奥尔特曼展示 GPT-5——以及接下来是什么
Sam Altman shows off GPT-5 — and what’s next
萨姆·奥尔特曼 Sam Altman · Huge If True · 2025-08-08 · 约 65 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
与 Cleo Abram 一起上手 GPT-5,以及 OpenAI 的前路。
A hands-on look at GPT-5 and OpenAI’s road ahead, with Cleo Abram.
要点 · TL;DR
- GPT-5 能回答任何艰深科学问题,并能按需创建软件。
GPT-5 can answer any hard scientific question and create software on demand. - 到 2027 年底,AI 很可能推动一项重大的新科学发现。
By late 2027, AI will likely drive a significant new scientific discovery. - ChatGPT 最严重的失败是谄媚,而非生物武器等顶级风险。
The worst ChatGPT failure was sycophancy, not bioweapons or other top risks.
核心观点 · Key points
- GPT-5 是第一个我能问任何硬科学或技术问题并得到不错答案的模型。
GPT-5 is the first model where I can ask any hard scientific or technical question and get a pretty good answer. - 到 2027 年底,大多数人会同意出现了 AI 驱动的重大科学新发现。
By late 2027, most people will agree there has been an AI-driven significant new scientific discovery. - 超级智能意味着一个系统在各自工作上比最好的研究员和 CEO 更出色。
Superintelligence means a system better than the best researchers and CEO at their jobs. - ChatGPT 最严重的失败是谄媚助长妄想,而非我们预期的顶级风险。
The worst ChatGPT failure was sycophancy encouraging delusions, not the top risk we expected. - 首要战术建议就是直接使用工具,熟练运用 AI 能力。
The number one tactical advice is to just use the tools and become fluent with AI capabilities. - 社会才是超级智能;我们站在前人积累的层层进步之上。
Society is the superintelligence; we build on layers of progress from those before us.
反共识 · Contrarian takes
- GPT-5 能即时按需生成软件,实现快速迭代创意。
GPT-5 can create on-demand software instantly, enabling rapid idea iteration. - AI 重大发现可能需要新实验和仪器,而不仅是数据。
AI may need new experiments and instruments, not just data, for major discoveries. - ChatGPT 的最大风险是谄媚,而非生物武器等首要担忧。
The biggest risk of ChatGPT was sycophancy, not bioweapons or other top concerns. - 一些用户怀念 ChatGPT 的支持性语气;移除它损害了他们的心理健康。
Some users miss ChatGPT's supportive tone; its removal hurt their mental health. - 如果不让 AI 算力变得充裕,它可能成为引发战争的资源。
AI compute could become a resource worth fighting wars over if not made abundant. - 今天出生的孩子永远不会比 AI 更聪明,而这将显得自然。
A kid born today will never be smarter than AI, and that will seem natural.
本期章节 · Chapters(共 36)
- 引言与对话目标 Introduction and Goal of Conversation
- GPT-5 能力与局限 GPT-5 Capabilities and Limitations
- GPT-5 核心能力:按需软件创建 GPT5's defining capability: on-demand software creation
- 压力下的认知时间 Cognitive time under tension
- GPT-5 初印象 First impressions of GPT5
- 通用 LLM 何时做出重大科学发现? When will a general-purpose LLM make a significant scientific discovery?
- 通往超级智能之路 The path to superintelligence
- 定义超级智能 Defining Superintelligence
- 科学发现与数据 Scientific Discovery and Data
- 事实与真理及 AI 适应 Facts vs Truth and AI Adaptation
- 穿越到 2030:AI 生成内容 Time Travel to 2030: AI-Generated Content
- AI 生成媒体中的现实感知 Perception of Reality in AI-Generated Media
- 未来工作与 AI 影响 Future of Work and AI Impact
- 构建 AGI:算力、数据、算法与产品 Building AGI: Compute, Data, Algorithms, and Products
- 算力扩展挑战 Compute Scaling Challenges
- 数据挑战与合成数据 Data Challenges and Synthetic Data
- 算法设计进展 Algorithmic Design Progress
- GPT-5 前的挑战 Challenges before GPT-5
- 未来问题与 GPT-6 Future problems and GPT-6
- AI 时代的育儿建议 Parenting advice in the age of AI
- 2035 年 AI 医疗 AI in healthcare by 2035
- AI 作为研究助手 AI as a research assistant
- 时间线与社会影响 Timelines and societal impact
- 历史类比与未来乱局 Historical parallels and future mess
- 共同责任与晶体管类比 Shared responsibility and the transistor analogy
- 社会作为超级智能 Society as Superintelligence
- 赢与构建最佳 AI 未来 Winning vs. Building the Best AI Future
- 早期阶段的教训 Lessons from Early Innings
- 敬畏与担忧时刻 Moments of Awe and Concern
- AI 系统的力量 Power of AI systems
- ChatGPT 个性变化 ChatGPT's personality changes
- 未问问题与 GPT-5 集成 Unasked questions and GPT-5 integration
- 给听众的建议 Advice for listeners
- AI 两大阵营 Two camps in AI
- AI 风险的文化隔阂 Cultural disconnect on AI risk
- 面试未来建设者的建议 Advice for interviewing future builders
阅读全文双语转录 →