AI 算力超级周期与 Claude Sonnet 4.5 的诞生

The AI Compute Super Cycle and the Making of Claude Sonnet 4.5

肖尔托·道格拉斯 Sholto Douglas · The MAD Podcast · 2025-10-02 · 约 70 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

Anthropic 研究员 Sholto Douglas 探讨 AI 加速进步、算力超级周期，以及 Sonnet 4.5 如何通过强化学习成为最佳编程模型。

Anthropic researcher Sholto Douglas discusses the accelerating pace of AI progress, the compute super cycle, and how Sonnet 4.5 became the best coding model through reinforcement learning.

要点 · TL;DR

• AI 进步是指数级的，所有基准测试都在快速提升。
  AI progress is exponential, with rapid improvements across all benchmarks.
• 强化学习是新的扩展轴，让模型学会说‘我不知道’等技能。
  Reinforcement learning is a new scaling axis, enabling models to learn skills like uncertainty.
• 长时间代理连贯性（30 小时）大幅扩展模型能力，例如构建完整聊天应用。
  Long agentic coherence (30 hours) dramatically expands model capabilities, e.g., building a full chat app.

核心观点 · Key points

• AI 的进步是指数级的；任何可衡量的基准都在快速提升。
  Progress in AI is exponential; any measurable benchmark improves rapidly.
• 强化学习是与预训练并列的新 Scaling 轴，使模型能学会说“我不知道”等技能。
  Reinforcement learning is a new scaling axis alongside pre-training, enabling models to learn skills like saying 'I don't know'.
• 更长的智能体连贯性（如 30 小时）极大扩展了模型能力，例如构建完整的聊天应用。
  Longer agentic coherence (e.g., 30 hours) dramatically expands what models can achieve, like building a full chat app.
• 由于可验证性和可重放性，编程是 AI 最易处理的领域，推动近期经济影响。
  Coding is the most tractable domain for AI due to verifiability and replayability, driving near-term economic impact.
• 当前的 LLM 训练流程还很原始，每个组件都有巨大的改进空间。
  Current LLM training pipelines are primitive and have huge room for improvement across every component.

反共识 · Contrarian takes

• 尽管有人声称平台期到来，但 AI 进步正在加速，而非放缓。
  Despite claims of a plateau, AI progress is accelerating, not slowing down.
• 最简单的 RL 方法对语言模型最有效；像 AlphaGo 那样的复杂方法常常失败。
  The simplest RL methods work best for language models; complex approaches like AlphaGo's often fail.
• 长链 token 推理从 RL 中自然涌现，无需显式结构或提示。
  Reasoning in long chains of tokens emerges naturally from RL without explicit structure or prompting.
• 莫拉维克悖论“有点假”；机器人技术的进步主要是数据和反馈循环问题。
  Moravec's paradox is 'a little bit fake'; robotics progress is mostly a data and feedback loop issue.
• Transformer 足以实现 AGI；不需要根本性的新架构。
  Transformers are sufficient for AGI; no fundamentally new architecture is needed.

本期章节 · Chapters（共 30）

• 引言与发布节奏 Introduction and pace of releases
• Sholto 背景与 Anthropic 之旅 Sholto's background and journey to Anthropic
• YouTube 影响与向他人学习 YouTube's impact and learning from others
• Gemini 早期与推理栈构建 Early days at Gemini and building inference stack
• 转投 Anthropic 与文化差异 Move to Anthropic and cultural differences
• AI 研究中的品味概念 The concept of taste in AI research
• 苦涩教训与简单性正则化 The Bitter Lesson and Simplicity Regularization
• AGI 时间线与研究理念差异 AGI timelines and research ethos differences
• Anthropic 为何聚焦编码 Why Anthropic focuses on coding
• Sonnet 4.5 与 SWE-bench 表现 Sonnet 4.5 and SWE-bench performance
• Cursor 与智能体编码兴起 Cursor and the rise of agentic coding
• 智能体 30 小时在做什么 What the agent does for 30 hours
• 长期连贯性作为突破 Long-term coherence as breakthrough
• 30 小时任务示例 Examples of 30-hour tasks
• Claude AI 演示与模型演进 Claude AI demo and model progression
• 实现更长自主运行的突破 Breakthroughs enabling longer autonomous operation
• 进展速度与突破 Pace of progress and breakthroughs
• 从预训练转向强化学习 Shift from pre-training to RL
• RL 与下一词预测对抗幻觉 RL vs Next-Token Prediction for Hallucination
• 语言模型 RL 终于奏效 RL on Language Models Finally Works
• 测试时计算与 RL 重叠 Test-Time Compute and RL Overlap
• 为何 LLM 的 RL 突破在当下 Why RL on LLMs Breakthrough Now
• 通过 LLM 和 RL 实现 AGI AGI via LLMs and RL
• 反论与辩论 Counter-thesis and debate
• 平台期还是指数级进步 Plateau or exponential progress
• 累积人类设计 vs.LLM 训练流程 Accumulated human design vs. LLM training pipeline
• Sonnet 4.5 与 GDP 评估 Sonnet 4.5 and GDP eval
• 建议：杠杆与机器人学 Advice: leverage and robotics
• 结束语 Closing remarks
• 尾声 Outro

阅读全文双语转录 →