2026.10.02 | Substack | LLM | How Sam Altman Uses AI

音频信息

音频标题: How Sam Altman Actually Uses AI to Run OpenAI and His Life

相关标的：undefined

议题：undefined

完整音频链接：http://mw-meeting.oss-cn-shanghai.aliyuncs.com/audio/1790911383022-How%20Sam%20Altman%20Actually%20Uses%20AI%20to%20Run%20OpenAI%20and%20-1790911331654-%5BjZh55CQwSh8%5D.m4a

音频纪要总结

Dev Day 专家访谈纪要

1. AI 对工程效率与未来工作的影响

1.1 编程模型能力快速提升后，为什么开发者数量和工作需求反而持续增长？

• 开发者从未像现在这样有能力完成更多工作，但行业和社会对开发者的要求也达到了前所未有的高度。
• AI 提升生产力后，并未使需求趋于饱和，反而激发了更多产品、服务和创造性工作的需求。
• 世界上存在庞大的潜在需求，人们对于技术产品的期望也在持续提高，因此开发者仍处于 AI 进步的核心位置。
• 这一趋势在积极意义上超出预期：AI 没有简单取代人，而是在扩大人类能够完成的工作范围。

1.2 AI 对就业的影响是否存在边界，能否延伸到其他领域？

• 部分不需要太多创造力、社会也不关心具体由谁完成的任务和岗位，未来确实会由 AI 承担。
• 但人类需要的新事物仍然非常多，因此不会缺少值得投入的新工作。
• AI 自动化的核心影响不是让工作整体消失，而是推动人们从低创造性任务转向新的需求和更高价值的工作。
• 创造人们真正需要的产品这一基本原则不会改变，但个人和小团队能够完成的工作量将发生根本变化。

1.3 为什么 AI 的发展仍然以人为中心？

• 人们并不真正关心机器本身做了什么，更关注作品背后的创作者、公司背后的人，以及理念或产品背后的主体。
• 即使大量智能体协作完成重大科学问题，公众也更关心如何理解成果、哪些人参与其中，以及成果对人类意味着什么。
• 例如，10,000 个智能体连续协作4天并解决重大问题后，外界讨论的重点仍可能是参与方式和成果解读，而不仅是机器完成了任务。
• 这说明 AI 的价值最终仍要通过人类目标、人类参与和社会意义来体现。

2. AI 时代的愿景：从“工业革命”转向“新文艺复兴”

2.1 AI 未来更像工业革命，还是更像文艺复兴？

• 工业革命从根本上以机器和规模化生产为中心，带来了经济增长和丰富的物质成果，但也让部分人成为高度专业化体系中的“齿轮”。
• 文艺复兴则更强调人的主体性、创造力，以及对人类潜力的重新认识。
• OpenAI 所期待的不是“AI 自动化一切、人类不再发挥作用”的未来，也不是人类像家猫一样由 AI 决定获得何种奖励和安排。
• 更理想的方向是：

1. 推动新一轮大规模科学发现；
2. 促进艺术创作和新公司成立；
3. 让更多人有能力开创新事物；
4. 让人们在享受经济增长的同时重新掌控生活；
5. 让个人可以按照自己的意愿安排时间和精力。

2.2 AI 是否正在改变人类对自身潜力的认识？

• 文艺复兴的重要特征之一，是人们重新连接历史思想，同时结合未来主义、科学与工程发现，进而拓展对人类潜力的想象。
• 当前围绕 AI 的讨论同样包含两个方向：

1. 对知识、文化与人类价值的重新理解；
2. 对未来科技和新能力边界的持续探索。

• AI 的重要价值不只是帮助人类完成更多任务，也包括帮助人们重新获得时间、注意力、创造空间和生活自主权。

3. 持久智能与主动智能：Dot 的产品定位和实践

3.1 什么是“持久智能”或“主动智能”？

• 核心理念是让智能体在后台全天候运行，与用户的系统和信息环境连接，主动发现并处理事务。
• 与传统聊天机器人不同，这类智能体不必等待用户明确发起请求，而是能够：

1. 持续理解用户背景；
2. 识别当前最需要关注的事项；
3. 主动筛选和整合信息；
4. 在合适的时间提醒用户；
5. 代表用户推进后续任务。

• 这一方向被认为是未来 AI 产品的重要发展领域。

3.2 Dot 如何帮助用户重新掌控时间和注意力？

• Dot 可以判断哪些事情需要用户立即关注，哪些事项可以延后，哪些任务可以直接由智能体处理。
• 其价值不是单纯提升处理速度，而是减少用户查看手机、Slack、邮件等渠道后产生的注意力中断。
• 对管理者而言，清晨通常是最适合深度思考和创造性工作的时间，但企业运营中的紧急情况又要求其随时关注信息。
• Dot 可以在紧急事务与非紧急事务之间进行分流，使用户既不会错过重大问题，也不必被大量日常信息占据注意力。
• 由此可以释放更多时间，用于创造性工作、复杂项目或陪伴家人。

3.3 Dot 在生产事故和日程管理中有哪些实际案例？

• 在 Dev Day 主题演讲期间发生了严重的 SEV 1 生产事故。
• Tibo 的 Dot 在演示开始前5分钟主动提醒：

1. 已识别用户正在参加 Dev Day 主题演讲；
2. 了解用户一直在为演示做准备；
3. 通过 Space 等系统了解了演示内容；
4. 判断该事故可能影响现场演示；
5. 主动询问是否需要尝试修复。

• 当前版本尚不能直接修复生产事故，但未来版本可能具备在生产环境中直接解决问题的能力。
• 在另一场景中，Dot 识别到用户连续忙碌一天、工作进度滞后，并在短暂空档中主动指出一项必须处理的关键事务，避免了潜在混乱。
• Dot 还可以从 Slack 中识别没有进入日历的会议，并在用户改签航班时提醒其存在行程冲突。

3.4 Dot 如何支持创意形成和持续迭代？

• 用户可以在散步、工作间隙或睡前，通过文本或语音随时记录原始想法。
• Dot 可以基于这些想法自动构建5至6个不同版本，接受反馈后继续迭代。
• 用户晚上留下较长的语音笔记后，第二天早晨不仅能看到想法已经被构建出来，Dot 还可能主动完成后续多个步骤。
• 这种方式降低了高不确定性项目的尝试成本，使过去因表达困难、执行复杂或耗时过长而被放弃的想法有机会落地。
• Dot 的意义在于形成近乎持续的“想法—构建—反馈—再构建”循环。

3.5 Dot 如何处理传统搜索无法解决的问题？

• 用户曾尝试通过 Codex 搜索一项 Slack 内容，连续调整搜索方式10至20分钟后仍未找到，最终放弃。
• Dot 注意到搜索失败后，在后台持续分析，并于次日找到目标内容。
• 原因是相关信息并非 Slack 文本，而是一张截图；Dot 对截图进行了图像识别，最终定位到正确内容。
• 这一案例体现了主动智能的核心优势：

1. 能够跨文本、图片和不同信息源处理内容；
2. 不会因为一次搜索失败而立即停止；
3. 可以在后台持续尝试；
4. 能在恰当时间将结果重新提供给用户。

• Dot 尤其适合处理 Slack、电子邮件和其他渠道构成的信息洪流。

3.6 Dot 当前的成熟度和技术基础如何？

• Dot 目前仍处于非常早期、相对粗糙的阶段，存在功能局限和使用问题。
• 产品可以在移动端、电脑端及独立应用程序等多个界面中使用。
• 后续计划以周为单位快速改善，但彻底理顺跨系统、跨信息源和跨终端协作仍需要时间。
• 从研究角度看，Dot 是一个复杂系统，需要模型、工具调用、信息连接和任务执行等多项能力。
• 最关键的基础之一是 Astra，即首先要让底层模型的智能水平跨过关键门槛。

4. Space 与人机协作办公体系

4.1 为什么要一次推出多款办公套件类工具？

• 这一计划并非最初就被设定，而是在实践中发现，现有软件总体上并不是围绕人类与 AI 紧密协作设计的。
• 传统软件通常支持两种模式：

1. 用户亲自使用；
2. AI 代替用户操作。

• 但当多人及其各自的 AI 需要共同协作时，传统软件中的文档、权限、流程和协作抽象便难以适用。
• 因此，Space 的核心探索方向是建立适合“多人+多个 AI”协同工作的全新空间。

4.2 用户与 Dot 在 Space 中如何分工？

• 当出现新想法时，用户仍倾向于自己完成初始写作，因为写作本身有助于思考和梳理逻辑。
• 初稿形成后，再由自己的 Dot 或其他人的 Dot 完成整理、扩展、更新和后续执行。
• 这意味着 AI 并非完全替代人的原创过程，而是承担协同、补充、维护和持续推进的职责。

4.3 什么是“实时文档”？

• “实时文档”会根据人员互动或外部环境变化自动更新，而不是保持静态。
• 多个 Dot 可以共同处理同一份文档，既共享信息，也能够协作生成新内容。
• 例如，7个人各自头脑中的想法可以通过多个 Dot 的协作被整合，最终形成任何单一个体都未完整提出的涌现式成果。
• 这种能力尤其适用于：

1. 新功能探索；
2. 跨团队规划；
3. 动态项目管理；
4. 公司经营信息汇总；
5. 持续变化的决策文档。

4.4 动态仪表盘如何改变管理方式？

• 过去管理者往往与多人共同使用统一仪表盘；现在可以拥有仅为个人服务的动态仪表盘。
• 该仪表盘会根据用户不断变化的关注点，对业务情况进行个性化总结。
• 它可以读取 Slack 等信息源，生成新的分析内容并持续更新。
• 管理者的关注点会随时间变化，例如：

1. 6个月前可能主要关注收入、增长和业务指标；
2. 当前则更多关注安全、对齐和安保挑战。

• 未来希望能够对管理者数年间关注重点的变化进行可视化，并进一步感知用户情绪，判断某条信息适合立即查看还是延后处理。

5. 平台、应用市场与开发者生态

5.1 Dev Day 发布内容为什么持续增加？

• 本届 Dev Day 共发布约22项新内容，明显多于去年约9项和前年约4至5项。
• 原计划发布的内容更多，但为了把主题活动控制在50分钟内，部分内容被删减。
• 发布数量增长体现了 OpenAI 对自身理念的实践：

1. AI 工具可以显著提高生产力；
2. 生产力提升会推动人们创造更多产品；
3. 市场本身也需要更多新产品和新服务；
4. OpenAI 自身同样应利用 AI 构建更多产品。

• 如果当前模式持续有效，未来将由更多团队和个人共同开发，产品发布数量还会进一步增加。

5.2 OpenAI 希望建立什么样的平台？

• 平台目标是让人们可以像 OpenAI 一样，在 ChatGPT 中开发应用。
• 用户可以在任何希望使用的场景中携带自己的订阅，而不是为每项服务重复订阅。
• 平台将建设应用市场，帮助用户接触开发者未来创造的各种产品和服务。
• 核心理念是平台所创造的生态价值应大于平台自身获取的价值。

5.3 OpenAI 是否已经找到成熟的应用商店模式？

• 当前尚未找到与传统应用商店相对应的成熟模式。
• 但平台已经帮助数百万人开发和创造新产品、新服务与新公司。
• 现阶段的成功不应只用应用商店收入衡量，更重要的是平台是否降低了开发门槛、扩大了创新规模。
• OpenAI 对应用商店模式仍在持续探索，但开发平台本身已经创造了显著价值。

5.4 AI 模型公司会不会最终只剩一家？

• 行业内大致存在两种观点：

1. 未来全球最终只会剩下一家主导公司；
2. AI 将推动公司和创业项目迎来复兴。

• OpenAI 更认同第二种观点，认为创业门槛会大幅降低，一人公司和小型团队也能创造出重要产品与服务。
• 模型实验室不可能覆盖全部行业、场景和创意，OpenAI 内部能够想到的愿景只占整体机会的一小部分。
• 平台的战略任务不是吞噬所有应用，而是让更多开发者利用同样的工具创造新价值。

5.5 为什么允许用户携带自己的 ChatGPT 订阅使用第三方应用？

• 传统 AI 应用开发者需要控制用户使用量，否则模型成本可能难以承受。
• 为降低成本，开发者可能被迫使用更便宜但体验较差的模型。
• 用户携带自己的订阅后，可以在订阅范围内使用所需模型，降低开发者的成本压力。
• 对用户而言，也不需要为不同 AI 服务注册大量独立订阅。
• 核心原则是：订阅属于用户，用户应该能够在任何想使用的地方方便地使用。
• 这一机制有望扩大应用层创新，并帮助开发者更直接地连接真正需要其产品的用户。

5.6 OpenAI 最期待应用层出现什么产品？

• 最期待的是当前尚未被 OpenAI 想到、也无法明确预见的新产品。
• OpenAI 相信外部开发者的优秀想法远多于公司内部。
• 目前这些新平台能力刚刚出现，尚未产生明确的代表性应用。
• 但过去两年已经证明，AI 模型公司并未吞噬一切，反而不断催生重要的新公司和优秀产品。
• AI 创业公司取得成功的速度，已经达到前所未有的水平。

6. 模型、Decisions API 与技术路线布局

6.1 Decisions API 的推出背景是什么？

• Diogo 在 OpenAI 工作期间便长期探索相关思路，公司也持续研究如何将其产品化。
• Luna 用户的首要诉求依然是模型更便宜、速度更快、能力更强。
• 即使用户已经非常喜欢现有模型，也仍会持续要求降低成本和提升速度。
• 市场对 Jev 的积极反馈进一步提升了 OpenAI 对这一方向的优先级。
• Decisions API 的目标之一，是让 AI 在常见使用场景中实现极高速度和极低成本。

6.2 OpenAI 如何应对新模型架构和新实验室的竞争？

• Jev 可能采取了与 OpenAI 不同的技术方法。
• OpenAI 不确定是否会从头构建完全类似的技术，因为公司还需要推进其他核心事项。
• OpenAI 的目标是在成本、速度、模态组合等帕累托曲线的不同位置，提供尽可能优秀、便宜、快速和能力强的模型。
• 未来一定会持续出现新的架构、方法和技术，任何一家实验室都不可能覆盖全部路线。
• OpenAI 将重点做好作为主线和核心工具的模型，同时通过平台为不同技术路线和应用创新提供空间。

7. AI 创业、组织规模与产品开发方式变化

7.1 当前是否处于 AI 创业的黄金时代？

• 当前是一个令人兴奋的创业时期，初创公司可以以前所未有的速度获得成功。
• AI 大幅降低产品开发和实验成本，使小团队甚至个人能够推进过去需要大型组织才能完成的项目。
• 对投资而言，最重要的判断因素仍然是创始人：

1. 是否具备突出能力；
2. 是否在推进宏大的项目；
3. 是否愿意承担足够大的风险。

• 技术方向固然重要，但优秀创始人仍然是投资决策的核心。

7.2 “模型实验室会终结应用公司”的说法为什么没有成为现实？

• 这类说法容易传播，是因为悲观叙事更容易吸引注意力，但实际上并未出现模型公司吞噬所有创业公司的情况。
• OpenAI 的明确战略是赋能开发者，而不是成为包揽所有业务的庞大垄断公司。
• OpenAI 向开发者提供与内部使用相同的工具，并持续扩大模型访问能力。
• AI 带来的机会空间极大，单一公司无法覆盖所有领域。
• AI 更像一种改变可能性边界的基础工具，OpenAI 希望将其交给所有人使用，而不是限制在自身内部。

7.3 AI 时代哪些创业原则不变，哪些原则需要重新审视？

• 不变的原则是必须创造人们真正需要的产品。
• 即使产品未来服务于智能体，最终仍然需要有人对相关产品或结果产生需求。
• 需要重新审视的方面包括：

1. 是否必须拥有传统意义上的联合创始人；
2. 一人或两人团队能够完成多少工作；
3. 创业融资结构是否仍适合新的成本结构；
4. 团队规模、产品验证和增长路径是否需要重构。

• 硅谷现有的创业流水线尚未充分适应个人和小团队生产力大幅提升的新现实。

7.4 AI 如何改变大型组织负责人的产品开发方式？

• AI 显著降低了从想法到原型再到发布的阻力。
• 借助 Dots，管理者可以利用零散时间或集中时间直接推进产品。
• 过去使用 Codex 构建的项目可能主要供个人使用；现在则有机会跨过正式发布门槛，形成插件、扩展程序或真正的公司产品。
• AI 不仅提升普通开发者效率，也使公司负责人重新获得直接参与产品创造的能力。
• 即使大量尝试中最终只有一项真正出色，只要总体创造的社会效用提高，这种开发模式就有价值。

8. Ultrafast：速度、反馈循环与推理成本

8.1 是否所有提示词操作都使用 Ultrafast？

• 回答者目前所有提示词操作都在 Ultrafast 上完成。
• 由于已经适应极快的响应速度，其本人很难再切换回普通模式。
• 同时也意识到，并非所有用户都能在所有任务中承担 Ultrafast 的使用成本。
• 为避免与普通开发者的真实体验脱节，其本人也在主动为使用方式设置限制。

8.2 Ultrafast 为什么会显著改变工作方式？

• 创作者应与正在创造的事物保持即时联系，速度和反馈循环对创造过程非常重要。
• 快速响应不仅节省等待时间，更会改变人的迭代思考方式。
• 无论是构建产品还是分析问题，低延迟都可以让用户：

1. 更快尝试不同方案；
2. 保持思维连续性；
3. 获得更密集的反馈；
4. 降低中途放弃的概率；
5. 在单位时间内完成更多迭代。

• 速度因此将成为继模型智能之后的另一个核心竞争维度。

8.3 高速模型是否会影响 AI 的普惠性？

• 过去 OpenAI 主要关注降低“IQ 的价格”，即在提升模型能力的同时持续降低使用成本。
• 当前公司已经意识到，速度也需要像智能能力一样被普惠化。
• 自研芯片将发挥重要作用，帮助更多用户获得高速推理服务。
• 模型研究、推理技术栈和基础设施也将共同推动速度提升。
• 长期目标不是取消高价前沿产品，而是持续提高每1美元所能获得的智能和速度。

8.4 Ultrafast 未来的速度和价格趋势如何？

• 当前用户可以付费获得 Ultrafast，但成本可能超出多数人的承受能力。
• 预计不久后可以以极具竞争力的价格实现约8倍速度提升。
• 同一时期，前沿产品可能已经达到接近100倍的速度，或出现类似“GPT7 Galaxy”的更高阶产品。
• 在智能水平、速度和其他维度上，市场始终会存在价格更高的真正前沿产品。
• 但每1美元所能创造的价值将持续快速提升，并呈现复利式增长，其速度可能超过传统摩尔定律。

结构化总结表格

业务/技术板块

核心内容

当前状态

未来方向

关键价值

Dev Day 发布

本届发布约22项内容

明显高于去年约9项、前年约4—5项

利用 AI 进一步扩大产品发布规模

验证 AI 对内部研发效率的提升

工程与就业

AI 提升开发者能力，同时推高社会需求

开发者和待完成工作均持续增加

自动化低创造性任务，拓展新工作

人类继续处于 AI 进步中心

AI 社会愿景

从机器中心转向以人为中心

反对“人类失去作用”的未来

推动科学、艺术、创业与个人自主

形成“新文艺复兴”

Dot

24/7 后台运行的主动智能体

处于早期阶段，功能仍较粗糙

主动筛选信息、处理事务、修复问题

释放时间和注意力，支持持续迭代

Dot 技术基础

模型、工具调用、跨系统信息整合

Astra 是关键智能基础之一

提升主动执行和生产环境处理能力

从问答工具升级为长期智能助手

Space

面向多人和多个 AI 的协作空间

已探索实时文档和动态仪表盘

建立 AI 原生办公与协作体系

支持多智能体涌现式创造

实时文档

根据互动和外部情况自动更新

多个 Dot 可共同维护同一文档

扩展到规划、决策和经营管理

将静态文档转化为动态工作系统

平台生态

ChatGPT 原生应用、插件、应用市场

尚未找到成熟应用商店模式

扩大开发者生态和用户触达

让平台创造的价值大于自身获取价值

订阅开放

用户可携带 ChatGPT 订阅使用第三方应用

新机制刚刚推出

降低开发者模型成本压力

减少重复订阅，促进应用创新

Decisions API

面向更快、更便宜的常见场景

因 Jev 市场反馈而提高优先级

优化速度、成本与能力组合

扩大高频 AI 应用的经济可行性

模型战略

覆盖成本、速度、模态的帕累托曲线

无法覆盖所有架构与技术路线

聚焦主线模型和核心工具

保持基础模型竞争力并容纳外部创新

AI 创业

一人和小团队能力显著增强

初创公司成功速度加快

重构团队、融资和产品开发方式

降低创业门槛，推动公司复兴

Ultrafast

极低延迟推理与快速反馈

当前成本较高，回答者已全面使用

推动速度普惠化及自研芯片落地

改变思考与迭代方式

速度规划

近期约8倍高速能力，前沿产品或接近100倍

高速能力仍存在价格门槛

持续提高每1美元的速度与智能价值

形成超过传统摩尔定律的复利提升

新版prompt提炼之后的转译

提问者：你所有的提示词操作都在 Ultrafast 上完成吗？

回　答　者：是的。7个人的想法协同在一起，可以创造出新的事物。它在某种程度上存在于我们每个人的头脑中，但最终呈现为一种涌现出来的成果。

提问者：我们现在正处于 Dev Day。如果4年前有人告诉我，编程模型会达到如今的能力水平，我可能会认为 Dev Day 不会再有这么多开发者。但现在的开发者似乎比以往更多，需要完成的工作也比以往更多。随着编程模型的能力不断提升，并持续观察其对工程领域的影响，你从中对工程的未来，以及更广泛的人工智能时代工作前景，有了哪些认识？

回　答　者：人们想要的东西越来越多，期望也非常高。我认为这是一件很好的事，意味着我们面前有一个值得探索的美好未来。开发者从未像现在这样有能力完成更多工作，但行业对他们的要求也从未如此之高。我认为这很好，也是社会发展中令人欣慰的一面。

提问者：这是否出乎你的意料？你此前预期过这种情况吗？

回　答　者：对于人工智能对就业的影响，我从来没有像其他人那样悲观。不过，世界上存在如此庞大的潜在需求，以及人类始终处于人工智能相关进展的中心，仍然在积极意义上有些超出我的预期。

提问者：你认为这种趋势存在极限吗？它能否顺利延伸至其他工作领域？

回　答　者：有些工作领域确实不需要太多创造力，我们也并不关心具体由谁来完成，因此其中许多任务或岗位将由人工智能承担。但世界上还有大量其他事物是人们需要的，因此不会缺少值得投入的新工作。我认为，这是近期发展带来的重要启示之一。

提问者：尽管我们经常讨论自动化，为什么人工智能的发展仍然如此以人为中心？

回　答　者：一个观察是，我们其实并没有那么在意机器做了什么。从某种意义上说，我们一直都是如此。我们真正关注的是其他人，想了解艺术作品背后的创作者、公司背后的人，以及某个理念或产品背后的人。我认为，这是一个令人欣慰的积极发现，只是我们或许早就应该意识到这一点。

提问者：我也认同。我一直在想，如果一个智能体在数据中心解决了纳维-斯托克斯问题，却没有人知道，它是否真的解决了这个问题？

回　答　者：这是一个很好的例子。10,000个智能体努力工作了4天，彼此之间实现了完美协作，最终完成了一件了不起的事情。但公众的反应几乎都没有集中在问题真正得到解决这件事上，更多讨论聚焦于如何解读这一成果，以及以不同方式参与其中的人。不过，这仍然是一件好事。

提问者：今天是 Dev Day。据我统计，你们今天发布了22项新内容。我参加了全部4届 Dev Day，去年发布的内容大约是9项，前年大约是4或5项，因此今年明显比往年更多。

回　答　者：等到明年再看。我们原本还有一批计划在今天发布的内容，但为了把活动控制在50分钟内，只能将其删减。我们一直努力践行自己的理念。如果我们认为这些工具能够显著提升人们的生产力，让人们创造更多，而人们也确实需要更多东西，那么我们自己就应该构建更多产品。

提问者：在这一过程中，新的瓶颈和最佳实践是什么？以我自己的感受来看，现在几乎什么都能构建，随之而来的问题便是：我们究竟应该构建什么，又该如何让所有内容保持协调一致？你如何看待这个问题？

回　答　者：推进……

回　答　者：我们正在进入一个所谓“持久智能”或“主动智能”的世界。无论人们如何称呼，其核心理念都是让智能体在后台全天候 24/7 运行，与用户的系统连接，主动为用户处理事务。这显然是一个值得探索的重要方向，我认为未来会出现令人惊叹的成果。

我们真正希望推动的另一件事是，如今已有大量想法具备落地条件，但我们仍然缺少发现并实现这些想法的创造精神。因此，我们希望真正开放生态系统，让人们能够像我们一样在 ChatGPT 中开发应用；让用户可以在任何希望使用的地方使用自己的订阅；建立应用市场，真正帮助用户接触到人们未来能够开发的各种产品与服务。

提问者：从您发布的所有平台功能以及您谈论平台的方式来看，这让我想起比尔·盖茨过去谈论 Windows 平台时的理念，即创造的价值多于自身获取的价值。您在主题演讲中也提到过这一点，我也观察到，每次开发者大会似乎都会涉及应用商店。你们已经多次尝试建设应用商店。从构建平台的过程中，您获得了哪些经验？您认为这次的平台有何不同？

回　答　者：我们显然还没有找到与应用商店相对应的成熟模式，但已经成功构建了一个平台，许多人正在此基础上创造巨大的价值。

在构建 AI 模型公司的群体中，大致存在两种观点。一种观点认为或期待，未来全世界只会剩下 1 家公司；另一种观点与我们相同，认为公司数量将迎来一场复兴，创业门槛也将大幅降低，使 1 人公司和小型团队公司都能创造出令人惊叹的新产品与服务。我们希望尽最大努力，通过平台帮助人们实现这一目标。

虽然我们尚未找到应用商店的成熟模式，这一评价完全公允，但我们已经找到了帮助数100万人开发和创造重要新产品与新公司的方法。

提问者：您在主题演讲中谈到了工业革命与文艺复兴之间的区别，刚才也再次提到这一点。如果必须聚焦于决定这两种未来走向的关键因素，无论是对 OpenAI 及其合作伙伴，还是对更广泛的生态系统而言，您认为这些因素是什么？

回　答　者：一种理解是，文艺复兴从根本上以人为中心，而工业革命从根本上以机器为中心。工业革命带来了许多积极成果，并在多个方面对世界产生了极其正面的影响。但我认为，对许多人而言，工业革命并未给个人生活带来完全积极的影响。

人们原本拥有丰富的生活，可以从事各种活动。当然，他们也面临困境，希望获得更多、更好的事物，而工业革命在某些方面确实满足了这些需求。但在许多情况下，人们也被变成了机器中的齿轮，只负责某项高度专业化的工作，并与许多让生活变得美好的部分失去联系。

当我听到有人说“AI 将实现一切自动化，人类将不再发挥任何作用；我们会像家猫一样，由 AI 决定如何奖励我们并告诉我们该做什么”时，我感到难过。

提问者：家猫，我也听人这样说过。

回　答　者：这显然不是我们的看法。我希望 AI 能够带来新一轮大规模的科学发现、艺术创作和公司创立浪潮，让每一个希望开创新事物的人都能参与其中。我也希望人们既能享受到经济增长和新事物带来的利益，又能重新掌控自己的生活，有能力做自己想做的事情，并按照自己希望的方式安排时间。

提问者：文艺复兴与工业革命之间令我印象深刻的一点是……

提问者：文艺复兴时期，人们开始以不同的方式理解“人”和“个体”的含义，不再仅仅将自己视为机器中的齿轮。您是否认为，我们如今对人类自身的认知也开始发生变化，并且可以与文艺复兴时期的变化相提并论？

回　答　者：文艺复兴时期发生了太多事情，很难将其归结为某1件事。但其中很重要的一部分，确实是人们开始重新连接久远过去的许多事物。与此同时，疯狂的未来主义、科学与工程领域的新发现也结合在一起，使人们感到人类的潜力得到了极大拓展。

现在，人们在谈论AI时，也会同时提到这2个方面，这正是我最初对这一类比产生兴趣的原因。与未来可能实现或已经能够实现的事情相比，下面这个例子或许非常微不足道，但我们今天发布的新功能Dot，是第1个让我重新掌控早晨时间和精神空间的产品。

我过去非常喜欢清晨，那是我1天中最喜欢的工作时段，也是创造力最强的时候。我通常会在这段时间处理当天真正深入、复杂的项目。但现在，我经营着一家大公司，总会有事情出现问题，夜间也会发生很多情况。如果早上不立即查看手机，有时不会有问题，但有时确实会发生严重事故。如果我忽视1小时，情况就会变得更加糟糕。

问题在于，一旦查看手机，各种事情就会占据我的大脑，使我失去深度工作的时间。现在，我相信Dot能够准确判断哪些事情需要我关注、哪些不需要。如果有非常紧急的情况，它会通知我；否则，它会替我处理。这样一来，我重新获得了用于创造性工作或陪伴孩子的时间和空间。这听起来似乎只是很小的变化，但确实显著改善了我的生活。

提问者：也就是说，它似乎替您分流了很多原本必须立即处理的事情。您可以放心地把其他事情留到以后，因为早上查看时，您有信心获得真正需要关注的信息。是这样吗？

回　答　者：是的。今天发生了2件与此相关的事情。主题演讲期间，我们遇到了严重的SEV 1生产事故。这正是进行现场演示并尝试同时部署多项新功能的风险。我当时正在使用curl。

Tibo在演示开始前5分钟收到了他的Dot发来的消息：“提醒一下，我看到你正在参加Dev Day主题演讲，也知道你一直在为此做准备。我还通过Space之类的地方查看了演示文稿，因此知道演示内容。这次SEV事故可能会导致演示无法正常进行。现在距离演示开始只剩几分钟，你希望我尝试修复吗？”

目前它还无法真正完成修复，但也许下1个版本就能直接在生产环境中解决问题。可以想象，要实现这样的效果，背后需要整合多少信息和能力。

就在来这里之前，我已经连续忙了1整天，正常工作进度严重滞后。就在我进入上一项安排之前，Dot对我说：“从现在到你要与Dan参加的活动之间，只有1小段空档，你必须关注这1件事。”如果没有它，我根本不会发现那件事，最终可能造成很大的混乱。仅仅是不再需要为此担心，就已经带来了截然不同的感受。

提问者：您的Dot叫什么名字？

回　答　者：Dotty。之前我曾把它叫作Alfred。目前很多人也在使用Dotty这个名字，但我还在考虑1个更好的名字。

提问者：我的Dot叫Boo。

回　答　者：这个名字很好。到目前为止，您的使用感受如何？我知道产品还处于非常早期的阶段，也比较粗糙。

提问者：我很喜欢。它已经成为我使用聊天功能的默认方式，我认为这是1个非常积极的信号。我之前也和您分享过相关反馈。因为我参与了测试，所以从早期阶段就开始使用它。我认为它目前仍有局限，而且在我的使用过程中也出现过一些问题。

回　答　者：我也有同感。它可以在许多不同的界面上使用，包括移动端，同时也有独立的应用程序。

回　答　者：电脑以及诸如此类的内容。我认为，妥善解决这类问题非常重要。从现在开始，它每周都会迅速改善，但要彻底理顺，肯定还需要一些时间。

我遇到过一个类似的例子。当时我想改签航班，它提示：“您改签后的航班时段内有一个会议。”但这个会议并不在我的日历中，它只是从 Slack 中读取到了相关信息。它能够整合你生活中各方面的背景信息，并以一种即使非常细心的人也很难准确做到的方式处理事务，这一点令人印象深刻，也是我认为它最有用的能力之一。

提问者：从研究角度来看，要实现这一点，需要具备哪些条件？

回　答　者：这是一个非常复杂的系统，涉及很多内容，但其中最重要的可能是 Astra。首先需要一个足够智能的模型，此外还有许多其他方面需要构建，但让模型的智能水平跨过关键门槛是核心环节。

提问者：您曾告诉我，不只是在日程安排或信息筛选等紧急事项上，它还有几次帮助您发现了一些原本无法发现的信息，或借助 AI 作出了原本无法作出的决策。您说，即使雇人也无法完成，因为这项工作过于复杂。您能否分享一些具体例子？

回　答　者：有一个我很期待的新功能，但要向别人详细解释清楚并让他们构建出来，会非常困难。即使能够说明，实现它的方式也有很多种，而我无法确定哪一种最好。

因此，每当我想到什么，或者散步时产生了想法，我都会随时给我的 Dot 记下笔记。随后，Dot 构建了5或6个版本，我可以提供反馈并持续迭代。只要有空闲时间或产生新想法，我就可以直接以原始形式输入，并获得和提供近乎持续的反馈，这种体验非常好。

类似的情况还有很多。有些想法的不确定性很高，或者由自己或他人构建需要投入太多精力，但现在可以持续迭代。晚上入睡前，可以留下一段很长的语音笔记，记录各种想法；早上醒来时，不仅相关内容已经构建完成，它甚至还继续完成了接下来的7个步骤。这是表达想法的一种非常好的方式。

此外，还有一些任务，我通常会放弃，或者过去已经放弃。有一次，我确信自己在 Slack 中看到过某项内容。Slack 的搜索功能很差，因此我通常使用 Codex 进行搜索。当时我想再次找到并引用那项内容，便花了10或20分钟尝试通过 Codex 搜索。我不断调整搜索方式，但始终没有找到。由于 Codex 通常不会失败，我甚至怀疑自己是否只是在梦里见过，实际上并不存在，最后便放弃了。

第2天，我的 Dot 提示：“我注意到您昨天尝试查找这项内容，但没有找到。我整晚都在继续思考，现在已经替您找到了。”我非常惊讶。原来那是一张截图，而不是 Slack 中的文本，我此前忘记了这一点。它一直耐心地寻找能够帮助我的事情，对每张截图进行图像识别，最终找到了相关内容。这非常出色，因为如果没有它，我已经放弃了。

我们经常感到 Slack、电子邮件以及其他渠道中有太多内容需要阅读，而它非常适合处理如此庞大的背景信息，并在恰当的时间提供恰当的内容。人们一直在讨论，需要有一个智能体来应对互联网带来的信息洪流，以及各种试图占据注意力、分散注意力的内容。在使用 Dot 的几周里，我明确感受到，尽管技术在很多方面都非常出色。

提问者：请介绍一下 Space。一次发布 3 款办公套件类工具，是一项很大的投入。

回　答　者：我们最初并没有打算这样做，但后来意识到，软件总体上并不是为与 AI 协作而设计的。我相信您也发现了类似的问题。用户可以自己使用软件，也可以让 AI 代为使用；但如果要实现紧密协作，特别是多人及其各自的 AI 共同协作，现有的许多抽象方式就难以适用。因此，这确实是一项雄心勃勃的投入，但我认为，人类与 AI 的协作空间非常值得探索。

提问者：您如何使用 Space？您是在 Space 中审阅文档，还是创建文档？对于其中由您输入的内容与由您的 Dot 代为完成的内容，您是否有明确预期？收到文档后，您也会从这一角度进行审阅吗？

回　答　者：如果有新想法，我喜欢用传统方式写作，因为这有助于我思考。因此，我仍会自己撰写，再让我的 Dot 或其他人的 Dot 完成后续工作。

我认为最令人惊叹的一点是，目前可能还没有合适的词来描述，暂且称之为“实时文档”。这类文档会随着人们的互动或外部情况的变化自动更新。多个 Dot 还可以共同处理同一份文档，不仅能够共享信息，还能由 7 个人的 Dot 协作创造出新的内容。这些内容在某种程度上原本存在于每个人的头脑中，但最终会形成一种涌现的成果。在探索新功能时，这一点非常有趣。

提问者：是否有某一份实时文档对您最有帮助？

回　答　者：我使用的是仪表盘。我想，每个人可能都有一个每天查看的仪表盘，用来按照自己关注的方式了解公司的经营情况。过去，我会与所有人或许多其他人共用一个仪表盘；现在，我拥有一个只为自己服务的仪表盘，它会根据我不断变化的关注点，对业务情况进行动态总结。它会查看我所有的 Slack 消息，为我生成新的内容并持续更新，让人感觉这确实是一份动态变化的文档。

提问者：您目前最关注什么？

回　答　者：最近，我经常思考如何应对摆在我们面前的安全、对齐和安保挑战。但在 6 个月前，我可能更关注收入、增长和指标等传统业务问题。

观察这些关注点的变化很有意思。我一直在思考如何呈现这种变化，因为我非常想知道，随着时间推移，我的 Dot 认为我每天最关注的事情是什么。目前，我们还没有很好的方式将其可视化，我也没能让我的 Dot 找到一种很酷的呈现方式。我希望能够看到这些关注点在数年间的变化。

提问者：我认为这非常有意思。我一直想要一个能够实时更新的仪表盘。

回　答　者：我希望它能感知我的情绪。例如，当我收到 Slack 消息但不确定是否要查看时，我会让 Dot 先查看，再判断我应该现在看还是稍后看。它可能会建议我稍后再看。我希望有一个能够实时了解我情绪的工具。

提问者：我也希望有这样的工具。请介绍一下 Decisions API。我认为，Decisions API 可能是今天最重要但又容易被忽视的内容之一，因为 Jev 是近期备受关注的新事物，也确实改变了我们内部构建产品的思路。Decisions API 是从 Jev 开始构思的，还是已经开发了一段时间？您如何看待它？

回　答　者：Diogo 曾在 OpenAI 工作，他很早就在探讨这一类想法，我们也一直在研究如何将其实现。用户对 Luna 的第1大诉求体现了人们始终希望产品更便宜、更快速、更出色。尽管用户非常喜欢这个模型，但对 Luna 的第1大诉求仍然是希望它更便宜、速度更快。因此，我们一直在思考如何进一步推动它朝这个方向发展。

看到市场对 Jev 的反响后，我们确实更愿意优先推进这个方向。Diogo 长期以来也一直启发我思考这类问题。未来，我们会尝试以不同方式推进相关工作，让技术在常见使用场景中实现极高速度和极低成本。

提问者：如果 Jev 这类新型模型出现，而 OpenAI 又可以增加 Decisions API，您认为这将如何改变新实验室与现有实验室之间的竞争？OpenAI 可以覆盖哪些领域，新实验室又可以专注于哪些方向？

回　答　者：据我了解，Jev 采用的方法有很大不同。我不确定我们能否从头构建类似的东西，也许可以，但我们还有其他事情需要推进。我们希望在成本、速度和模态构成的帕累托曲线上的每一个位置，都提供最好、最便宜、最快、能力最强的模型。

未来还会出现新的方法、新的架构和新的技术，我们不可能覆盖所有方向。但对于作为主线和核心工具的模型，我们会努力做好交付。

提问者：我一直在向您建议推出 Codex 原生应用，现在终于实现了，其中还包括插件扩展。用户还可以将 ChatGPT 订阅中的 tokens 带给应用开发者。对于我这样的开发者而言，这改变了原有的成本考量。过去通常需要考虑可以允许用户使用多少 AI，才不会让自身难以承担成本；也需要考虑是否应使用价格更低的模型，但这又会导致体验变差。如果用户可以带着自己的订阅使用应用，那么他们就可以在订阅范围内使用自己想用的任何模型。在这个新环境下，您最期待应用层能够构建出哪些应用和产品？

回　答　者：我最期待的是那些我们尚未想到的产品。我们这样做有很多原因，其中一个最重要的原因，也是我们举办大型 Dev Day 活动的原因，就是我们始终坚信，OpenAI 之外的好想法远多于内部。

我目前还不知道自己最喜欢的 ChatGPT 原生应用或插件扩展会是什么，但我非常确信，未来一定会出现一些我真正喜欢、却无法在今天想象出来的产品，而且目前 OpenAI 内部也没有人能够明确预见这些产品。这正是推进这项工作如此重要的原因之一。

我们希望让人们能够发挥这种创造力，也希望帮助开发者与真正认为这些技术有用的用户建立联系。至于让用户随时随地携带自己的 ChatGPT 订阅，这是我们自己也希望拥有的能力。人们正在构建许多出色的产品，也出现了很多我们想尝试的新应用。

回　答　者：我不喜欢为了使用不同服务而注册100个订阅。我也希望在尝试某项产品时，能够充分使用AI能力。我们希望让用户可以非常方便地随处使用自己的订阅。订阅属于用户，用户应该能够在任何想用的地方使用。

提问者：目前应用层是否有人在这方面做得很好，开发出了有趣的应用？

回　答　者：由于这些新技术刚刚出现，目前还没有。但整体来看，尽管2年前许多人担心AI模型公司会吞噬一切，市场上仍不断涌现大量重要的新公司和优秀的新产品。我从未见过初创公司能够如此迅速地取得成功。

提问者：您在加入OpenAI之前曾参与Y Combinator。这是否改变了您对创业的看法？当人们询问如何创办初创公司、如何实现增长等问题时，您的思考是否发生了变化？

回　答　者：我很希望再次成为初创公司投资人，因为现在确实是一个非常令人兴奋的时期。看到人们目前正在做的事情，并与其中一些创始人交流，是非常难得的体验，尽管我能够投入的时间远少于自己所希望的。这确实是一个黄金时代。

提问者：除了创始人本身，您还会关注哪些因素？什么样的项目会让您产生投资意愿？

回　答　者：我一直是一名非常注重创始人的投资人。和其他人一样，我也有自己特别感兴趣的项目，但我的投资方式始终是寻找能力最强、正在推进宏大项目的创始人，并愿意与他们共同承担大量风险。

提问者：市场上似乎始终存在一种说法，认为模型实验室将吞噬一切。在DevDay期间，最容易引发传播的推文往往是“OpenAI刚刚终结了某家初创公司”之类的内容。

回　答　者：这类推文之所以容易传播，是因为人们喜欢悲观论调，但这种事情实际上从未发生。

提问者：为什么这种事情从未发生？

回　答　者：首先，我们坚定地认为，赋能开发者，而不是试图成为1家庞大的邪恶巨头公司，是非常重要的。因此，我们今天采取了许多其他AI实验室不会采取的举措，明确表示将向开发者提供与我们自己使用的完全相同的工具。我们始终在推动模型访问能力的边界，因此，我认为我们在行业中对这件事有着独特而坚定的承诺。我们确实非常重视开发者，这在一定程度上也源于我对初创公司的热爱，以及我自身的从业背景。

此外，我们正在经历一场新的文艺复兴，还有太多领域有待探索。人们将创造出令人难以置信的成果，而我们自身能够想到的愿景只会占其中极小的一部分。这个世界从根本上获得了一种全新的赋能工具。我甚至不知道该如何类比，也许它就像火一样，真正改变了可能性的边界。或许是我们将它从山上带了下来，但我们希望把它交给所有人，让大家用它创造新的事物。

提问者：创业领域存在许多基本经验，例如始终应该有1位联合创始人，甚至Y Combinator的投资结构也是基于过去的成本结构形成的。如果从头开始设计或重新思考这些做法，您认为会发生哪些变化？

回　答　者：有些原则可能始终不变。例如，必须创造人们真正需要的产品，这条建议在1,000,000年前也是正确的。

提问者：或者创造智能体需要的产品？

回　答　者：是的。不过，至少按照我们今天对世界的理解，归根结底仍然需要有人想要某种东西。有些原则并未改变，但还有很多方面已经完全不同，例如1个人、2个人或其他规模的团队能够完成多少工作。我认为，硅谷现有的创业流水线还没有对此进行充分的重新审视。

提问者：作为一家大型公司的负责人，您通常可能没有时间亲自开发产品，但现在似乎已经能够这样做。能够在日常工作之外开发产品，对您的领导方式和所能完成的事情产生了怎样的影响？

回　答　者：首先，我们并不是一家规模庞大的公司。大型科技公司的员工数量可能是我们的100倍。

回　答　者：我始终相信，只要稍微降低一些阻力，人们就能创造出更多东西。现在有了 Dots，无论是利用零散时间，还是投入较长时间，我都可以通过这种交互方式完成大量工作。我感觉自己会做一些不再只是像过去使用 Codex 时那样、仅供个人使用的小项目。我希望能向全世界发布一些东西，真正创造出一项很酷的新事物，并推动它跨过发布门槛。

提问者：您是打算把它发布在 samman.com 上，还是将其做成一款真正的 OpenAI 产品？

回　答　者：我的意思是，我会尝试开发一个插件、扩展程序或类似的东西，然后发布出去。也许我还会推出一款真正的 OpenAI 产品。现在似乎一切皆有可能。

提问者：但您仍然需要作出决定。您原本可以直接安排一个团队来做这件事，而现在可以自己开发一个插件，这会带来什么额外价值？

回　答　者：今天就是一个例子。您提到，每届 Dev Day 发布的内容都在增加。如果没有 AI 的帮助，我们不可能完成所有这些工作。如果这种方式行之有效，明年我们会安排团队去做，我也会亲自做，更多人同样会参与，我们将推出数量惊人的优秀成果。即使最终只有一项真正出色的成果也很好，但我希望我们为世界创造的总体效用能够提升。

提问者：您的所有提示词都使用 Ultrafast 处理吗？

回　答　者：是的。

提问者：这非常惊人。

回　答　者：我一直在思考这件事。我担心从某些方面来说，我们可能会逐渐脱离那些使用我们工具进行开发的人，因此一直在努力为自己设置各种限制。我非常清楚，并非所有人都能始终在所有任务中使用 Ultrafast，但我自己已经很难再切换回去了。

提问者：您仿佛已经生活在未来，而我们所有人或许也能略微体验到这种未来。在 Ultrafast 的世界里，使用这些工具的方式有何不同？

回　答　者：大约在 15、20 年前，Brett Victor 做过一场精彩的演讲，讨论了他所认同的工作原则。我的转述可能不完全准确，其中一个观点大致是，创作者应当与自己正在创造的事物保持即时联系，而速度和反馈循环至关重要。

我此前没有意识到，快速获得响应会对我的迭代思考过程产生如此大的影响。无论是尝试构建某种事物，还是思考一个问题，这种速度都会带来极为显著的差异。

提问者：长期以来，您一直将 AI 视为一种普惠化力量，让任何人都能创造事物。在 Ultrafast 的世界，或 Asterism 这类更昂贵模型的世界中，人们可以付出更高成本来获得更强的智能。您如何看待这一点？

回　答　者：从以往来看，我们主要关注的是降低“IQ 的价格”，而且我认为我们在这方面做得很好。我们打造了能力大幅提升的模型，同时模型成本每隔几个月或每年都会显著下降。

现在，速度显然会成为一个非常重要的新维度，我们正在为此进行投入。我们自己的芯片将发挥重要作用，让每个人都能使用速度极快的推理服务。此外，模型还会推动研究、推理技术栈及其他各方面的改进。我们现在已经意识到，必须像推动模型智能普惠化一样，投资推动速度的普惠化。

提问者：那么，我什么时候可以在所有提示词中使用 Ultrafast？

回　答　者：你现在就可以为此向我们付费，让我们来完成，但你可能会觉得自己负担不起。我认为，在不久的将来，我们能够以极具竞争力的价格实现8倍速度提升；但届时，我们也将拥有快100倍的产品，还会出现类似GPT7 Galaxy之类的产品。

在我看来，理解这一趋势的正确方式是：在智能水平、速度或其他方面，始终都会存在价格更高的真正前沿产品。但可以确定的是，我们会延续此前的做法，持续提升每1美元所能创造的价值。这种提升将以复利式的方式不断累积，其力度甚至足以让摩尔定律相形见绌。

原文转译

Do you do all your prompting on Ultrafast? I do. That's sick. Yeah. That's crazy. You know, seven people's thoughts can working together create something new that kind of existed in all of our heads to some degree, but was like an emergent thing. Hey, this is Dan Shipper. You might be wondering, what happened to AI and I? This is the same show, we are just rebranding it to the Every Podcast. I'm really excited for our first episode with Sam Altman. Here it is. Sam, welcome to the show. Thanks for having me. Psyched to have you here. So, we're here at Dev Day. I think if you had told me four years ago that the coding models were at this level of ability... They're so good now. I would have been like, wow, I don't know if there would be a lot of developers at Dev Day. But there's more developers than ever, it seems like, and more work to do than ever. So, what has the like riding the increasing capabilities of coding models and seeing the impact on engineering taught you about the future of engineering and uh more broadly like the future of work with AI? I think people just want so much more stuff. People's expectations are so high. Uh, and I think that's a wonderful thing. I think this is why we'll all have this incredible future in front of us to explore. But developers, they've never been able to do more, but there's never been a higher bar either, and I think that's great. I think it's like a nice thing about society. Is that something that surprised you? Did you expect that to happen? I think I was never like as doomy about the job stuff as other people, but it still has surprised me somewhat on the upside of just how much latent demand existed in the world and how and also how much like people have remained at the center of everything that is happening with AI. Do you think there's a a limit to that or do you think that translates pretty cleanly to other areas of work? I think there will be some areas of work where there's actually not much creativity required and we don't care about the person doing it. And many of those tasks or jobs will be done by AI. But I think there's so much other stuff people want in the world, there's going to be no shortage of new things to go to go do. I think that's been one of the lessons of recent time. Why do you think it's so human-centered even though we talk a lot about automation? I mean, one one observation is we just we don't care that much about what the machines do. And in some sense we never have. And we're just like we're obsessed with other people. We want to know about the person behind the art or the person behind the company or the idea or the product. And I think this is a like a wonderful positive surprise. that we probably should have all known before. I agree. It's I've been thinking like if an agent solves a Navier Stokes in a data center but no one finds out, did it really solve it, you know? I think that was like a great example where, you know, 10,000 agents worked super hard for four days and perfect coordination with each other and did this amazing thing. And almost none of the public reaction was about that actually, the problem being solved. A lot of talk about the interpretation, a lot of talk about like the people involved in it in different ways. But, you know, that was like a, that was a good thing. Yeah. So, it's Dev Day. You launched by my count 22 new things today, uh, which I've been at all four Dev days and uh I think last year was like nine and the year before that was like four or five. So it's a lot more than usual. Is that Just wait for next year. Okay. I I'm tired. I'm tired. Um, we cut a bunch of things that we were going to launch today just because you have to like fit it into 50 minutes. Uh but, you know, like we try to practice what we preach and if we are saying these tools can make people much more productive and build more and that people want more stuff, like we better build more ourselves. What are the new kind of bottlenecks and best practices for doing that? Because I I can see for myself and and I assume you guys are seeing this too, like now that you can build everything, you're like, what what should we build and and how do we make it all coherent? How do you think about that? I mean, moving, moving... into this world of sort of persistent intelligence, active intelligence, whatever people want to call it. This idea that you're going to have these agents that are running in the background 24/7 connected to your systems doing stuff proactively for you. This is clearly a direction that feels important to explore. I think we're going to see amazing stuff happen there. Another thing we really wanted to do is there's just there are now so many ideas that are capable of being built and what we are missing is the creative spirit to figure them out and build them. So we wanted to really open up our ecosystem, let people build apps in ChaGBT and like the same way that we get to do them. Let people bring their subscription wherever they'd like, have the marketplace and and really try to let people get connected to all the stuff that people will now be able to build. Yeah, I mean, I I just looking at all the platform features that you launched and the way that you talk about platforms, it reminds me a lot of the way that Bill Gates used to talk about Windows as a platform where you create more value than you capture. What have you learned, you know, you said this in your keynote and I've observed it as well. There's at Dev there's always an app store. You you've been taking swings at app stores for a while. What have you learned about building a platform and what do you think makes this one different? I I mean, we we clearly have not yet figured out the App Store equivalent, but we have figured out how to build a platform that a lot of people are building incredible value on top of. I I think there are really kind of two schools of thought among people building AI model companies. There are people that believe or hope that there's only going to be one company in the world. And then there are people like us that believe we are going to see a renaissance in the number of companies and just an absolute lowering of the bar to people, one-person companies, small team companies creating incredible new products and services. And we want to go as hard as we can with the platform to enable people to do that. And although I think we haven't figured out the App Store yet, uh that's totally fair. I think we have figured out how to help like millions of people build and create significant new products and companies. If you had to narrow in, you you said you you talked at the keynote about an industrial revolution versus Renaissance which you just which you just mentioned now. If you had to narrow in on the key things that make the difference between those two worlds and maybe open AI is partner, but maybe even the broader ecosystem, what do you think those things are? One take on this would be that the Renaissance was fundamentally centered around people and the Industrial Revolution was fundamentally centered around machines. There were a lot of good things about the Industrial Revolution. It had a, you know, hugely positive impact on the world in in many ways. But I think a lot of people did not have a purely positive personal impact from that. And this feeling that um people had these like rich lives and did all these things and yeah, of course they like struggled and wanted more and better stuff and that's the part the Industrial Revolution did deliver in some ways, but they were kind of turned into in many cases, this like cog in a machine doing some extremely specialized thing and kind of like disconnected from a lot of the parts that made life good. I get sad to hear people talk about like, oh, AI is going to automate everything. There's going to be no role for humans. We're going to be like house cats and the AI will decide how to reward us and tell us what to do and Um house cats. I've heard people say that. Um, that's certainly not what we think. Yeah. Um, what I am hopeful for is that this is just a huge new wave of scientific discovery, art, company formation, like everybody that wants to start a new thing. Um, and that people do get the benefits of all the economic growth and the new stuff happening here. But also people get to back in control their lives and ability to kind of do the things they want and spend their time the way they want. One of the things that strikes me about the Renaissance versus the Industrial Revolution is In the Renaissance, you have a sort of different understanding of what it means to be a person and like an individual, uh, as a Yeah. as opposed to sort of the cog in the machine. Do you feel like you're starting to see any difference in how we think of ourselves as human that's comparable to the way that we changed how we think of ourselves during the Renaissance? I mean, so much happened during the Renaissance. It's hard to say there's like this one thing. But but but definitely a big part of it was about people like sort of reconnecting to a lot of things that had in the in the distant past. And you do hear and then and then also like this kind of combination of that with this crazy futurism and discovery of science and engineering and just sort of this this feeling that human potential had massively expanded. And you do hear people talk about both of those things simultaneously with AI, which is kind of how I got interested in this as an analogy in the first place. I mean, this is like a very trivial example relative to what is going to be possible or already is possible. But the dots, the new thing we launched today is the first thing that has given me back control of my mornings and like my mental space in the mornings. I used to love the early mornings. It's my favorite time of the day to work. It's like maximum creativity. That's when I would like do the really deep complicated project of the day. And then now I like run this like big company and something is going wrong all of the time. And overnight, a lot of stuff happens. And if I don't check my phone right away, then, you know, sometimes it was fine and sometimes there was like a real disaster that if I ignored for an hour, it just got much worse. The problem had been that as I checked my phone, I polluted my brain with all these other things and I lost the deep work period. But now, I trust my dot to go have good judgment about what I need to think about and not think about if there's something very urgent, otherwise handle things for me. And I get this space for kind of like creativity or to just hang out with my kids, uh back. And that has been like, you know, it sounds small but like a significant improvement to my life. That's what I was going to ask is so it look it seems like it's deflecting a lot of things that you might have had to deal with immediately that you can just like leave for later because you're confident you get the thing you need when when you check it in the morning. Yeah. I mean two Two things that happened to today about that. Um during the keynote, we had like a big sev one production. This is the risk of live demos and trying to deploy many new features at once. I was curling. I was... Um so, so that happened. Uh Tibo got a message from his dot five minutes before the demo saying, just to let you know, um there's I see that you're in the Dev Day keynote. I know you were working on it. I know what the demo is because I was able to like look at the presentation, um in space or whatever. And they're, you know, this sev is probably going to make the demo not work. Would you like me to try to fix it? There's like a few minutes to go. And now it couldn't fix it yet. But like maybe the next version will be able to go fix that in prod. And like think about that whole like all the stuff that had to come together to make that work. Um just right now before I came here, I've been in like back-to-back stuff all day. Very, very behind on all of my normal stuff. Um my dot just before I had walked into the last thing said, you have one little break between now and this thing you're going to go to with Dan. You have to pay attention to this one thing. I never would have found that otherwise and like it would have been a big mess. And just the fact that I like don't have to worry about that feels feels super different. What's your dot's name? Dotty. Dotty. I try, I had it Alfred earlier. And then like many people are calling it Dotty for now. But I'm trying to like think of a great name. Yeah, yeah. Mine's name is Boo. Uh That's a good one. Yeah, it's a good one. Um What do you think so far? I know it's like super early and still rough. I I like it. It has become the default way that I use chat, which I think is a very promising thing. Me too. And I I shared with you like I think, you know, I'm I'm in the beta. so I was in the the early period. So I I think it's it, you know, it has limits. It's like it has been a little bit buggy for me. So Yeah. There's a lot of different surfaces that it can be used on, like mobile and it has its own app. computer and like all that kind of stuff. and I think those kinds of things getting that cleaned up is a big deal, but yeah. That that I I think this is the kind of thing where it's going to get like better rapidly each week from here, but it's going to definitely take some time to smooth it all out. I had a similar example where I was trying to reschedule a flight and it was like, hey, you have a meeting during this flight that you're rescheduling, but it wasn't on my calendar. It had just like read it in slack. Yeah. And then it was like, yeah. That that thing about the ability to bring context together from across your entire life and like really do things in ways that would be hard even for an attentive person to get right has been super impressive. That that's been one of the things I found most useful about it. What had to happen from a research perspective to make this possible? This is like a very complicated system with a lot of stuff, but may maybe most important was Astra. We needed like a you needed a smart enough model. And then there were a lot of other things you had to build too, but but getting the model over the intelligence bar was a key part. You told me that there's a there have been a couple times, not just for like scheduling or or or things that are urgent like uh filtering, but there have been a couple times where it's helped you see something or make a decision with an AI that you could not have. You said you couldn't have hired someone to do that. Like it it'd just be too complicated. Do you have any specific examples you can share? There there was one like new feature that I was kind of excited about. And it would have been very hard to explain it to people in enough detail to get it built. Um and even then there were like a lot of ways to do it and I wouldn't have been confident in which one. And I just was like I would just like constantly jot notes as I thought of something or as I was like on a walk or whatever to my dot. And then the dot went off and built like five or six versions. I was able to give feedback. I was able to iterate and it was just like whenever I had a spare moment or had an idea, and the kind of like the ability to just like dump that in raw form, the ability to like get and give pretty continuous feedback on it was very cool. And then, so there's like many categories like that where like uh an idea that would be you'd be so unsure about or it would be too much effort for you or for someone else to build. You can actually just kind of like continually iterate on it and and it's it's like it's pretty cool to go to sleep and like just leave a long voice note as you're kind of drifting off with a bunch of thoughts. And like wake up in the morning, not only is the thing built, but it like went off and did like the next seven steps. It's the best. And it's like, all right, this is like a very cool way to express stuff. And then and then also like there there's a set of tasks that I would normally give up on or that I would have given up on before. There was a thing that I was sure I had seen in slack. And slack search is obviously terrible. so I mostly use Codex to search it and I had been searching for this thing because I wanted to find it again, I wanted to reference it. And I spent like a good 10 or 20 minutes trying to search with Codex and find this and I kept trying different things and didn't find it. and I was like, man, like maybe I dreamed it and this didn't really happen because Codex doesn't usually fail. And I kind of gave up on it. And then the next day, um, my dot was like, hey, I noticed you were trying to do this thing and you didn't find it. But I like kept thinking about it overnight and I found it for you. Wow. I was like, what? And it was a screenshot. I had forgotten in my head it was a screenshot, not text and slack. And it just like sat there very patiently looking for something to do to be helpful to me. and like looked at like image wrecked every screenshot and found the thing. And I was like, that's really, yeah, that's really good. I would have given up on that. I think there's a you're we're constantly feeling like there's way too much to read in slack on email and all that kind of stuff and it's it's perfect for that amount of context to just like get you the right thing at the right time. People have talked about the need to have an agent deal with the just onslaught of the internet and all of the various things trying to like hack our attention and distract us and whatever. And I've definitely learned in the few weeks that I've been using dots, that although technology is very wonderful in many ways, the kind of like crazy attention hacking of social media and phones and notifications had gotten way worse than I realized. for me personally. You hear a lot about AI tools on this show, but the one that's truly at the center of your work is your CRM. It's where all of your customer contacts lives and where the work either happens or gets missed. Addio delivers the work that you never had time to focus on before and more. It picks up leads at 2:00 a.m. It catches renewals before they slip, and drafts the follow-up before you've even thought to send one. Ask Addio anything about your business from this week's forecast to which deals have gone quiet, and get the answer in seconds. And because it's extensible with an API and an MCP, your agents work right alongside your team to build pipeline, advance deals, and grow accounts around the clock. Try Addio for free at addio.com/every. That's addio.com/every. And now, back to the episode. Tell me about space. It's it's a big bet to do three like office suite type tools in one launch. We didn't set out wanting to do this. And then we realized, and I'm sure you've realized your own version of this too, that software was really not designed on the whole for collaborating with AI. Yeah. And you can use it yourself, or you can have AI use it on your behalf, but to really collaborate closely, especially when it's multiple people and multiple people's AIs all working together, like a lot of the current abstractions we have break down. So, this is definitely an ambitious bet, but, uh, but I think the the space of human AI collaboration is really important to explore. How are you working with it? Like, are you reviewing docs in space or are you making docs? And do you have expectations for how much of it you're typing versus how much of it your dot is doing for you? And also for the the with the docs that you get? Like are you are you reviewing it from that perspective too? If there's like a new idea, um, I like writing the old fashioned way because it helps me like think about it. So I still do that and then maybe have my dot or other people's dot like do the follow-ups, but one of the things that I have found most amazing about it, and I don't think we have a word for this yet, but for lack of a better word, these like live documents. Um, creating a doc that is going to automatically update as people interact or as other things change in the world. And something that like many of our dots can work with together, not only to like share information, but but actually like, you know, seven people's dots can working together create something new that kind of existed in all of our heads to some degree, but was like an emergent thing. That's been very fun as we think about like new features to explore. Is there a specific live doc that has been most useful for you? I have switch, like the dashboard, I think everybody's probably got like a dashboard that they look at every day for how the company's doing in the way that they care about. And you know, I used to share one with everybody else or many other people. And I now have one that is kind of like a changing summary of the business just for me based off of like the evolving things that I'm caring about or not caring about. Like looks at all my slack messages, makes me a new thing, like constantly updates it. And it's like, oh, this just feels like a living document. That's sick. What what do you care about now? A lot on my mind recently is just how we're thinking about the safety, alignment and security challenge in front of us. Yeah. Um, but, you know, six months ago, I probably cared more about revenue and growth and metrics, uh, of this sort of traditional part of the business. So, watching that evolve, I I've actually been thinking about how I can like, because I'd be very curious to see what my dot thinks I care about each day the most over time. And we don't currently have a good way to visualize that. I have not been able to get my dot to figure out a cool way to do that. And I I would love to see that over the course of years. I've been. I think that's really cool. I've been wanting one that just like, it's a live dashboard. sort of my mood uh, because I I have my dot like when I get a slack message and I'm like, I don't know if I want to look at this slack message. I have my doc go look at it and then I'm like, should I look at this now or later? And it's like, you should look later. Um, and I kind of want like a live thing of my mood so it knows. Totally. Yeah, yeah. I I want that too. I would use that. Um, very cool. Tell me about decisions API. I feel like the decisions API is one of the sneaky most important things today because Jev is this new hot thing that has honestly changed how we think about building things internally. Is this decisions API, did you start that from Jev or has that been in the works for a while? How do you think about it? Diogo uh used to work at OpenAI. and um, Diogo's awesome. and he was talking about this class of ideas for a long time. Um, and we had been working on trying to figure out how to make it work. We we like the number one request of Luna and this is like an example of how people always want like more cheaper, faster, better. But the number one request with Luna, even though people love that model, was I want it to be even cheaper and even faster. And so we had been thinking about how we could push it further in that direction. Um, definitely seeing the response to Jev made us more excited to prioritize things in that direction. Um, but and like Diogo has inspired me for a long time on thinking about this. Um, but I think you will see us try to push you know, in different ways, in the spirit of make the technology super fast and super low cost for common use cases. How do you think about you know, if something like Jev comes out, it's a kind of new type of model and then OpenAI can can add like a decisions API. How do you think about how that changes the race between new labs and labs and and what what OpenAI can cover and what a new lab can focus on. As I understand the approach of Jev, it's pretty different and I don't think we can go, I don't, well, maybe we could go do something from scratch like that. But I think we have other things to go do. you know, like we want to make the best, cheapest, fastest, most capable model at every point on the Pareto curve of cost and speed and modality. But I think there will be new approaches, new architectures, new technologies that are things we just can't cover all of. But for the the kind of main line workhorse, like we will try to deliver that. So I've been like in your ear a lot about Codex native apps for a little while now. You finally did it. And and you finally did it. I'm psyched. Um so plugin extensions. Um, yeah, I I and and also being able to bring your tokens from your chat subscription to an app developer, like it changes the equation for someone like me where normally you're you're thinking about, okay, how much AI can I let my users use so that it doesn't bankrupt me and then you're like, can I use should I use uh lower price models and then the experience is worse. But if someone can bring their subscription, I'm like, great, you can use whatever model you want under your subscription. What are you most excited about for the kinds of applications and products that can be built at the application layer in this new world. I mean, I'm most excited for the ones that we haven't thought of. Like there's many reasons that we wanted to do this, but one of the biggest is and why we do this big Dev day thing is just that this this very deep belief that there are way more good ideas outside of our building than inside. and I don't know what I'm going to love the most as like a ChatGPT native app, a plugin extension, but I am very confident that there will be things that I really do love that I cannot imagine today and that no one at OpenAI really could imagine with conviction today. and that's why I mean, that's one of the reasons why this is so important to do. Like we we want to find a way to get that creative spirit into people's hands and we want to find a way to help builders connect with people who are going to find the technology useful. In terms of bringing your ChatGPT subscription with you everywhere, like we just want that ourselves. There's all this great stuff that people are building. There's all these new apps we want to try them. Um, And and I like, I hate having to sign up for 100 subscriptions. And I also like when I try something, I'm like, hey, I, you know, like I'd like to be able to use this with like full juice AI. Um, so we we want to make it just super easy for people to take it with them. and it's your subscription like it's you should get to use wherever you want. Do you think anyone at the app player is doing this well today? Like building interesting apps? Well, not with the new stuff just because it's brand new, but but yes, I mean, I think there's an unbelievable for all of the fear people had a couple of years ago about, man, AI model companies are just going to eat everything. There is an unbelievable amount of significant new companies, great new products building this stuff all of the time. I I've never seen startups get as successful as quickly. Does that change, you know, you you were part of Y Combinator before Open AI, does that change how you think about when people ask you about, how do I start a startup or how do I think about how to grow, like all that kind of stuff? I would love to be a startup investor again. Um, I mean, because like, oh man, what a fun time. Like the the watching what people are doing right now, getting to spend time with some of these sounders, which I don't get nearly as much as I like. It is it is just a golden age. What are you looking for? I mean, obviously besides the founder themselves, like what are you looking for in the shape of things that you'd be like, oh man, I want to invest in that. Um, I was always a very founder-driven investor. Like I I I have my pet projects like anybody else. But my my whole MO as an investor was like, find the absolute best founders I can that are working on the ambitious projects and be willing to take a lot of risk with them. Yeah, I I feel like there's always that narrative of the Netlabs are going to eat everything and uh Dev days like the the the tweet that always works is like Open AI just killed X startup or whatever. The tweet works because people like doomerism, but it never actually happens. Yeah, why is that? Why does it never happen? Yeah. Well, A, I think we have a lot of conviction that empowering builders and not just trying to be like the one giant evil Megacorp is important. So, like, I think we took a lot of steps today that no other AI lab would do about saying, we're going to give you all of the same tools we use. We've always pushed the frontier on on model access. And so, I think we're like uniquely committed in the industry to doing this. And we we really do love developers and I think that's partly I just also love startups and that's where I came from. But I also think, speaking of like a new Renaissance, it's just there is so much to explore. People are going to do such incredible stuff. We are going to have such a tiny fraction of that vision ourselves that, you know, like the world has a fundamentally new enabler. This is like I mean, I don't even know what this is like fire or something. This is like a real change to what's possible. You know, maybe we like brought it down from the mountain, but we want to give it to everybody and we want people to like do stuff with it. There's a lot of um basic startup lore, like you should always have a co-founder or, you know, even the structure of how Ytor does investments that is based on a previous, for example, cost structure. Yeah. If you were going to start those things from scratch or think about them from scratch, like what do you think changes? There are a few things that are probably just like invariance. You know, you've got to like make something people want. That advice was good a million years ago. Or agents want. You got new kings. Yeah. Well, yes, but at least by our understanding of the world today, like, at some level people want something. Yeah, yeah, yeah. You got it. Um, so there's some stuff that just hasn't changed. And then there's a lot of other things like how much work you can get done with one person or two people or whatever that feels totally different. And I don't think there has been enough re-examination of that in the sort of startup conveyor belt of Silicon Valley yet. As a person who's running a gigantic company and generally probably wouldn't have time for building stuff and it now seems like you are, how does that change how you are able to lead or what you're able to do that you can like make stuff on the side? I mean, we're not a gigantic company, first of all. Like there's probably, I don't know, like maybe the big tech companies have like 100 times more people than we do. like that. But I am way, I I I I really believe in if you lower friction a little bit, people build way more. And I have really felt that now with dots and these like, you know, little chunks of time I have or even a like a long chunk of time where I can like have this interactive thing and get a lot done. Uh, I I I feel like I will do something that is not just like a little side project for myself like I used to do with Codex, but like I hope I launch something to the world. Like I hope I actually make a cool new thing and like get it over the threshold. And like does that go out on like samman.com or you talking about like a real like open eye product? Uh, I meant like I would try to like build a plugin extension or something and put it out. Yeah. That's that's Maybe I'll maybe I'll ship a real open AI product. And like Feels like all bets are off now. And so, you know, you have to make that decision, do I? You know what? I will do that. I'm inspired now. What's do you have I no, I don't know. I'm going to go think about it. now I want to do something. I love it. I love it. But you you do have to make the decision of like you could just put a team on that. So what do you get to what does that add that instead of having to put a team on it, you can now be like, oh, I'm going to make a plugin. Uh well, I think today was an example of this. Like you talked about how the amount of stuff every dev day has gone up. Like we could not have done all of this stuff without having AI to help us. And if this works, like next year we will put the teams on it and I will do something and more people will do something and we'll have like a crazy amount of amazing stuff. Or just one really great thing. That'd be fine too. But like hopefully the total amount of like utility we put out into the world goes up. Do you do all your prompting on Ultrafast? I do. That's sick. Yeah. That's crazy. No, I've been I've been thinking about this. Um I I worry that in some ways we kind of get out of touch with people building on our tools and we try we've tried to like put all these constraints on ourselves. You know, I'm like very aware that not everybody is going to be able to use ultrafast all the time for everything. but I have such a hard time dragging it back over. Given that you're like living in that future that I think, you know, all of us can experience a little taste of it. What is what is different about the ultra fast world in terms of how you can use these tools? There was an amazing talk like, I don't know, 15, 20 years ago from Brett Victor about kind of the principles that he thought people like how he did his work. And one of the things, I'm going to get this a little bit wrong, but it was something like creators should have an immediate connection to what they're creating and how speed and the feedback loop was so important. And I did not realize what a difference that makes to my kind of iterative thinking loop to have responses come in so quickly, but but in terms of trying to build something, trying to think through a problem, it's like a wild difference. I know you've been thinking for a long time about AI as a democratizing force to allow to allow anyone to build stuff. and in the sort of ultrafast world or in the asterism more expensive model world, I know you're also thinking about that means something that people can pay more for better intelligence. So like like what what are you thinking there? So, historically, what we have mostly focused on and I think done a great job of is bringing down the like price of IQ. And we have made hugely more capable models that get dramatically, dramatically cheaper every few months or every year. Now, there's this new vector that is clearly going to be very important of speed. And we are putting the investments in place. Like our own chip is going to be an incredible contributor to letting us get very fast inference in the hands of everybody, to say nothing of what the models will do to improve research and our our inference stack and everything else. And I think we now realize that we've got to go invest in the democratization of this like we have for model intelligence. So, when can I expect to use ultra fast for all my prompts? you can pay us for it and do it now. Like uh I don't think I can afford it. Um, I think we can deliver 8x at a great price in the not distant future. But then we'll have 100x faster. And like, you know, and they'll also be like, you know, GPT7 Galaxy or whatever. Um, and so I I think, I I think the right way to think about this is, as far as I can see, there will always be a real frontier in terms of intelligence or speed or whatever else at a higher price. But I think you can also bet that we will keep doing what we've done of just a ridiculous, like put Moore's lot of shame, relentless focus on driving value per dollar up and up and up, kind of in this big compound way. Amazing. Sam, thank you so much. Thank you. Thank you all for coming. I didn't realize there was going to be like a live audience, but that was fun. See ya.