像 Tapestry 这样的开源 AI 平台是主权和多样性的关键。 Open-source AI platforms like Tapestry are key to sovereignty and diversity.
核心观点 · Key points
LLM 擅长语言，但不是通往人类级智能的路径；它们缺乏世界模型和规划能力。 LLMs are great for language but not a path to human-level intelligence; they lack world models and planning.
世界模型预测行动的后果；通过搜索进行规划对于智能体至关重要。 World models predict consequences of actions; planning via search is essential for intelligent agents.
JEPA 通过在表示空间中进行预测来学习抽象表示，而非像素——更高效且可扩展。 JEPA learns abstract representations by predicting in representation space, not pixels—more efficient and scalable.
LLM 本质上不安全，因为它们无法预测自身行动的后果；具有世界模型的目标驱动型 AI 更安全。 LLMs are intrinsically unsafe because they cannot predict consequences of their actions; objective-driven AI with world models is safer.
需要像 Tapestry 这样的开源 AI 平台来实现主权和多样性，类似于 Linux 取代专有 Unix。 Open-source AI platforms like Tapestry are needed for sovereignty and diversity, similar to Linux replacing proprietary Unix.
AI 的未来是拥有世界模型和规划的目标驱动型系统，而不仅仅是扩展 LLM。 The future of AI is objective-driven systems with world models and planning, not just scaling LLMs.
反共识 · Contrarian takes
LLM 本质上不安全，因为它们无法预测行动的后果。 LLMs are intrinsically unsafe because they cannot predict consequences of actions.
VLA（视觉-语言-行动）模型现在被视为失败。 VLA (vision-language-action) models are now seen as a failure.
扩展 LLM 不会带来人类级 AI；需要范式转变。 Scaling LLMs will not lead to human-level AI; a paradigm shift is needed.
当前对 LLM 安全性的担忧被夸大；真正的危险是滥用，而非 AI 接管。 Current LLM safety concerns are overblown; real danger is bad usage, not AI takeover.
像 OpenAI 这样的闭源 AI 公司是今天的 Sun Microsystems，注定会被开放平台超越。 Closed-source AI companies like OpenAI are the Sun Microsystems of today, destined to be overtaken by open platforms.
模仿学习和合成数据不是高效学习的路径；世界模型能实现零样本任务解决。 Imitation learning and synthetic data are not the path to efficient learning; world models enable zero-shot task solving.
本期章节 · Chapters（共 31）
引言与对 LLM 的怀疑Introduction and skepticism of LLMs
什么是 AMI？What is AMI?
Meta 研究与产品的矛盾Tension between research and product at Meta
Meta 战略转变与研究影响Meta's Shift in Strategy and Impact on Research
世界模型对比 VLA 与 LLMWorld Models vs. VLA and LLMs
通过预测学习世界模型Learning World Models by Prediction
机器人演示与模仿学习Robotics Demos and Imitation Learning
泛化与世界模型Generalization and World Models
合成数据与视频模型Synthetic Data and Video Models
数据效率与规模之争Data efficiency and scaling debate
世界模型的应用Applications of world models
未来里程碑与全球主导Future milestones and world domination
世界模型范式转变时间表Timeline for world model paradigm shift
Tapestry：主权 AI 开放平台Tapestry: open platform for sovereign AI assistants
开源与专有模型之争Open Source vs Proprietary Models
LLM 的优势与局限LLM Strengths and Limitations
与共同获奖者关于 AI 风险的分歧Divergence from Co-recipients on AI Risk
杰夫的顿悟与计算Jeff's Epiphany and Calculation
约书亚的担忧与商业动机Yoshua's Concerns and Commercial Motives
新架构与 LLM 的安全性Safety of New Architectures vs LLMs
LLM 局限与目标驱动 AILLM limitations and objective-driven AI
领导 Meta FAIR 的反思Reflections on leading FAIR at Meta
从研究到产品的管道Research to product pipeline
短期压力与研究文化Short-term pressure and research culture
对年轻研究者的影响Impact on younger researchers
离开 Meta 与在 FAIR 的角色Leaving Meta and Role at FAIR
Meta 对 LLM 的支持与转变Support and Shift to LLMs at Meta
规模收购与 LLM 聚焦Scale Acquisition and LLM Focus
开源 Llama 2 的内部辩论Internal Debate on Open-Sourcing Llama 2
自监督学习的一致观点与转变Consistent View and Change of Mind on Self-Supervised Learning
联合嵌入架构与防止崩溃Joint Embedding Architecture and Collapse Prevention