AI 前沿：DeepSeek、Claude Opus 4.5 与全球竞赛

AI State of the Art: DeepSeek, Claude Opus 4.5, and the Global Race

内森·兰伯特 Nathan Lambert · Lex Fridman 播客 · 2026-01-31 · 约 265 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

Lex Fridman 与 Sebastian Raschka 和 Nathan Lambert 探讨 AI 最新突破，涵盖 DeepSeek 时刻、开源模型以及中美实验室的竞争格局。

Lex Fridman discusses the latest AI breakthroughs with Sebastian Raschka and Nathan Lambert, covering the DeepSeek moment, open-weight models, and the competitive landscape between US and Chinese labs.

要点 · TL;DR

• 缩放定律仍然有效，但重点从预训练转向后训练和推理缩放。
  Scaling laws persist but shift from pre-training to post-training and inference scaling.
• RLVR 是后训练中实现推理和工具使用的关键突破。
  RLVR is the key post-training breakthrough for reasoning and tool use.
• 中国开源模型因宽松许可而流行，但美国模型在质量上仍领先。
  Chinese open-weight models gain traction via permissive licenses, but US models lead in quality.

核心观点 · Key points

• 缩放定律在预训练中仍然成立，但低垂果实已被摘取；当前进步主要来自后训练和推理缩放。
  Scaling laws still hold for pre-training, but low-hanging fruit is gone; gains now come from post-training and inference scaling.
• 基于可验证奖励的强化学习（RLVR）是后训练中最大的突破，实现了推理和工具使用。
  RLVR (reinforcement learning with verifiable rewards) is the biggest post-training breakthrough, enabling reasoning and tool use.
• DeepSeek 等中国开放权重模型因宽松许可和强劲性能而受欢迎，但美国模型在质量上仍领先。
  Chinese open-weight models like DeepSeek are popular due to permissive licenses and strong performance, but US models still lead in quality.
• Transformer 架构仍占主导；创新在于注意力机制调整（MLA、GQA）和混合专家模型。
  Transformer architecture remains dominant; innovations are in attention tweaks (MLA, GQA) and mixture of experts.
• 数据质量和筛选对预训练至关重要；合成数据和精心混合可提高效率。
  Data quality and curation are critical for pre-training; synthetic data and careful mixing improve efficiency.

反共识 · Contrarian takes

• 预训练缩放并未消亡，只是目前性价比低于推理缩放和强化学习。
  Pre-training scaling is not dead; it's just less cost-effective than inference scaling and RL for now.
• 中国公司发布开放权重模型是为了扩大影响力，而非单纯商业目的；这一趋势可能持续数年。
  Chinese companies release open-weight models to gain influence, not just for business; they may continue for years.
• Qwen 模型上的 RLVR 提升可能部分源于数据污染，而非纯粹的能力解锁。
  RLVR gains on Qwen models may be partly due to data contamination, not pure skill unlocking.
• LLM 生成的摘要常丢失语气和核心洞见；人类撰写的内容保留独特价值。
  LLM summaries often lose voice and core insights; human-written content retains unique value.
• 凡事依赖 LLM 可能阻碍学习；刻意挣扎和离线学习对成为专家仍然必不可少。
  Using LLMs for everything may hinder learning; deliberate struggle and offline study are still essential for expertise.

本期章节 · Chapters（共 101）

• 引言与嘉宾 Introduction and Guests
• DeepSeek 时刻与国际竞争 DeepSeek Moment and International Competition
• 中国开源模型的可持续性 Sustainability of Open-Weight Models from China
• 中国开源模型与市场激励 Chinese open-weight models and market incentives
• DeepSeek 的定位与竞争 DeepSeek's position and competition
• 中国企业的不同激励 Different incentives among Chinese companies
• 用户习惯与多订阅 User habits and multiple subscriptions
• 2025-2026 赢家预测 Predictions for 2025 and 2026 winners
• 模型使用偏好 Model usage preferences
• GPT-5.2 长上下文与中国模型 GPT-5.2 long context and Chinese models
• 用 LLM 编程：工具与经验 Programming with LLMs: tools and experiences
• 书籍与从零学习 Books and learning from scratch
• 用 LLM 的阅读习惯 Reading habits with LLMs
• 开源 LLM 格局 Open LLM models landscape
• 工具使用及其意义 Tool Use and Its Significance
• 开源模型爆发原因 Reasons for Open Model Explosion
• 开源模型的有趣想法 Interesting Ideas from Open Models
• KV 缓存与注意力机制调整 KV Cache and Attention Mechanism Tweaks
• Transformer 架构概览 Transformer Architecture Overview
• 从 GPT-2 到如今的演进 Evolution from GPT-2 to Today
• 后训练与系统级改进 Post-training focus and system-level improvements
• 预训练、后训练与推理的缩放定律 Scaling laws across pre-training, post-training, and inference
• 预训练缩放与成本 Pre-training scaling and cost
• 后训练与发布周期 Post-training and release cycles
• 看好所有缩放方向 Bullish on all scaling directions
• 预训练、中训练与后训练定义 Pre-training, mid-training, and post-training definitions
• 预训练的合成数据与数据提取 Synthetic data and data extraction for pre-training
• 数据质量与算力 Data Quality vs. Compute
• 提升数据质量 Improving Data Quality
• 意外的高质量数据源 Unexpected High-Quality Data Sources
• 训练数据的保密与许可 Secrecy and Licensing of Training Data
• 数据护城河与领域缩放 Data as Moat and Domain-Specific Scaling
• LLM 摘要中的声音与洞察 Voice and Insight in LLM Summaries
• 为不同用户设计 AI Designing AI for diverse users
• 科技巨头与 AI 声誉 Big Tech and AI reputation
• 在 AI 中寻找自主性 Finding agency in AI
• 自主性与忽视 AI Agency vs ignoring AI
• 开发者对 AI 的享受度调查 Survey on developer enjoyment with AI
• 平衡享受与 AI 使用 Balancing enjoyment and AI use
• AI 作为结对程序员 AI as a pair programmer
• 延迟满足与 AI Delayed gratification and AI
• AI 学习的黄金区间 Goldilocks zone of learning with AI
• 后训练与 RLVR Post-training and RLVR
• 推理缩放与思维链 Inference Scaling and Chain-of-Thought
• 后训练配方与 RLVR Post-Training Recipe and RLVR
• 中训练与推理轨迹 Mid-training and reasoning traces
• RLHF 作为收尾 RLHF as finishing touch
• RLVR 的算力需求 Compute requirements for RLVR
• RLVR 与 RLHF 缩放 RLVR vs RLHF Scaling
• 教育与学习推荐 Education and Learning Recommendations
• 从 Transformers 库学习 Learning from Transformers Library
• 学习路径与研究机会 Learning Path and Research Opportunities
• 后训练主题概览 Overview of Post-Training Topics
• 培养研究品味 Developing research taste
• 教育的短暂数字窗口 The brief digital window in education
• 角色训练研究的算力需求 Compute requirements for character training research
• 权衡：新颖想法与职业路径 Trade-offs: novel ideas vs. practical career paths
• 职业建议：学术界与工业实验室 Career advice: academia vs. industry labs
• 学术界与工业界 Academia vs. Industry
• 硅谷泡沫与炒作 Silicon Valley bubble and hype
• 文生图缩放与替代架构 Text-to-image scaling and alternative architectures
• 文本扩散模型与自回归模型 Text Diffusion Models vs. Auto-Regressive Models
• 工具使用的未来 Future of Tool Use
• 工具使用中的开源与闭源模型 Open vs. Closed Models in Tool Use
• 持续学习的定义与重要性 Open vs Closed Models and Tool Use
• 持续学习与上下文学习 Continual Learning Definition and Importance
• 记忆机制 Continual Learning vs In-Context Learning
• 长上下文的创新 Memory Mechanisms
• 世界模型与 LLM Innovations in Long Context
• 机器人技术与缩放 World Models and LLMs
• 机器人挑战与安全 Robotics and Scaling
• AGI 与 ASI 时间线 Robotics challenges and safety
• AI 与软件开发未来 Timelines to AGI and ASI
• 用 LLM 的软件工程 AI and Software Development Future
• 经济影响与工具使用 Software engineering with LLMs
• 需要新想法 Economic Impact and Tool Use
• 能力平台期与经济影响 Need for New Ideas
• LLM 用于学习与个性化建议 Plateauing Capabilities vs. Economic Impact
• AI 初创公司的整合 LLMs for learning vs. personalized advice
• IPO 与 AI 公司未来 Consolidation in AI startups
• API 市场竞争 IPOs and future of AI companies
• Meta 路径与 Llama 未来 API Market Competition
• Llama 演进与内部问题 Meta's Path and Llama's Future
• 扎克伯格角色与开源辩论 Llama's Evolution and Internal Issues
• 社区反弹及其影响 Mark Zuckerberg's Role and Open Source Debate
• 美国填补 Llama 空白 Community Backlash and Its Impact
• Adam 项目：起源与使命 US Filling the Gap Left by Llama
• 政府与行业支持 The Adam Project: Origin and Mission
• AI 行动计划与开源 Government and Industry Support
• 教育与人才培养 The AI Action Plan and Open Source
• 开源模型与全球访问 Education and Talent Development
• 集中化与开放性 Open Source Models and Global Access
• 英伟达主导与硬件趋势 Centralization vs. Openness
• 英伟达创新与黄仁勋角色 Nvidia's Dominance and Hardware Trends
• 网络与算力缩放 Nvidia's innovation and Jensen's role
• 神经网络作为突破 Networking and Compute Scaling
• 未来世界：机器人与设备 Neural Networks as Breakthrough
• 失业的社会影响 Future World: Robots and Devices
• 对未来的希望 Social impact of job loss
• 后奇点场景中的人与 AI Hope for the future
• Human vs AI in a post-singularity scenario Human vs AI in a post-singularity scenario

阅读全文双语转录 →