格雷格·布罗克曼谈会自我改进的 AI

Greg Brockman on AI that improves itself

格雷格·布罗克曼 Greg Brockman · Big Technology · 2026-04-01 · 约 73 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 总裁谈自我改进的 AI、规模扩张，以及公司的走向。

OpenAI’s president on self-improving AI, scaling, and where the company is headed.

要点 · TL;DR

• AGI 几年内到来，AI 将处理几乎所有智力型计算机任务。
  AGI expected within a few years; AI will handle most intellectual computer tasks.
• 因算力限制，OpenAI 优先发展推理模型而非世界模型。
  OpenAI prioritizes reasoning models over world models due to compute limits.
• 算力是主要瓶颈，需求远超供给。
  Compute is the main bottleneck; demand far exceeds supply.

核心观点 · Key points

• AGI（通用人工智能）将在几年内到来，AI 能处理几乎任何智力性计算机任务。
  AGI will arrive within a couple years, with AI handling almost any intellectual computer task.
• 由于算力有限，OpenAI 优先发展 GPT 推理模型，而非 Sora 等世界模型。
  OpenAI prioritizes GPT reasoning models over world models like Sora due to limited compute.
• 超级应用将整合聊天、编程和浏览，成为个人 AGI（通用人工智能）助手。
  The super app will unify chat, coding, and browsing into a personal AGI assistant.
• 预训练的改进会放大下游收益；Scaling（规模扩张）仍然至关重要。
  Pre-training improvements multiply downstream gains; scaling remains crucial.
• 算力是主要瓶颈；需求远超供给。
  Compute is the primary bottleneck; demand far outstrips supply.

反共识 · Contrarian takes

• OpenAI 最可怕的时刻是 ChatGPT 发布后感到“我们赢了”——这是一种危险的自满。
  The scariest moment at OpenAI was after ChatGPT launch, feeling 'we won'—a dangerous complacency.
• Codex 不仅面向程序员，而是面向所有人；它关乎解决问题，而非编写代码。
  Codex is for everyone, not just coders; it's about solving problems, not writing code.
• 数据中心用水量微乎其微——例如阿比林设施的年用水量相当于一个家庭。
  Data center water usage is negligible—e.g., Abilene uses as much as a household per year.
• AI 将增进人际联系，腾出时间加深纽带，而非取代工作。
  AI will increase human connection and free up time for deeper bonds, not replace jobs.
• OpenAI 的 1100 亿美元算力押注并非鲁莽，而是对必然需求的审慎回应。
  OpenAI's $110B compute bet is not YOLO; it's a calculated response to inevitable demand.
• 经济增长将由 AI 算力驱动，而不仅仅是模型智能。
  The economy's growth will be driven by AI compute, not just model intelligence.

本期章节 · Chapters（共 38）

• AGI 时间线与 ChatGPT 后氛围 AGI Timeline and Post-ChatGPT Vibe
• 战略转向超级应用 Strategic Shift to Super App
• 迪士尼类比与算力限制 Disney Analogy and Compute Constraints
• 文本推理 vs 世界模型之争 Bet on text reasoning vs world models
• 超级应用愿景 Debate on world models vs text reasoning
• 超级应用愿景 Vision of the super app
• 追赶现实软件工程 Super App Vision
• OpenAI 氛围转变：从赢家到挑战者 Catching up in real-world software engineering
• Spud：新预训练基础模型 Vibe shift at OpenAI: from winner to challenger
• AI 使用的质变与量变 Spud: a new pre-trained base model
• 起飞阶段与自动化 AI 研究员 Qualitative and Quantitative Shifts in AI Usage
• 起飞与进展的风险 Takeoff Phase and Automated AI Researcher
• 安全与对齐 Risks of Takeoff and Progress
• AGI 定义与进展 Safety and Alignment
• 2025 年 12 月转折点 AGI Definition and Progress
• 模型如何实现飞跃 December 2025 Inflection Point
• 从程序员专用 Codex 到人人可用 How Models Made the Leap
• Open Claw 与智能体愿景 From Codex for Coders to Codex for Everyone
• 成为智能体舰队 CEO Open Claw and the Vision for Agents
• 人类能动性与问责 Becoming CEO of a Fleet of Agents
• 模型演进下一步 Human agency and accountability
• 为何尚未发生？ Next steps in model evolution
• 数学 vs 文字：对失业的担忧 Why hasn't this happened yet?
• AI 对人类连接的影响 Math vs words: concern about job displacement
• 智能体用例与预训练必要性 AI's impact on human connection
• 推理时代对英伟达 GPU 的需求 Agentic use cases and pre-training necessity
• 预训练何时足够？ Need for Nvidia GPUs in inference era
• 数据中心投资背后的数学 When is pre-training enough?
• 构建新类别的确定性 Math behind data center investment
• 收入流与企业采用 Certainty in building new category
• 基础设施押注与算力限制 Revenue streams and enterprise adoption
• 智能体与软件的亲身体验 Infrastructure bets and compute constraints
• AI 对个人的积极影响 Personal experience with agents and software
• 对就业和数据中心的担忧 AI's positive impact on individuals
• 政治捐款与单一议题关注 Concerns about jobs and data centers
• 应对 AI 恐惧 Political donations and one-issue focus
• 为未来准备的建议 Addressing fears about AI
• Advice for preparing for the future Advice for preparing for the future

阅读全文双语转录 →