Sam Altman 谈 AGI、GPT-4 与 AI 的未来

Sam Altman on AGI, GPT-4, and the Future of AI

萨姆·奥尔特曼 Sam Altman · Lex Fridman 播客 · 2023-03-25 · 约 144 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

Sam Altman 讨论了 AGI 早期受到的嘲笑、ChatGPT 的关键时刻，以及 RLHF 在让 AI 变得有用中的作用。

Sam Altman discusses the early mockery of AGI, the pivotal moment of ChatGPT, and the role of RLHF in making AI useful.

要点 · TL;DR

• RLHF 用极少数据对齐模型，使其更可用。
  RLHF aligns models with little data, making them vastly more usable.
• 早期公开部署有助于社会适应并发现缺陷。
  Early public deployment helps society adapt and discover flaws.
• AI 应放大人类意志，而非被视为生物。
  AI should amplify human will, not be treated as a creature.

核心观点 · Key points

• 基于人类反馈的强化学习（RLHF）用极少的数据就能让模型对齐人类偏好，大幅提升可用性。
  RLHF aligns models to human preferences with remarkably little data, making them vastly more usable.
• 早期公开部署 AI 能让社会逐步适应，并帮助发现能力与缺陷。
  Deploying AI early and publicly allows society to adapt and helps discover both capabilities and flaws.
• 对齐与能力紧密相关；更好的对齐通常带来更强的能力。
  Alignment and capability are closely linked; better alignment often leads to better capabilities.
• 智能和能源的成本将大幅下降，从经济和政治上改变社会。
  The cost of intelligence and energy will dramatically fall, transforming society economically and politically.
• AI 应是放大人类意志的工具，而非生物；将其视为生物是危险的。
  AI should be a tool that amplifies human will, not a creature; treating it as a creature is dangerous.

反共识 · Contrarian takes

• GPT-4 并非 AGI（通用人工智能）；它缺乏真正通用智能的深度和广度。
  GPT-4 is not an AGI; it lacks the depth and generality of true general intelligence.
• 近期最大的 AI 风险是虚假信息和经济冲击，而非超级智能的对齐问题。
  The biggest near-term AI risks are disinformation and economic shocks, not superintelligent alignment.
• OpenAI 的利润上限结构对于抵制资本主义压力、优先考虑安全至关重要。
  OpenAI's capped-profit structure is crucial to resist capitalist pressures and prioritize safety.
• AI 不会消除人类工作的需求；人们将找到新的创造性表达形式。
  AI will not eliminate the need for human work; people will find new forms of creative expression.
• AI 带来的变化速度将超过制度适应能力，正如硅谷银行挤兑事件所示。
  The speed of change from AI will outpace institutional adaptation, as seen with the SVB bank run.

本期章节 · Chapters（共 56）

• 引言：被误解的兽人 Introduction: The misunderstood orc
• 主持人介绍：关键时刻 Host introduction: The critical moment
• GPT-4概览与转折点 GPT-4 overview and pivotal moment
• ChatGPT与RLHF ChatGPT and RLHF
• 预训练数据与数据整理 Pre-training data and data curation
• 预测模型行为与科学理解 Predicting model behavior and scientific understanding
• ChatGPT对话格式与挑战 ChatGPT's dialog format and struggles
• 两个问题：模型弱点与公开构建 Two separate issues: model weaknesses and building in public
• 细微之处与小问题的重要性 Nuance and the importance of small issues
• GPT-4发布的安全考量 Safety considerations for GPT-4 release
• 对齐过程的见解 Insights on alignment from the process
• RLHF与社会对齐 RLHF and societal alignment
• 系统消息与可操控性 System message and steerability
• 提示工程与人类类比 Prompt engineering and human parallels
• 对编程的影响 Impact on programming
• 对话界面与迭代编码 Dialogue interfaces and iterative coding
• 系统卡与安全透明度 System card and safety transparency
• AI边界的民主进程 Democratic process for AI boundaries
• OpenAI的责任与参与 OpenAI's responsibility and involvement
• 无限制模型与言论自由绝对主义 Unrestricted model and free speech absolutism
• 公众认知与标题党压力 Public perception and clickbait pressure
• 审核工具与拒绝系统 Moderation tooling and refusal system
• 从GPT-3到GPT-4的技术飞跃 Technical leaps from GPT-3 to GPT-4
• 模型规模与100万亿迷因 Model size and the 100 trillion meme
• 从文本重建人类 Reconstructing humanity from text
• 参数数量与性能 Parameter count vs performance
• LLM作为通往AGI的路径 LLMs as path to AGI
• AI作为人类增强工具 AI as tool for human amplification
• 国际象棋类比与人类缺陷 Chess analogy and human flaws
• 积极AI轨迹与对齐担忧 Positive AI Trajectory and Alignment Concerns
• AI起飞速度与安全 AI Takeoff Speeds and Safety
• 辩论AGI与GPT-4 Debating AGI and GPT-4
• 开源风险与安全 Open source risks and safety
• 组织结构与转型 Organizational structure and transition
• OpenAI独特结构与AGI担忧 OpenAI's unique structure and AGI concerns
• 关于AGI的一致与分歧 Agreements and Disagreements on AGI
• AI开发中的压力与偏见 Pressure and Bias in AI Development
• 自我批评与谦逊 Self-Critique and Humility
• 以用户为中心的方法与旅行计划 User-Centric Approach and Travel Plans
• 对变化与AI的紧张 Nervousness About Change and AI
• 岗位替代与影响 Job Displacement and Impact
• AI对就业与工作的影响 Impact of AI on Jobs and Work
• 全民基本收入哲学 Philosophy on Universal Basic Income
• AI带来的经济与政治变革 Economic and Political Changes with AI
• 推出与撤回模型 Rolling out and pulling back models
• 定义真相与错误信息 Defining truth and misinformation
• 仇恨内容与责任 Hateful content and responsibility
• 越狱与安全 Jailbreaking and security
• 发布速度与团队文化 Shipping velocity and team culture
• 招聘优秀团队 Hiring great teams
• 微软合作 Microsoft partnership
• 萨提亚·纳德拉在微软的领导力 Satya Nadella's Leadership at Microsoft
• 硅谷银行倒闭 Silicon Valley Bank Collapse
• 与AGI互动与拟人化 Interacting with AGI and Anthropomorphism
• 将生物性投射到AI工具 Projecting creatureness onto AI tools
• 内省与意义 Introspection and meaning

阅读全文双语转录 →