从个人 AI 到 Codex：OpenAI 的内部演变

From Personal AI to Codex: Inside OpenAI's Evolution

格雷格·布罗克曼 Greg Brockman · Tetragrammaton · 2026-02-28 · 约 197 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

OpenAI 联合创始人探讨 ChatGPT 在亲密个人应用中的使用、他自己对 Codex 的晚期采用，以及从解雇事件中吸取的教训。

OpenAI co-founder discusses how ChatGPT is used for intimate personal applications, his own late adoption of Codex, and the lessons learned from the firing incident.

要点 · TL;DR

• 规模法则依然成立；更大的模型带来可预测的提升。
  Scaling laws still hold; bigger models yield predictable improvements.
• 在 AI 驱动的经济中，计算资源将成为一项基本人权。
  Compute access will become a basic human right in an AI-driven economy.
• 基于强化学习的后训练是教会模型复杂任务的关键。
  Post-training with reinforcement learning is key to teaching models complex tasks.

核心观点 · Key points

• 缩放定律持续有效；更大的模型带来可预测的提升。
  Scaling laws continue unabated; bigger models yield predictable improvements.
• AI 将驱动算力经济；算力获取是未来的基本权利。
  AI will become a compute-powered economy; compute access is a future basic right.
• 后训练结合强化学习是教会模型复杂任务的关键。
  Post-training with reinforcement learning is key to teaching models complex tasks.
• OpenAI 的使命是构建惠及所有人的 AGI，需要商业与非营利双轨并行。
  OpenAI's mission is to build AGI that benefits all, requiring both business and nonprofit arms.
• 解雇事件教会我们及早直面冲突的重要性。
  The firing event taught the importance of addressing conflict early and directly.

反共识 · Contrarian takes

• ChatGPT 并非实体，而是可由用户偏好塑造的多元智能体集合。
  ChatGPT is not an entity but a plurality of agents shaped by user preferences.
• AI 模型不必擅长讲笑话；当前重点首先是知识工作。
  AI models do not need to be perfect at jokes; focus is on knowledge work first.
• 真正的瓶颈是算力供给，而非模型架构或数据。
  The real bottleneck is compute supply, not model architecture or data.
• 通过多样性实现韧性比单一对齐的超级智能更重要。
  Resilience through diversity is more important than a single aligned superintelligence.
• 埃隆·马斯克曾要求完全控制 OpenAI；拒绝是基于原则而非个人。
  Elon Musk wanted full control of OpenAI; the refusal was principled, not personal.

本期章节 · Chapters（共 75）

• ChatGPT 个人使用 Personal Use of ChatGPT
• 模型改进与扩展 Model Improvements and Scaling
• 对 Sam Altman 的看法 Thoughts on Sam Altman
• 解雇事件 The Firing Incident
• 解雇与即时反应 The Firing and Immediate Reaction
• 董事会政变的周末 The weekend of the board coup
• OpenAI 的冲突解决与决策 Conflict resolution and decision-making at OpenAI
• Elon 的离开与质疑 Elon's departure and skepticism
• AG1 广告 AG1 advertisement
• 与微软的合作 Partnership with Microsoft
• 与微软的合作 Partnership with Microsoft
• ChatGPT 的个性与情感 ChatGPT's personality and emotions
• 突破速度与意外能力 Rate of breakthroughs and unexpected capabilities
• AI 对信息类型的优势 AI's strength with information types
• Greg 的编码与领导 Greg's continued coding and leadership
• 早期构建与首批用户 Early days of building and first users
• ChatGPT 发布与惊喜 ChatGPT launch and surprise
• Sora 与世界模型 Sora and world models
• 扩展定律与科学发现 Scaling Laws and Scientific Discovery
• AI 优先组织与公司转型 AI-forward organization and company transformation
• AI 在生物学与推理中的应用 AI in biology and reasoning
• OpenAI 的财务方面 Financial side of OpenAI
• 计算作为基本人权 Compute as a basic human right
• 一个改变的信念 A belief that changed
• OpenAI 与 Stripe：不同的公司 DNA OpenAI vs Stripe: different company DNA
• 回顾 OpenAI 的创立 Looking back at OpenAI's founding
• AI 格局：合并与赢家 AI landscape: mergers and winners
• 对齐与价值观多样性 Alignment and Diversity of Values
• 无监督情感神经元 Unsupervised Sentiment Neuron
• AI 与语言学 AI and Linguistics
• 记忆与学习 Memory and Learning
• AI 作为现状破坏者 AI as a Destructor of Status Quo
• Greg 的 MIT 经历 Greg's MIT Experience
• MIT 的知识产权与黑客文化 MIT's IP wealth and hacker culture
• 离开哈佛与父母支持 Leaving Harvard and parental support
• 在北达科他州乡村长大 Growing up in rural North Dakota
• 早期学业加速 Early academic acceleration
• 高中修大学课程 University courses in high school
• 即兴表演与 AI Improv and AI
• 第一个网站：反向图灵测试 First Website: Reverse Turing Test
• 图灵的愿景与计算的作用 Turing's vision and the role of compute
• 早期 AI 怀疑与图灵测试启发 Early AI skepticism and the Turing test inspiration
• 编程视觉与象棋的挑战 The challenge of programming vision vs. chess
• 计算民主化 Democratization of compute
• OpenAI 的第一天 OpenAI's first day
• OpenAI 的创立 Founding of OpenAI
• 与 Ilya Sutskever 合作 Working with Ilya Sutskever
• Greg 如何进入 Stripe How Greg ended up at Stripe
• 童年与身份 Childhood and identity
• 离开 Stripe Leaving Stripe
• 化学竞赛与书籍 Chemistry Competition and the Book
• 撰写量子力学书籍 Writing a book on quantum mechanics
• 编码与其他活动的比较 How coding compares to other activities
• 数学作为宇宙的基石 Math as the fabric of the universe
• 编码心流状态的感觉 The feeling of coding in flow state
• 调试与可观测性 Debugging and Observability
• 氛围编码及其影响 Vibe Coding and Its Impact
• 还有理由用旧方式编码吗？ Is There Still a Reason to Code the Old Way?
• 编码会变得不必要吗？ Will Coding Become Unnecessary?
• AI 结果中的人类责任 Human accountability in AI outcomes
• 技术的愿景与影响 Vision and Impact of Technology
• 代理 AI 及其应用 Agentic AI and its applications
• ChatGPT 与语言模型的机制 Mechanics of ChatGPT and language models
• 模型架构与训练系统 Model architecture and training system
• 预训练与后训练 Pre-training and post-training
• 预训练概述 Pre-training Overview
• 从数据到实验：AlphaGo 的教训 From Data to Experiment: The AlphaGo Lesson
• AI 泡沫与计算需求 AI Bubble and Compute Demand
• 计算作为基本人权 Compute as a basic human right
• 竞争对手与差异化 Competitors and differentiation
• 基准测试与实际使用 Benchmarks and real-world usage
• 基准测试作为代理指标 Benchmarks as proxy metrics
• OpenAI 的研究与部署 Research and deployment at OpenAI
• 对 Clawbot/Open Claw 的看法 Thoughts on Clawbot / Open Claw
• 关于向世界发布 AI On releasing AI to the world

阅读全文双语转录 →