从扑克到 DeepMind：穆斯塔法·苏莱曼谈 AI 起源

From Poker to DeepMind: Mustafa Suleyman on the Origins of AI

穆斯塔法·苏莱曼 Mustafa Suleyman · This Week in Startups · 2023-08-18 · 约 75 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

穆斯塔法·苏莱曼分享与戴密斯·哈萨比斯在扑克桌上创立 DeepMind 的故事，并回顾 AI 发展的漫长历程。

Mustafa Suleyman shares the story of founding DeepMind with Demis Hassabis over a poker game, and reflects on the long journey of AI development.

要点 · TL;DR

• DeepMind 被谷歌收购是为了获得大规模算力投资。
  DeepMind was acquired by Google to secure massive compute investment.
• 硬件扩展而非仅算法推动 AI 每年 10 倍进步。
  Hardware scaling, not just algorithms, drives AI progress at 10x per year.
• 个人 AI 必须与用户对齐并付费，以免成为产品。
  Personal AI must be user-aligned and paid for to avoid being the product.

核心观点 · Key points

• DeepMind 被谷歌收购是出于对大规模算力投资的迫切需求。
  DeepMind's acquisition by Google was driven by the need for massive compute investment.
• 硬件创新而非仅算法推动了 AI 进步，算力每年增长 10 倍。
  Hardware innovation, not just algorithms, has driven AI progress with 10x compute per year.
• 个人 AI 必须与用户对齐并付费，以避免用户成为产品。
  Personal AI must be user-aligned and paid for to avoid being the product.
• AI 将导致大规模就业替代，需要再培训和全民基本收入。
  AI will cause significant job displacement, requiring retraining and UBI.
• 谷歌的保守策略和广告模式可能难以应对 AI 的说服性。
  Google's conservative approach and ad model may struggle with AI's persuasive nature.

反共识 · Contrarian takes

• 由于算力 Scaling（规模扩张），AI 进步比人们想象的更可预测。
  AI progress is more predictable than people think due to compute scaling.
• 谷歌在 2020 年就拥有 ChatGPT 级别的技术，但因保守而未发布。
  Google had ChatGPT-level tech in 2020 but chose not to launch due to conservatism.
• AI 直接给出答案可能去中介化开放网络，损害内容创作者。
  The open web may be disintermediated by AI giving direct answers, harming content creators.
• AI 导师将教育民主化，全球竞争取决于动力而非特权。
  AI tutors will democratize education, making global competition about motivation, not privilege.
• 微软同时投资 OpenAI 和 Inflection 体现了平台之平台的策略。
  Microsoft's investment in both OpenAI and Inflection shows a platform-of-platforms strategy.

本期章节 · Chapters（共 26）

• 开场闲聊与介绍 Opening banter and introduction
• DeepMind 的起源 Origins of DeepMind
• 扑克诈唬策略 Poker Bluffing Strategy
• Open Phone 广告 Open Phone Ad
• 赛后讨论与 DeepMind 起源 Post-Tournament Discussion and Origins of DeepMind
• 早期困境与向 Peter Thiel 推销 Early struggles and the pitch to Peter Thiel
• 谷歌的方法与图像识别现状 Google's approach and the state of image recognition
• 转向强化学习与 Larry Page 的兴趣 Transition to reinforcement learning and Larry Page's interest
• 融资与早期里程碑 Raising investment and early milestones
• 产品速度与 Crowdbotics 广告 Product velocity and Crowdbotics ad
• 硬件演进与计算扩展 Hardware evolution and compute scaling
• 硬件限制与数据中心设计 Hardware constraints and data center design
• DeepMind 收购与独立之争 DeepMind acquisition and independence debate
• DeepMind 招募精英人才 Recruiting elite talent at DeepMind
• 后悔卖给谷歌吗？ Regret selling to Google?
• DeepMind 在谷歌内部的实用成果 DeepMind's practical wins inside Google
• 为何 DeepMind 不碰搜索和 YouTube Why DeepMind didn't touch Search and YouTube
• 模型质量随时间退化 Model quality degradation over time
• Carta 广告 Carta advertisement
• 谷歌对 AI 的谨慎态度 Google's cautious approach to AI
• 谷歌面临的挑战：广告模式与开放网络 Challenges for Google: ad model and open web
• 个人 AI 愿景 Vision for Personal AI
• 资金与硬件 Funding and Hardware
• AI 风险：失业与安全 Risks of AI: Job Displacement and Safety
• 失业与转型 Job displacement and transition
• 书籍介绍与预购 Book Introduction and Pre-order

阅读全文双语转录 →