐进的过程，而非突然的事件，并且需要数十年的研究。 AGI will be a gradual progression, not a sudden event, and will require decades of research.
反共识 · Contrarian takes
仅靠语言不足以构建世界模型；大多数知识来自感官互动，而非文本。 Language alone is insufficient for building world models; most knowledge comes from sensory interaction, not text.
预测像素的生成模型在视频上失败；在表征空间中进行联合嵌入更优。 Generative models that predict pixels fail for video; joint embedding in representation space is superior.
强化学习效率低下，应尽量减少；模型预测控制更适合规划。 Reinforcement learning is inefficient and should be minimized; model predictive control is preferred for planning.
AI 末日论者的恐惧基于错误假设；超级智能不会是一个单一事件，也不会天生具有支配性。 AI doomers' fears are based on false assumptions; superintelligence will not be a single event or inherently dominant.
大语言模型的成功归功于自监督学习，而非自回归预测；它们缺乏系统二的推理能力。 LLMs' success is due to self-supervised learning, not autoregressive prediction; they lack system-2 reasoning.
AI 中的偏见是不可避免且主观的；开源多样性是唯一解决方案，而非集中式去偏。 Bias in AI is inevitable and subjective; open source diversity is the only solution, not centralized debiasing.
本期章节 · Chapters（共 39）
专有 AI 与开源的危害Dangers of Proprietary AI and Open Source
自回归 LLM 的局限Limitations of Autoregressive LLMs
莫拉维克悖论与缺失能力The Moravec Paradox and Missing Capabilities
内部世界模型与规划Internal world model and planning
自监督学习：重建 vs 联合嵌入Self-supervised learning via reconstruction vs. joint embedding
自监督学习与抽象Self-supervised learning and abstraction
视频表征与物理合理性Video representation and physical plausibility
分层规划Hierarchical planning
LLM 与物理世界理解LLMs and physical world understanding
对自回归 LLM 和自监督学习的怀疑Skepticism on autoregressive LLMs and self-supervised learning
图灵测试与 LLM 局限Turing Test and LLM Limitations
联合嵌入 vs 生成模型理解世界Joint Embedding vs Generative Models for World Understanding
LLM 中的常识与世界知识Common sense and world knowledge in LLMs
LLM 的幻觉问题Hallucinations in LLMs
越狱与长尾分布Jailbreaking and Long-Tail Distribution
LLM 中的原始推理Primitive Reasoning in LLMs
在世界模型之上构建推理Building Reasoning on Top of World Models
AI 中的系统 1 与系统 2System 1 vs System 2 in AI
如何在 LLM 中实现系统 2How to Achieve System 2 in LLMs
未来对话系统：通过优化思考Future dialog systems: thinking by optimization
正则化与表征Regularization and Representations
放弃生成模型与强化学习Abandoning Generative Models and RL
用奖励模型训练Training with Reward Models
对 Gemini 和审查的批评Criticism of Gemini and Censorship
开源赋能多样化 AI 应用Open Source Enables Diverse AI Applications
大科技公司面临生成式 AI 挑战Big Tech's challenges with generative AI
开源模型与 Meta 的 Llama 的未来Future of Open-Source Models and Meta's Llama
智能是多维的Intelligence is multidimensional
AI 末日论者的假设错误AI doomers' assumptions are false
通过更好设计实现 AI 安全AI safety through better design
AI 作为武器 vs 核武器AI as a weapon vs. nuclear weapons
AI 末日论者的心理Psychology of AI doomers
对技术变革的恐惧Fear of Technological Change
世界模型与机器人学World Models and Robotics
AI 作为增强人类智能的工具AI as a tool to augment human intelligence
历史类比：印刷术与奥斯曼帝国Historical analogy: printing press and Ottoman Empire
对就业和未来职业的影响Impact on jobs and future professions
相信人性本善与开源 AIBelief in human goodness and open source AI