超级智能将至——但要「以人为本」

Superintelligence is near — but make it humanist

穆斯塔法·苏莱曼 Mustafa Suleyman · Decoder · 2026-06-08 · 约 72 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

微软 AI 负责人谈 AI 伙伴、边界，与「以人为本的超级智能」。

Microsoft AI’s chief on companions, boundaries, and humanist superintelligence.

要点 · TL;DR

• 微软必须自研前沿模型，避免过度依赖 OpenAI。
  Microsoft must build its own frontier models to avoid over-reliance on OpenAI.
• 超智能即将到来，得益于计算和数据对数线性扩展。
  Superintelligence is near due to log-linear scaling of compute and data.
• AI 的目的是让人们更健康、更快乐、更有能力。
  AI's purpose is to make people healthier, happier, and more capable.

核心观点 · Key points

• 微软长期必须拥有自己的前沿模型，不能仅依赖 OpenAI。
  Microsoft must own its own frontier models long-term, not depend solely on OpenAI.
• 超级智能即将到来，由算力和数据的对数线性扩展驱动。
  Superintelligence is coming soon, driven by log-linear scaling of compute and data.
• 蒸馏是短期胜利；真正的前沿进步需要原创研究。
  Distillation is a short-term win; true frontier progress requires original research.
• AI 的目的是让人们更健康、更快乐、更有能力——这是检验标准。
  AI's purpose is to make people healthier, happier, and more capable—that's the test.
• 任务会被自动化，而非整个工作；效率往往创造更多工作。
  Tasks will be automated, not entire jobs; efficiency often creates more work.
• 消费级 AI 价值真实但不足；企业级展现出更清晰的产品市场契合度。
  Consumer AI value is real but insufficient; enterprise shows clearer product-market fit.

反共识 · Contrarian takes

• 白领任务将在 12-18 个月内被 AI 完全自动化，而非工作岗位。
  White-collar tasks will be fully automated by AI in 12-18 months, not jobs.
• 模型没有意识；赋予其意识是危险且拟人化的。
  Models are not conscious; attributing consciousness is dangerous and anthropomorphic.
• 奇点还需数十年，过于推测；应关注近期进展。
  The singularity is decades away and too speculative; focus on near-term progress.
• 微软的自给自足使命意味着独立构建前沿模型，而非依赖 OpenAI。
  Microsoft's self-sufficiency mission means building frontier models independently, not relying on OpenAI.
• 消费者对 AI 的反感真实存在；价值交换尚不明确。
  Consumer AI antipathy is real; the value exchange is not yet clear enough.
• 手机将被解构；到 2030 年代 AI 将分布于环境设备中。
  The phone will be disintermediated; AI will live across ambient devices by 2030s.

本期章节 · Chapters（共 27）

• 引言与嘉宾 Introduction and Guest
• 批评 Anthropic 与意识主张 Critique of Anthropic and Consciousness Claims
• 消费硬件与企业分发传闻 Rumors about consumer hardware and enterprise distribution
• 纳德拉的英特尔类比与伙伴关系演变 Satya Nadella's Intel analogy and partnership evolution
• 前沿模型预算与自给自足 Budget approval for frontier model and self-sufficiency
• 微软独立性与 OpenAI 合作 Microsoft's Independence and Partnership with OpenAI
• 决策框架与组织节奏 Decision-Making Framework and Organizational Rhythm
• 定义超级智能、AGI 与奇点 Defining Superintelligence vs AGI vs Singularity
• 编码与其他领域的验证 Coding vs Other Domains for Validation
• 模型训练方法差异 Differences in Model Training Approach
• 不蒸馏现有模型的原因 Reason for Not Distilling Existing Models
• 蒸馏的法律与知识产权问题 Legal and IP Concerns with Distillation
• 循环 Transformer 与全栈创新 Loop Transformer and Full-Stack Innovation
• 对蒸馏与数据抓取的挫败感 Understanding Frustration About Distillation and Data Scraping
• 数据策展与付费 Data Curation and Payment
• 企业 AI 与消费 AI：价值与反感 Enterprise vs Consumer AI: Value and Antipathy
• AI 目标：健康与幸福 Purpose of AI: health and happiness
• 工作与任务的区别 Jobs vs tasks distinction
• 任务与 AI 的长期影响 Tasks and Long-term Impact of AI
• 社会影响与进步 Societal Impact and Progress
• 企业采用与令牌使用 Enterprise Adoption and Token Usage
• 形态实验与计算架构 Form Factor Experimentation and Computing Architecture
• 可编程徽章的酷想法 Cool idea for a programmable badge
• LLM 是通往 AGI 或超级智能之路吗？ Are LLMs the path to AGI or superintelligence?
• 模型的意识与活力 Consciousness and aliveness of models
• Anthropic 的拟人化及其危险 Anthropic's anthropomorphism and its dangers
• 结束语 Closing remarks

阅读全文双语转录 →