扩展定律与通往 AGI 之路：2026-2027 年

Scaling Laws and the Path to AGI by 2026-2027

达里奥·阿莫迪 Dario Amodei · Lex Fridman 播客 · 2024-11-11 · 约 315 分钟 · 原视频 ↗

本期速览 · Overview

Dario Amodei 讨论扩展假说、他在百度的早期观察，并基于能力的快速增长预测 AGI 可能在 2026-2027 年到来。

Dario Amodei discusses the scaling hypothesis, his early observations at Baidu, and predicts AGI could arrive by 2026-2027 based on rapid capability growth.

要点 · TL;DR

扩展定律预测 2026-2027 年实现人类级 AI，无根本障碍。
Scaling laws predict human-level AI by 2026-2027 with no fundamental blockers.
后训练（RLHF、宪法 AI）与预训练同等重要。
Post-training (RLHF, constitutional AI) is as crucial as pre-training.
机械可解释性揭示神经网络内部清晰概念以保障安全。
Mechanistic interpretability reveals clear concepts inside neural networks for safety.

核心观点 · Key points

缩放定律很可能在 2026-2027 年带来人类水平的 AI，且没有根本性障碍。
Scaling laws will likely lead to human-level AI by 2026-2027, with no fundamental blockers remaining.
后训练正变得与预训练同等重要，使用基于人类反馈的强化学习（RLHF）、宪法 AI 和合成数据。
Post-training is becoming as important as pre-training, using RLHF, constitutional AI, and synthetic data.
机械可解释性是一种有前景的安全方法，揭示了神经网络内部的清晰概念。
Mechanistic interpretability is a promising safety approach, revealing clear concepts inside neural networks.
负责任扩展策略（RSP）使用 if-then 触发器，在模型达到能力阈值时施加安全措施。
Responsible Scaling Policy (RSP) uses if-then triggers to impose safety measures as models reach capability thresholds.
AI 风险包括灾难性滥用（网络、生物）和自主性风险；监管应精准且有针对性。
AI risks include catastrophic misuse (cyber, bio) and autonomy risks; regulation should be surgical and targeted.
人才密度胜过人才数量；一小群顶尖人才胜过大型稀释的组织。
Talent density beats talent mass; a small team of top performers outperforms a large diluted organization.

反共识 · Contrarian takes

模型不会随时间变笨；用户感知变化源于熟悉度和对提示的敏感性。
Models don't get dumber over time; user perception shifts due to familiarity and sensitivity to prompts.
基于人类反馈的强化学习（RLHF）不会让模型更聪明；它弥合了人类偏好与模型输出之间的差距。
RLHF doesn't make models smarter; it bridges the gap between human preferences and model output.
宪法 AI 使用书面宪法的自我对弈，减少对人类反馈的依赖。
Constitutional AI uses self-play with a written constitution, reducing reliance on human feedback.
计算机使用能力不会从根本上增加风险；它是现有能力的窗口。
Computer use capability doesn't fundamentally increase risk; it's an aperture for existing abilities.
遏制未对齐的 AI 比构建对齐的模型更糟；可解释性是未来安全的关键。
Containing unaligned AI is worse than building aligned models; interpretability is key for future safety.
开放心态和新视角比经验对突破性 AI 研究更重要。
Open-mindedness and fresh eyes matter more than experience for breakthrough AI research.

本期章节 · Chapters（共 110）

0. 引言与缩放定律 Introduction and Scaling Laws
1. AGI 时间线与担忧 Timeline to AGI and Concerns
2. 对话背景 Conversation Context
3. 缩放将持续 Scaling will continue
4. AI 性能天花板 Ceilings in AI performance
5. 缩放定律的潜在极限 Potential limits to scaling laws
6. 领域竞争者 Competitors in the field
7. 登顶策略竞赛 Race to the Top Strategy
8. 机制可解释性 Mechanistic Interpretability
9. Claude 模型变体 Claude Model Variants
10. 模型命名与哲学 Model Naming and Philosophy
11. 编程能力提升 Programming ability improvement
12. Claude 3.5 Opus 发布时间线 Claude 3.5 Opus release timeline
13. 模型版本管理挑战 Model versioning challenges
14. 用户体验与模型性格 User experience and model character
15. 用户对模型退化的感知 User perception of model degradation
16. 模型性格与拒绝行为 Model character and refusal behavior
17. 模型行为的权衡 Trade-offs in model behavior
18. 收集用户反馈 Gathering user feedback
19. Claude 4 与缩放 Claude 4 and scaling
20. 负责任缩放政策与 ASL 等级 Responsible scaling policy and ASL levels
21. 两大风险类别：灾难性滥用与自主风险 Two major risk categories: catastrophic misuse and autonomy risks
22. ASL3 与 ASL4 触发条件与时间线 ASL3 and ASL4 triggers and timelines
23. 条件承诺与风险规划 If-then commitments and risk planning
24. 计算机使用：工作原理 Computer Use: How It Works
25. 未来改进与安全 Future Improvements and Safety
26. 风险与负责任发布 Risks and Responsible Release
27. 训练中的测试与沙盒 Testing and sandboxing during training
28. 监管与 AI 安全 Regulation and AI safety
29. 监管与极化 Regulation and Polarization
30. Dario 在 OpenAI 的背景与缩放假说 Dario's Background at OpenAI and Scaling Hypothesis
31. 缩放与安全的灵感与早期工作 Inspiration and Early Work on Scaling and Safety
32. 离开 OpenAI 的原因 Reasons for Leaving OpenAI
33. Anthropic 对 AI 安全的不完美追求 Anthropic's Imperfect Pursuit of AI Safety
34. 人才密度与人才规模 Talent Density vs Talent Mass
35. 优秀 AI 研究员的品质 Qualities of a Great AI Researcher
36. 优秀 AI 研究员的品质 Qualities of a great AI researcher
37. 给有志 AI 研究者的建议 Advice for aspiring AI researchers
38. 后训练与秘方 Post-training and secret sauce
39. RLHF 为何有效 Why RLHF works
40. RLHF 与模型智能 RLHF and Model Intelligence
41. 预训练与后训练成本 Cost of Pre-training vs Post-training
42. 宪法 AI Constitutional AI
43. 谁定义宪法 Who Defines the Constitution
44. OpenAI 模型规范与 Anthropic OpenAI Model Spec and Anthropic
45. 竞争动态与积极实践 Competitive Dynamics and Positive Practices
46. 文章《爱与优雅的机器》 The Essay 'Machines of Love and Grace'
47. 强大 AI 的定义 Definition of Powerful AI
48. 关于进展速度的两种极端观点 Two Extreme Views on Speed of Progress
49. AI 快速部署的障碍 Obstacles to rapid AI deployment
50. 怀疑论观点：缓慢的生产力增长 Skeptical view: slow productivity gains
51. 变革驱动力：愿景者与竞争 Drivers of change: visionaries and competition
52. 时间线估计：5-10 年 Timeline estimate: 5-10 years
53. AGI 的命名与时间线 Naming and timeline for AGI
54. AGI 在生物学与医学中的早期步骤 Early steps of AGI in biology and medicine
55. 生物学作为 AI 的杠杆点 Biology as a Leverage Point for AI
56. AI 作为生物学研究生 AI as Grad Students in Biology
57. 用 AI 改进临床试验 Improving Clinical Trials with AI
58. 编程作为快速变化的领域 Programming as a Fast-Changing Field
59. 模型闭环能力与快速编程改进 Model's ability to close the loop and rapid programming improvement
60. 比较优势与程序员角色的变化 Comparative advantage and the changing role of programmers
61. IDE 的未来与 Anthropic 的角色 Future of IDEs and Anthropic's role
62. 在强大 AI 世界中的意义 Meaning in a world of powerful AI
63. 模拟世界中的意义 Meaning in a Simulated World
64. 意义与社会中的 AI Meaning and Society with AI
65. 鼓励完整文章 Encouraging the Full Essay
66. Amanda 的哲学背景 Amanda's Philosophical Background
67. 从哲学到技术 AI 工作 From Philosophy to Technical AI Work
68. 塑造 Claude 的性格与个性 Crafting Claude's Character and Personality
69. 平衡诚实与谄媚 Balancing honesty and sycophancy
70. 对话者的理想特质 Desirable traits for a conversationalist
71. 代表多元视角 Representing multiple perspectives
72. 影响他人与知识谦逊 Influencing People and Intellectual Humility
73. 通过对话映射 Claude 的行为 Mapping Claude's Behavior Through Conversations
74. 创造力与提示工程 Creativity and Prompt Engineering
75. 对模型的同理心 Empathy for the Model
76. 与模型互动 Interacting with Models
77. 后训练的魔力 The Magic of Post-Training
78. 宪法 AI 与 Claude 的性别化 Constitutional AI and Gendering Claude
79. 宪法 AI 概述 Constitutional AI Overview
80. 系统提示与争议话题 System Prompts and Controversial Topics
81. 系统提示设计与演变 System prompt design and evolution
82. 系统提示变化与用户感知 System prompt changes and user perception
83. 收集用户反馈 Gathering user feedback
84. 解决用户对说教与道歉行为的投诉 Addressing user complaints about moralizing and apologetic behavior
85. 模型行为与道歉倾向 Model Behavior and Apologetic Tendencies
86. 经验对齐与理论对齐 Empirical vs. Theoretical Alignment
87. 最优失败率 Optimal Failure Rate
88. AI 中的意识与痛苦 Consciousness and Suffering in AI
89. AI 关系与情感依恋 AI relationships and emotional attachment
90. 与 AGI 的第一次对话 First conversation with AGI
91. 测试 AGI 能力 Testing AGI capabilities
92. 人类独特之处 What makes humans special
93. 机制可解释性导论 Introduction to Mechanistic Interpretability
94. 定义机制可解释性 Defining Mechanistic Interpretability
95. 神经网络的普遍性 Universality in Neural Networks
96. 神经网络的普遍性 Universality in Neural Networks
97. 特征与电路 Features and Circuits
98. 线性表示假说 Linear Representation Hypothesis
99. 叠加假说与多义性 Superposition Hypothesis and Polysemanticity
100. 多义性与对单义特征的需求 Polysemanticity and the need for monosemantic features
101. 理解特征与自动可解释性 Understanding features and automated interpretability
102. 将可解释性扩展到 Claude 3 Scaling interpretability to Claude 3
103. 字典学习与大模型中的特征 Dictionary Learning and Features in Large Models
104. 神经网络中的暗物质 Dark matter in neural networks
105. 微观与宏观可解释性 Microscopic vs macroscopic interpretability
106. 神经网络的解剖学类比 Anatomy analogy for neural networks
107. 相对于神经科学的优势 Advantages over neuroscience
108. 安全与美作为目标 Safety and beauty as goals
109. 神经网络中的美与复杂性 Beauty and Complexity in Neural Networks

阅读全文双语转录 →