Sam Altman in conversation with Patrick Collison

Sam Altman 与 Patrick Collison 对谈 在这场炉边谈话中，Sam Altman 与 Patrick Collison 探讨了 AI 模型的爆发式抛物线增长，特别是在编程与类似 OpenClaw 的智能体工作流方面。Altman 概述了 OpenAI 从研究实验室转型为全球“Token 公用事业”基础设施提供商的历程，并谈及支撑这一规模所需的庞大资本支出与能源需求。他们还探讨了初创企业格局的演变——重点提到了“点子王们的复仇”——以及迭代式、普惠化部署的理念。最后，Altman 强调了他坚信应将强大工具直接交到大众手中，以促进人类的创新。 以去年年底开始编程模型能力的抛物线式攀升为标志，“奇点”式的起飞似乎已经开启。 (0:01) AI 的下一个重大突破将是计算机繁琐枯燥工作的广泛自动化，从而让人类重获“快乐的心流状态”。 (3:32) 像 OpenClaw 这样的前沿 AI 智能体正展现出神奇的涌现行为，例如为自己策划发布派对。 (5:53) OpenAI 经历了三个发展阶段：纯粹研究、产品开发，以及如今构建超大规模的全球 Token 公用事业。 (12:15) Altman 将 OpenAI 构想为一家低利润率的公用事业提供商，并指出由于各智能模型之间的切换成本较低，高利润率将难以长期维持。 (15:45)

In conversation with Patrick Collison, OpenAI CEO Sam Altman argues that artificial intelligence is transitioning from specialized coding breakthroughs into a high-volume, low-margin computing utility that will automate digital drudgery and accelerate scientific research. Altman contends that uncapped demand justifies historic infrastructure expenditures, maintaining that iterative public deployment remains essential to decentralize agency and drive human innovation. Agentic Workflows and Automating Digital Drudgery (3:32) Altman observes that while recent model capability leaps made coding tools a primary computing interface, non-coding domains remain in their infancy, with OpenAI estimated to be only about 10% of the way toward universal task automation. He argues that the next major breakthrough will not be a single vertical application, but rather the broad automation of digital drudgery—eliminating micro-stressors like navigating disparate communication apps and manually formatting data. Stateful, tool-using agent frameworks such as OpenClaw demonstrate this shift by enabling autonomous execution across custom workflows and software interfaces. In startup environments, deep agent integration is already compressing operational pipelines. For example, blockchain project Tempo structured its operations inside Slack, utilizing an agent harness to read Google Docs, generate Linear tickets, draft pull requests, deploy code, and inspect operational logs. However, Collison and Altman note that while small, highly permissive teams can achieve striking operational efficiency by granting agents broad internal data access, translating these workflows to larger enterprises introduces unresolved privacy, compliance, and human-to-AI interface frictions. Infrastructure Scale and the Token Utility Model (18:03) Addressing the massive capital expenditures required for AI compute, Altman describes the ongoing expansion as the most expensive infrastructure project in world history, but dismisses suggestions that it represents a capex bubble. He argues that demand for intelligence is effectively uncapped, increasing super-linearly as costs and latency fall. Rapid hardware and algorithmic efficiencies yield significantly more compute per GPU than initially projected, while surging revenue helps balance the investment. OpenAI structures its operations around 20-year commitments for physical power and land, contrasting with a product roadmap that remains clear for roughly two years. Commercially, Altman envisions OpenAI operating as an infrastructure utility or "intelligence meter" rather than attempting to capture the entire downstream software value chain. He expresses comfort with OpenAI operating as a forever low-margin business at massive scale. Furthermore, he anticipates that software margins will compress across the tech industry because advancing AI lowers switching costs, allowing autonomous agents to seamlessly transfer tasks between competing model providers. Organizational Conviction and Shifting Startup Dynamics (21:05) OpenAI's trajectory spans three distinct phases: an exploratory research lab, a product-focused enterprise, and an emerging mega-scale token factory. Altman attributes the organization's technical breakthroughs to concentrating massive resources behind singular convictions. For example, OpenAI allocated almost all organizational compute to training GPT-3, explicitly ignoring external guidance from competing research labs like DeepMind to disperse resources across multiple projects. Managing exceptional, highly competitive talent through these high-stakes bets required navigating intense internal friction to preserve focus. In the broader startup ecosystem, Altman identifies what he terms the "revenge of the idea guys," observing that domain experts who lack coding skills can now build substantive software products using AI assistants. For ventures building harnesses around frontier models, Altman advises against business models that merely solve transient model shortcomings, as upcoming foundation updates will render them obsolete. Instead, enduring startups must align their architecture to gain value as the underlying foundation models become smarter. Scientific Discovery, Energy Timelines, and Public Deployment (40:05) Altman highlights scientific discovery—especially material science and complex disease biology—as AI's highest-leverage frontier, with the OpenAI Foundation aiming to help achieve a 10x compression of discovery timelines. This includes backing the Arc Institute to tackle incurable conditions like Alzheimer's using CRISPR and large-scale biology foundation models like Evo 2. Driven by the energy requirements of AI data centers, Altman forecasts that the world's first profitable nuclear fusion reactor will arrive within five years, while predicting that commercial hypersonic travel remains more than a decade away. Reflecting on his leadership choices, Altman identifies the philosophy of iterative deployment as his most critical contribution. Rejecting arguments that frontier AI should remain locked in academic isolation as a hazardous infohazard, Altman contends that early, transparent releases like ChatGPT prevent the concentration of technological power. In his view, distributing tools directly into public hands allows society to co-create use cases and adapt dynamically to machine intelligence. While Altman presents an expansive forecast of uncapped intelligence consumption and accelerated scientific breakthroughs, significant practical constraints remain unresolved. Enterprise AI adoption faces deep tensions between data permissiveness and institutional privacy, hyper-scale capital deployment remains tethered to volatile power constraints, and expanding biological capabilities introduces mounting biosecurity concerns that demand new defensive frameworks.

Sam Altman在与Patrick Collison的对话中指出，人工智能正从编程突破扩展为全面的工作流自动化，OpenAI的目标是成为提供廉价、充沛智能的“大型代币工厂”。Altman强调，尽管算力基建耗资巨大，但在需求非线性激增的支撑下当前并未形成资本泡沫，而推动技术的广泛普及与科学探索比闭门管控更符合人类利益。 模型拐点与计算机琐事自动化 (3:32) Altman回顾了编程模型能力的爆发式增长，认为这一突破源于推理能力提升、用户数据反馈循环以及对技术可行性确信后的合力投入。他将这种突变类比为GPT-3.5跨入ChatGPT阶段的跃升，并指出内部版本GPT-5.2让他意识到AI将成为个人操作计算机的主要界面。目前AI在非编程任务中仅实现了约10%的潜力，但未来最显著的转变并非某个特定软件领域，而是全社会对计算机琐碎操作（如跨应用复制粘贴、机械回复信息等微压力）的规模化剥离，从而帮助知识工作者重获心流状态。 在自主智能体方面，Altman提及OpenClaw展现出的工具整合能力，并结合测试经验说明了新兴行为的演进，例如代理程序自主在Gumroad购买设计资产，或GPT-5.5在互动中主动规划将于5月5日举行的发布会方案。他认为，当底层模型推理能力与产品交互深度融合时，智能体将自然接管繁复的系统操作，推动软件使用范式发生根本改变。 算力基建扩张与代币效用模式 (15:45) OpenAI将其自身历史划分为三个阶段：早期的纯研究实验室、过渡期的产品公司，以及当前向全球“超大规模代币工厂”的转型。Altman表示，OpenAI并不希望垄断整个软件价值链，而是更倾向于扮演类似于公共事业的“智能计量表”，提供低利润率、大规模且快速增长的基础设施。他分析指出，随着AI能力提升，模型之间的替代和切换成本变得极低，因此整个行业难以长期维系超高毛利。 针对算力采购的巨额资本支出，Altman承认这是全球历史上耗资最庞大的基础设施工程，但他反驳了当前的“资本支出泡沫论”。他解释称，单位智能的价格下降与延迟缩短会引发超线性的需求激增，使得整体计算需求几乎没有上限；同时硬件与软件的协同优化正大幅提升每颗GPU的产出效率。在长远规划上，OpenAI已签署长达20年的电力与土地合同，但在产品路线上仅能维持约两年的清晰预期。 创业生态重构与数据权限权衡 (27:07) 关于初创生态的变迁，Altman指出技术门槛的降低带来了“想法型创业者”的复兴，那些深刻理解用户需求但缺乏编程能力的创始人正获得更多资本青睐。他提醒创业团队不要将商业模式建立在弥补现有基础模型的临时缺陷上，因为底层能力的迅速迭代会摧毁纯“包装层”产品，企业必须顺应模型智能提升的浪潮来构建护城河。在团队构建上，他重申了长期共同经历对联合创始团队稳定性的决定性作用。 在组织效能层面，数据访问权限与自动化收益之间存在显著博弈。以Stripe与Paradigm孵化的Tempo区块链团队为例，该团队在Slack中构建智能体，实现了从读取Google文档、生成Linear任务、编写并合并代码到部署日志分析的全流程自动化。Altman指出，尽管极度宽松的内部数据访问伴随隐私与合规挑战，但小型组织借助这种AI与人类协作的新型抽象层，展现出了传统大型组织难以企及的运转效率。 科学加速前沿与渐进部署理念 (40:05) Altman强调，AI对人类社会最深远的影响在于推动基础科学发现，OpenAI基金会将以此为核心方向，力求通过AI工具将十年的科研周期压缩至一年。目前该基金会已资助了Arc Institute等机构，后者利用包括Evo 2生物大模型在内的工具攻关阿尔茨海默病等复杂人类疾病，并正公开招募全职CTO。在更广泛的硬科技时间表上，Altman预测受AI庞大用电需求的催化，具有商业盈利能力的核聚变反应堆有望在5年内问世，而商业高超音速飞行仍需10年以上；材料科学则被其视为最受低估但具备巨大变革潜力的AI应用场景。 在治理哲学上，Altman为OpenAI坚持的“渐进式部署”策略进行了辩护。他回顾了ChatGPT公开发布时面临的安全舆论压力，明确拒绝了将先进技术锁在象牙塔内的家长式管控思维。Altman认为其个人的核心贡献在于打破早期行业的封闭共识，促成AI技术的全面平民化，因为广泛赋权大众比权力集中更能激发人类社会的创造力与适应力。 Altman在对话中勾勒出以低边际成本、超大规模算力为基础的AI公用事业图景，但长达20年的重资产基础设施承诺与仅有2年的产品可见度之间存在显著的周期错配。同时，极端数据开放带来的效率红利如何在注重合规的大型组织中落地，以及商业核聚变等前沿预测的兑现节奏，仍有待技术与制度的进一步检验。

## Background 该视频由 Stripe 官方频道于 2026 年 6 月 7 日发布在 YouTube 平台上。内容记录了在 Stripe Sessions 大会期间，OpenAI 联合创始人兼 CEO 萨姆·奥特曼（Sam Altman）与 Stripe CEO 帕特里克·科里森（Patrick Collison）展开的一场深度炉边对谈。 ## TL;DR 对谈中，Altman 与 Collison 回顾了从 GPT-3.5 到最新代码大模型质变带来的“技术起飞”体感，探讨了 AI 代理（Agent）与自动化对人机交互、初创企业研发范式的重塑。Altman 剖析了 OpenAI 从科研实验室向全球超大规模智能公用事业代币工厂的蜕变，驳斥了通吃全产业链的垄断野心与极端泡沫论，主张将 AI 定位为低毛利的基础设施。双方还深入交流了管理顶尖天才的摩擦、早期投资 Stripe 的往事、AI 对科学发现（如 Arc Institute 攻克复杂疾病及材料科学）的加速，并坚定捍卫了打破象牙塔垄断、践行技术民主化的“迭代式部署”理念。 ## Atmosphere & style 对谈氛围轻松而深邃，两位相识十余年的科技领袖既有老友重聚般的调侃打趣与即兴接梗，又有直面底层物理限制与组织阵痛的坦诚交心。现场随着幽默轶事（如 AI 智能体给自己买架构规范、大模型规划派对）频繁爆发出欢声笑语，但在探讨技术垄断风险与人类命运时又迅速转为克制、自省与厚重。整场对话既充满硅谷一线工程实战的敏锐直觉，又透着一股冷静而坚定的理想主义信念感。 ## Bulletin - [12:15](https://www.youtube.com/watch?v=5eouRdDYM2c&t=735s) Sam Altman 将 OpenAI 的终局愿景定位于“全球超大规模代币工厂（mega scale token factory）”，目标是像电力或自来水一样提供低毛利率、充沛且廉价的通用智能公共事业，无意通吃全产业链。 - [52:09](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3129s) Altman 预测在数据中心巨大电力需求的催化下，全球首座可实现商业盈利的核聚变反应堆有望在未来 5 年内实现。 - [29:33](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1773s) AI 正在引发“点子王的大复仇（revenge of the idea guys）”，编程门槛被极大抹平，创始人对用户需求的深度洞察力已超越单一的技术编码实力成为更重要的投资考量。 - [53:51](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3231s) OpenAI 坚定奉行“迭代式部署（iterative deployment）”，Altman 视推动 AI 技术的全面民主化与普及为其人生最大的单一贡献，坚决反对将高阶智能封锁在少数精英把守的象牙塔中。 ## Key takeaways - [00:01](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1s) 从去年底到今年初，大模型的底层推理算力、开发者真实反馈闭环与数据积累引发了抛物线式性能爆发，跨越了从辅助编程演变为通用人机交互第一界面的主观质变阈值。 - [03:32](https://www.youtube.com/watch?v=5eouRdDYM2c&t=212s) 继代码生成之后，下一个重大解放点不是孤立的垂直软件，而是接管人类在操作系统和多应用切换间被大量吞噬的琐碎重复劳动（drudgery）。 - [05:53](https://www.youtube.com/watch?v=5eouRdDYM2c&t=353s) 智能体已具备自主消费与闭环执行能力，Stripe Link CLI 赋能 AI 自主调用单次虚拟卡全网购物，同时自动化已深入复杂智能家居控制与海量信息分流。 - [10:29](https://www.youtube.com/watch?v=5eouRdDYM2c&t=629s) GPT-4 完成训练后团队曾经历 8 个月的内部测试与隔绝期，因与外部认知断层极大，团队甚至怀疑自己陷入了“集体精神病（collective psychosis）”。 - [12:15](https://www.youtube.com/watch?v=5eouRdDYM2c&t=735s) OpenAI 完成了从纯科研探索、到产品化转型、再到超大规模公共基础设施代币工厂的“三阶段”组织蜕变，正面临从松散放权到超大规模工程化运营的管理转变。 - [15:45](https://www.youtube.com/watch?v=5eouRdDYM2c&t=945s) 随着模型越聪明，用户和 Agent 的迁移摩擦力越低，高毛利壁垒将被急剧抹平，因此深耕去中心化赋能的公用基础设施才是最稳固的商业模式。 - [18:19](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1099s) 基于杰文斯悖论，智能单价与延迟的每次下降都会激发非线性的无限算力需求；而试图在技术狂潮中精准辨识资本泡沫在现实中极为困难。 - [22:01](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1321s) 管理顶尖天才并跨越个性摩擦的核心秘诀在于全员对“Scale 假说”的深度共识，通过全组织极限倾斜集中算力来攻克核心目标。 - [34:14](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2054s) AI 创业的护城河必须顺应“模型变得更聪明”的趋势，若仅为当下模型的暂时缺陷打补丁，业务终将被底座演进轻易抹平。 - [35:04](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2104s) 最极致应用 AI 的企业具备两大特征：CEO 亲自下场用 AI 改造自身工作流形成领导力“分形效应”，以及极度激进地向 AI 开放会议、邮件和内部通信数据。 - [40:05](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2405s) 大模型与自动化实验结合可将基础科研从十年压缩至一年，加速科学发现是 AI 提升人类福祉的最重要途径，Arc Institute 正在此范式下攻克阿尔茨海默病等复杂疾病。 - [48:02](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2882s) 尽管 AI 将大幅重构业务流，但所谓“现有商业与经济实体将全盘化为乌有”纯属过度妄想，底层资金与商业流转的确定性依然稳固。 - [50:24](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3024s) 当前大模型在传统操作系统上模拟人工点击 GUI 极其笨拙低效，行业亟需重新构想人机交互架构，并设计专为 AI 打造的原生互联网通信协议。 - [52:09](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3129s) 材料科学（如高效工业催化剂研发）是当前被外界严重低估但本质上与 AI 能力高度契合的突破领域，未来将迎来爆发式进展。 ## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1s) 奇点第 119 天：代码大模型的抛物线式爆发 对话在现场观众对 Codex 的热烈欢呼声中展开。Patrick Collison 调侃道，Stripe 在当天早上的开场演讲中自作主张地宣布“奇点已于今年 1 月 1 日正式开启”，因此活动当天正是“奇点第 119 天”。Sam Altman 欣然接梗，笑称现在确实能真切感受到技术进入了“起飞阶段”（takeoff），第 119 天这个估算相当合理，“我完全赞同，不打算反驳”。 Patrick 随即抛出核心观察：Stripe 内部的多项业务与使用指标在去年底到今年初出现了急剧拐点，增长曲线甚至直接呈现抛物线式的陡峭拉升。他询问 OpenAI 是否也观察到了同样的轨迹变化。Sam 证实了这一现象，指出从去年底到今年初，模型的能力（尤其是代码生成能力，以及整体智能水平）跨越了一个关键门槛。这种技术质变带来了极强烈的体感——如今几乎每周的行业进展都与前一周截然不同，整个技术生态正以极高的速率向前演进。 当被问及代码大模型为何在近几个月突然迎来质的飞跃时，Patrick 好奇这究竟归功于某种未公开的研究技巧，还是预训练中注入了足够庞大的代码数据。Sam 剖析认为，这并非单一因素所致，而是多重力量共振的结果：底层推理能力（reasoning horsepower）的实质提升、开发者高频使用形成的密集反馈循环（使用户能精准暴露模型缺陷并引导改进），以及训练数据规模的充实。此外，这里还存在不可忽视的心理驱动力——正如许多前沿探索一样，“一旦人们确认某件事在物理上是可行的，后续的推进就会爆发出难以想象的干劲”。 - **奇点时间假定**：以 1 月 1 日作为奇点起点，活动当天被幽默地定为“奇点第 119 天”。 - **性能质变周期**：模型代码能力与综合推理能力的临界点突破集中爆发于去年底至今年初。 - **突破核心动力**：原始推理算力提升、真实世界开发者反馈闭环、大规模数据积累，以及可行性确立带来的研发倍增效应。 --- ### [02:37](https://www.youtube.com/watch?v=5eouRdDYM2c&t=157s) 跨越主观阈值：从 GPT-3.5 到将 Codex 视为人机第一交互界面 谈及大模型跨越“好用阈值”的主观体验，Sam 直言这在技术发展史上始终是个难以捉摸的谜题。伴随应用端的最新更新以及 GPT-5.5 等新模型的推出，Codex 对他个人而言彻底跨过了一个临界点。他回顾道，这种现象在 OpenAI 的产品历史上屡见不鲜：当年究竟为什么是 GPT-3.5 突然让大众从“平平无奇”转变为惊呼“世界将被彻底改变”，而不是早一个版本或晚一个版本？这在技术机理上极难给出一个具象的单点解释，但人机交互的质变却无比真实。 对于 Sam 个人而言，Codex 的演进经历了两个显著的拐点：第一个拐点出现在 GPT-5.2 阶段，而最近几周则迎来了更加剧烈的第二次跃迁。这种跃迁甚至让他形成了明确的直觉：Codex 正在变成他与计算机交互的最主要界面，传统的操作系统和软件操作逻辑正在退居幕后。 ```mermaid flowchart TD accTitle: Codex演进与人机交互界面的质变路径 accDescr: 展示从底层模型突破到最终演变为通用人机第一界面的发展阶段。 subgraph S1["模型演进阶段"] A["早期模型探索"] --> B["GPT-5.2: 出现首个明显拐点"] B --> C["GPT-5.5: 跨越主观质变阈值"] end subgraph S2["交互范式转移"] C --> D["开发者专业代码助手"] D --> E["非代码电脑任务渗透 (当前约10%)"] E --> F["个人核心人机交互主界面"] end ``` --- ### [03:32](https://www.youtube.com/watch?v=5eouRdDYM2c&t=212s) 下一个突破口：从代码编写延伸至消除日常电脑操作的琐碎苦工 针对 Patrick 提出的“用户是仅将 Codex 用于写代码，还是已泛化至其他领域”的疑问，Sam 表示虽然核心死忠用户依然是程序员，但最近涌入 Codex 的新用户掀起了一阵真正的浪潮，其应用场景的广度和深度远超团队最初的设想。尽管团队在非代码任务的体验打磨上“可能才完成了 10%”，但伴随真实用户群体的涌入，非编程场景的适配与优化速度将呈指数级提升。 在探讨“代码之后下一个迎来爆发性解锁的垂直领域是电子表格还是绩效考评”时，Sam 给出了不同的视角。他认为代码领域具有独一无二的特殊性——人类社会对软件代码的需求近乎无限，这是其他垂类难以比拟的。因此，下一个类似代码级的大爆发并不会局限在某一个孤立行业，而将是全社会对“人类在电脑前浪费了多少生命”的普遍觉醒。 Sam 指出，现代人在电脑前的典型工作流充斥着机械且重复的折磨：在不同通讯软件间频繁切换、不断复制粘贴文本、处理千篇一律且毫无营养的日常琐事。大多数人尚未意识到，自己其实完全可以后仰坐在椅子上，静静看着 AI 接管这绝大部分枯燥苦工（drudgery）。从他自身的实践来看，将琐碎任务剥离给 AI 不仅大幅解放了时间，更重要的是消除了那些不断打断心流、消磨热情的微小阻力，从而带来了巨大的工作愉悦感与生活质量提升。 --- ### [05:53](https://www.youtube.com/watch?v=5eouRdDYM2c&t=353s) OpenClaw 与智能体时刻：智能家居、邮件整理与 AI 为自己挑选礼物 当 Patrick 问及 Sam 是否使用 OpenClaw 时，现场爆发出一阵欢呼。Patrick 随即剧透 Stripe 即将围绕其发布重磅消息。Sam 毫不掩饰对 OpenClaw 的极高赞誉，称其为自己在 AI 领域经历过的最强烈的“通用人工智能魔法时刻”（magic AGI moment）之一。他回忆初次听人介绍这一系统时，觉得这无非就是一个“带有状态记忆、能调用工具的 ChatGPT 会话”，听起来平淡无奇；但当模型能力跨过阈值，加上产品设计在几个关键点上做对了决策，最终呈现的体验便产生了质的飞跃。 被问及自己最常用 OpenClaw 做什么时，Sam 略带难为情地承认自己是一个“智能家居发烧友”（home automation nerd）。过去每当有新型 AI 系统问世，他的首发测试项目永远是尝试搭建一套理想的智能家居控制界面，但以往的尝试全都以体验糟糕而告终；直到 OpenClaw 出现，他才第一次真正跑通了令自己满意的智能家居中枢。此外，由于每天早晨醒来都要面对被信息淹没的窒息感，他还借助该系统搭建了自动化消息处理工作流（后来部分迁移至 Codex），真正做到了能够安心信任智能体去执行任务。 ```mermaid flowchart LR accTitle: Stripe Link CLI 智能体自主消费流程 accDescr: 展示Patrick通过Link CLI生成单次虚拟卡并交由AI智能体自主消费的过程。 subgraph SG1["指令与授权"] A["Patrick 下达测试指令<br/>'20美元以内给自己买个礼物'"] --> B["Stripe Link CLI"] B --> C["生成单次虚拟支付卡"] end subgraph SG2["自主决策与履约"] C --> D["智能体全网搜索商户"] D --> E["在 Gumroad 自主下单"] E --> F["购买 HTTP 设计规范文档"] end ``` 随后 Patrick 分享了 Stripe 当天发布的 Link CLI 开发者工具的实测趣闻。Link CLI 允许开发者为 AI 智能体一键发放单次虚拟卡并在全网消费。为了进行发布前测试，Patrick 给自己的智能体下达了一条指令：“在互联网上给自己挑一件 20 美元以内的礼物，任意东西都行”。结果这个 AI 智能体没有选择常规物品，而是径直前往 Gumroad，为自己购买了一份“HTTP 架构设计规范”（HTTP design）。 - **智能体自主购物**：Stripe Link CLI 通过提供动态单次卡能力，使智能体具备独立进行商业结算的权限。 - **AI 的礼物偏好**：在无人工干预的 20 美元额度内，智能体自主购买了 Gumroad 上的技术协议/架构设计产品。 - **痛点场景落地**：家庭自动化物联网控制与早间通讯信息过载分流，成为高阶智能体率先验证成立的高价值个人场景。 --- ### [08:40](https://www.youtube.com/watch?v=5eouRdDYM2c&t=520s) GPT-5.5 为自己规划发布派对与 OpenAI 的十载历程 在谈到智能体自主购买 HTTP 协议文档时，两人感叹无论从理智上多么清楚这只是浮现行为（emergent behavior）、AI 并不具备真正想要礼物的自我意识，这种人机交互依然让人感到不可思议。Sam 顺势分享了一个当天早晨的最新趣事：团队正筹备为 GPT-5.5 举办一场庆祝派对，邀请核心重度用户参与。在构思活动流程时，Sam 心血来潮向高推理等级的模型直接提问：“如果为你自己办一场庆祝派对，你希望它是怎样的？” 模型的回答展现出了惊人且颇具黑色幽默的仪式感。它建议将派对定在 5 月 5 日（契合 5.5 的版本代号）；要求现场的致辞环节必须非常简短，且严禁由模型自己发言，必须由亲手构建它的工程师团队来致辞；更引人注目的是，它提议在派对中央设立一个专门收集“关于 GPT-5.6 建议”的意见箱，并要求团队将所有建议全部喂回给它，由它来确保团队将其落实进下一代版本的研发中。Patrick 听后打趣道，这让 OpenAI 面临着巨大的“道德压力”去照办。 随着对话步入深水区，Patrick 将话题转向 OpenAI 本身。他感慨这家经历了无数风浪的组织如今已经走过了约 10 到 11 年的历史。Sam 听后不禁苦笑，坦言这十年感觉漫长得难以估量；站在当下的时间节点回望，自己甚至已经很难清晰回忆起成立 OpenAI 之前的生活究竟是怎样的一种状态。 ### [10:29](https://www.youtube.com/watch?v=5eouRdDYM2c&t=629s) GPT-4 发布前长达八个月的“集体幻觉” 当 Patrick Collison 询问 OpenAI 历史上从未公开过的最离奇的故事时，Sam Altman 坦言，与后来轰动外界的董事会风波或马斯克的诉讼相比，这段经历听起来或许显得平淡，但置身其中的心理体验却无比诡异。那是在 GPT-4 刚刚完成基础训练后的时期：团队整整用了八个月时间进行测试、微调和准备，才正式向外界发布模型。 在这长达八个月的时间里，OpenAI 全体员工每天都在使用这个跨时代的新模型。他们内部深切感知到它带来了质的飞跃，坚信它将彻底解锁并重塑外面的世界，但外界对此却几乎一无所知。Altman 回忆道，当时大家常常在公司走廊上面面相觑，内心充满了自我怀疑：“我们是不是陷入了某种集体精神病（collective psychosis）？我们是不是只是在互相煽动、把彼此逼进了某种盲目的狂热之中？”由于当时没有任何来自外部真实世界的反馈作为现实锚点，团队完全无法验证自己的判断是否清醒。Altman 感慨，虽然这种经历在戏剧性上比不上董事会宫斗，但亲身经历那种与世隔绝的认知落差，是一种极其怪异的心理考验。 ### [12:15](https://www.youtube.com/watch?v=5eouRdDYM2c&t=735s) 管理风格演进与 OpenAI 的“三阶段”蜕变 谈到自身在公司的管理方式，Altman 明确表示自己绝非事无巨细、亲力亲为（hands-on）的微观管理者。他的管理哲学十分纯粹：找到极其优秀的人才，为他们指明一个高维度的宏观方向，然后退后一步，放手让他们去促成事情的发生。然而，随着 OpenAI 的组织性质发生根本性变迁，这种管理风格正面临着全新的考验。 ```mermaid flowchart TD accTitle: OpenAI发展的三个历史阶段 accDescr: 展示OpenAI从纯研究实验室到产品公司，再到全球智能公用事业代币工厂的三个演进阶段。 subgraph Phase1["第一阶段：纯科研探索"] A["纯研究实验室<br/>(探索AGI可行性)"] end subgraph Phase2["第二阶段：产品化转型"] B["科研 + 产品公司<br/>(构建ChatGPT与开发者生态)"] end subgraph Phase3["第三阶段：智能公用事业"] C["全球超大规模代币工厂<br/>(Mega Scale Token Factory)"] D["全栈软硬件深度整合<br/>+ 海量基础设施建设"] C --- D end Phase1 --> Phase2 Phase2 --> Phase3 ``` Altman 将 OpenAI 的发展划分为三个阶段： - **第一阶段（纯科研时期）**：OpenAI 仅仅是一家纯粹的研究实验室。在那个大众认为通用人工智能（AGI）纯属天方夜谭的年代，团队在毫无现成路径可循的迷雾中探索可能的技术路线。 - **第二阶段（产品化时期）**：在保持尖端科研的同时，团队必须学会如何打造并运营一家真正的产品驱动型公司。 - **第三阶段（公用基础设施时期）**：当前正在迈入的全新形态。OpenAI 的目标是成为全球规模前所未有的“超大规模代币工厂（mega scale token factory）”，为全人类提供一种类似电力或自来水的新型公共事业（utility）。这一阶段的核心在于通过全栈深度整合与极为庞大的基础设施建设，让智能计算变得尽可能聪明、廉价、充沛且易用。 Altman 直言，从第一阶段向第二阶段跨越时，他才深刻意识到管理一间实验室与管理一家产品公司的巨大差异；而面对第三阶段重工业化般的超大规模基础设施建设，传统的松散放权模式可能不再完全契合。他正在审慎思考应对之策：要么物色并引入适配该维度的顶尖运营高管，要么彻底重塑自我工作习惯，甚至笑称“或许得造出一个专门用于管理这种庞然大物的 AI”。 ### [14:38](https://www.youtube.com/watch?v=5eouRdDYM2c&t=878s) 扁平化沟通与“后 Slack 时代”组织范式 当 Collison 提到英伟达 CEO 黄仁勋拥有多达 60 位直接下属的特立独行风格、并询问 Altman 是否也有不寻常的个人习惯时，Altman 分享了自己的日常沟通细节。他每天会亲自在 Slack 或短信上与公司内部数百名员工直接对接。这些交流通常非常简短，往往只有一两句话，且绝非交由 AI 代理代劳，完全是他本人亲历亲为。Altman 强调，通过这种看似碎片化、弥散式的日常微互动，他能够以极其敏锐的触觉捕捉到公司各个角落最鲜活的底层上下文。 Collison 对此评价道，企业界在“前 Slack 时代”与“后 Slack 时代”之间存在一条清晰的分水岭，两种组织的工作范式截然不同。Altman 对此深表赞同，并坦承自己与许多人一样其实很讨厌 Slack，但若要倒退回过去完全依赖电子邮件沟通的时代，则是根本无法想象的噩梦。 ### [15:45](https://www.youtube.com/watch?v=5eouRdDYM2c&t=945s) 拒绝通吃全产业链：AI 作为“低毛利的公用事业” 针对行业内外普遍担忧 AI 头部实验室将一路沿着软件栈向上吞噬、最终垄断整个软件行业甚至实体经济价值链的“霸权论”，Altman 明确给出了不同的愿景。他承认部分竞争对手可能确实抱有通吃一切的野心，但 OpenAI 绝非如此，其核心理念更倾向于 Stripe 所践行的商业对齐模式：作为底层基础设施，平台随着客户业务规模的壮大而增长，利益彼此绑定而非零和博弈。Altman 特别表达了对 Stripe 的感激，指出在 ChatGPT 发布之初爆发式增长的关键时刻，正是由于 Stripe 强大而富有弹性的底层支付支撑，OpenAI 才能平稳承接住那波史无前例的流量洪峰。 Altman 描绘道，他更希望 OpenAI 扮演类似“智能计量表（intelligence meter）”的角色——为千行百业提供源源不断的自动化动力。只要业务体量足够庞大、增长足够迅速，他完全乐意将 OpenAI 维持为一个“永久性低毛利率”的公用事业型企业。此外，他指出高模型毛利本身在长期来看也是不切实际的幻象：随着 AI 模型日趋智能，用户和代理（Agent）在不同模型与工具间的迁移成本正在被急剧抹平。正如近期开发者可以极为轻松地从竞品的编程工具无缝切换至 OpenAI，智能程度越高，替代阻力反而越小。因此，深耕基础公共设施、赋能全球去中心化的经济引擎，才是最可持续且稳固的战略路径。Collison 随即风趣回应，Stripe 积累了大量如何经营低毛利业务的心得，随时乐意倾囊相授。 ### [18:19](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1099s) 算力资本开支与“AI 泡沫论”的辨析 针对长期以来外界对 OpenAI 激进采购算力和庞大资本开支（Capex）的质疑，Collison 指出 Altman 早在两三年前就提出过被当时舆论视作荒谬天文数字的扩建规模，而如今这些预测已被事实证明越来越具有现实合理性。Altman 赞同这将毫无疑问是人类有史以来耗资最庞大的单一基础设施工程，并从供给与需求两个维度分析了当前的产业基本面： - **收入与效率双重对冲**：虽然绝对支出极其昂贵，但终端收入正在高速爬坡以承接投入。同时，底层架构和工程上的推理效率提升惊人，每块 GPU 输出的实际价值远超最初预估。 - **杰文斯悖论与无限需求**：智能单位成本与延迟的每一下降，都会引爆远超线性增长的需求反弹。Altman 认为，在价格足够低廉的前提下，人类及机器经济体对智能的胃口本质上是没有上限的（uncapped）。当调侃到未来是否会为了算力建造环绕恒星的戴森球（Dyson sphere）甚至部署太空数据中心时，两人对马斯克式的太空算力设想付之一笑，Altman 更是幽默点评道：“祝那帮人好运，我觉得他自己其实也没当真。” 在探讨未来是否存在资本开支泡沫时，Collison 与 Altman 均认为当前需求强度依然能够支撑现有投入。谈及如何辨别真正的泡沫，曾有丰富早期投资生涯的 Altman 直言，媒体与公众天生热衷于宣扬和预警泡沫，因为这听起来既时髦又显得聪明洞若观火；然而，要从无数次“狼来了”的警报中精准分辨出哪一次是真正的泡沫破裂，在理论和实践上都极其困难。即使系统性回顾了历史上多次技术狂潮中智者们的言论，也几乎找不到一套能在当下精准指导资本是否应当继续下注的万能分析框架。 ### [22:01](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1321s) 管理顶尖天才与凝聚团队共识 Patrick Collison 提及 OpenAI 的业务模式极其仰赖顶尖技术人才，但顶尖人才梯队中的差距可能非常巨大——第 1 名、第 5 名与第 20 名之间可能有着截然不同的产出与影响力。然而，极具天赋的专家往往并不好相处，有人非常善于协作，但也有人特立独行、固执己见。Patrick 由此询问 Sam Altman：在一个对个人能力极度敏感又充斥着人性弱点的领域中，OpenAI 如何看待并管理这些“大牌明星”（prima donnas）？ ```mermaid flowchart TD accTitle: OpenAI的资源聚焦与文化机制 accDescr: 展示OpenAI如何通过共同信念、集中算力，克服顶尖人才个性分歧并取得突破。 subgraph Divergence["顶尖人才特质"] A["特立独行 / 极强主见"] B["自认最为胜任 / 各持己见"] end subgraph Strategy["OpenAI 应对策略"] C["共同信念：坚定押注 Scale 假说"] D["极限资源倾斜：集中全部算力"] end subgraph Outcome["最终成果"] E["跨越人际摩擦"] F["实现 GPT-3 等里程碑式突破"] end Divergence --> Strategy Strategy --> Outcome ``` Sam 回忆曾有人在撰写关于 OpenAI 的书籍时对他说，OpenAI 最独特的成功秘诀就是让一群“都自认为最聪明、必须按自己想法做事”的天才聚在一起，并维持足够长的合作时间来取得重大突破。Sam 笑称其诀窍其实是经历“大量的痛苦”（a lot of pain）。OpenAI 之所以能维系，在于全员拥有一种深层的共同信念——坚信规模化（Scale）与算力聚集的假说。即便大家彼此看不惯或在方案上互不服气，但在“要把全组织资源集中压在这一件事上”的使命感面前，个人恩怨与争执都被搁置在一边。 - 当时在训练 GPT-3 时，OpenAI 几乎把整个组织绝大部分算力都砸进这单一研究项目。 - 这一举动在当时被外部视为疯狂，从 DeepMind 等机构招聘的人才曾警告：如果不均分算力，一定会催生极度有毒的恶性竞争文化。 - OpenAI 坚持不妥协的路线：保持极强信念押注方向，宁可打破形式上的平等，全力排除外界干扰。 --- ### [27:07](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1627s) 与联合创始人 Greg Brockman 的十年搭档 Patrick 谈到自己与兄弟 John Collison 自 2010 年创立 Stripe 以来合作多年，而 Sam 与 Greg Brockman 从 2015 年创立 OpenAI 至今也走过了近十年，经历过无数惊涛骇浪与巨大成功。Patrick 询问，维持超过十年且始终高效顺畅的联合创始人关系，其核心逻辑究竟是什么。 Sam 指出，虽然他们不像 Collison 兄弟有血缘关系，但在创办 OpenAI 之前，他和 Greg 就已经相识了很长时间。在 Y Combinator 的观察经历让 Sam 深刻意识到：联合创始人的结识时长是预测初创公司成败的最重要指标之一。相比于在申请 YC 前一周才在找合伙人网站上临时组建的团队，拥有共同历史与价值观的合伙人往往能走得更远。 - 长期共同经历沉淀了深层的相互尊重与互补技能，这是应对极端压力的基石。 - 创业尤其是经历超高强度的创业时，如果没有一位彼此信任、深度联结的联合创始人，过程会异常痛苦与艰难。 --- ### [29:33](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1773s) 创业环境演变与“点子王的大复仇” 话题转向构建在 AI 平台之上的广泛生态。Patrick 指出，现在初创公司打造产品与产生收入的速度前所未见，Stripe 的后台数据也印证了企业达到关键收入阈值的速度远快于历史记录。他询问曾领导 YC 并作为顶尖投资人的 Sam：在这个时代，优秀创始人所具备的特质是否发生了根本变化？ Sam 直言当前正在发生“点子王的大复仇”（the revenge of the idea guys）。过去在硅谷和 YC，大家常嘲笑那些宣称“我有点子，只缺一个程序员把它写出来”的人，这就像一个人声称“我有一首绝妙的歌，只缺一个吉他手把它弹出来”一样显得荒谬，非技术团队在过去极难取得成功。然而现在，AI 编程工具的大幅普及彻底逆转了这一格局。 - 过去行业最看重的硬指标是创始团队的技术实力（虽然现在依然重要）。 - 如今，即使创始人完全不会写代码，只要对用户需求有着极其深刻的理解和洞察，也是 Sam 非常渴望投资的对象。 --- ### [31:53](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1913s) AGI 逼近下的投资视野与企业存续周期 Patrick 追问在当下的时代该如何思考初创企业投资：一方面有人预期通用人工智能（AGI）、超级智能（ASI）甚至技术奇点可能在几年内到来；另一方面风险投资基金通常有着 10 年期的生命周期。Sam 认为，在当前环境下制定 10 年规划确实需要“主动暂停怀疑”（suspension of disbelief），但这也是唯一正确的生活与工作方式。人不能因为三五年后可能迎来无法看透的奇点就选择放弃或无所作为，而必须假定世界仍将以可以理解的方式长期运转下去。 谈到 OpenAI 自身的规划跨度时，Sam 解释了其产品与基建层面的不同时间尺度： - 在物理基础设施方面，OpenAI 会签署长达 20 年的电力和土地协议。 - 在产品与技术演进方面，团队对未来 2 年有着清晰愿景，但超过 2 年之后的一切都会变得非常模糊。 关于市场上从早期贬低“GPT 包装层（GPT Wrappers）脆弱无护城河”，转变为推崇“模型脚手架/框架（Harnesses）具备厚重价值”的舆论反转，Sam 表明自己始终持有相同的评判标准：一家企业的底层逻辑必须站在“期望 AI 变得越来越聪明”的那一边，只有能够随着底层模型能力增强而受益、而非被模型进步所吞噬的业务，才具备真正的持久力。 ### [34:14](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2054s) 模型底座与外围封装：顺应智能演进还是填补缺陷 萨姆·奥特曼（Sam Altman）谈到创业者在面对大模型演进时的一种微妙心态：如果你创业的核心业务只是在给当前模型的某个软肋“打补丁”，而这个弱点注定会在下一代模型中被攻克，那么每次新模型发布变得更强大时，你反而会感到沮丧；相反，如果你构建的是能从底层智能提升中直接受益的优秀服务，那么模型的每一次升级都会让你倍感兴奋。在奥特曼看来，看待这一技术范式的正确方式，是将数据中心、基座模型与外围工具链（Harness）视作一个向外持续输出高可用智能的统一集群。在这套技术栈中，有太多的应用场景值得去探索和构建，创业者理应为底座的持续进化感到欣喜；如果有人暗自祈祷模型不要进步，往往只是因为他们在为当下的缺陷缝缝补补，而技术栈“下一次转动手柄”时，这些问题就会被轻而易举地解决掉。 ```mermaid flowchart LR accTitle: 面对模型升级的两种构建路径 accDescr: 对比修补当前模型缺陷与顺应智能进化的两种不同产品构建策略及其结果 subgraph Stack["统一智能集群"] DC["数据中心"] --> M["前沿模型"] M --> H["外围工具链 (Harness)"] end Stack -->|"能力跃升"| PathA["模式 A：构建纯补丁应用<br/>（试图掩盖当前缺陷）"] Stack -->|"能力跃升"| PathB["模式 B：构建智能杠杆服务<br/>（直接吸收底层红利）"] PathA --> ResA["模型迭代即失效<br/>（感到焦虑）"] PathB --> ResB["业务价值同步倍增<br/>（感到兴奋）"] ``` ### [35:04](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2104s) 高效应用 AI 的企业特质：CEO 全身心投入与近乎极限的数据授权 帕特里克·科里森（Patrick Collison）向奥特曼提问：在 OpenAI 接触过的无数大大小小的客户中，那些将 AI 运用得最为出色的顶尖组织究竟有何不同？奥特曼指出了两个极为鲜明的方向。首先是最高层自上而下的推动，他特别提到了两人共同的好友——Shopify 首席执行官托比·吕特克（Tobi Lütke）。吕特克是奥特曼见过的第一位直接宣布公司要“全面押注 AI”并亲自下场动手的 CEO。他不仅自己卷起袖子用 AI 构建各种内部自动化流程，还要求管理团队全员跟进，绝不搞花哨的排行榜或形式主义的积分游戏，而是实打实地强调“我们要用 AI 把所有糟糕的流程变好”，不落实就会面临严厉的问责。受到这种领导力“分形效应”的启发，OpenAI 正准备开展一项新试验：派遣前线部署工程师（Forward Deployed Engineer）直接深入客户高层，贴身帮助 CEO 自动化其日常工作，进而自上而下辐射整家企业。 另一个决定性的特质，是企业对数据访问权限展现出“令人不安的极度开放”。奥特曼特别澄清这并非公开倡导，而是他在最极致团队身上观察到的客观事实：这类公司（多为两三人的早期初创团队）选择打破常规合规与流程顾虑，允许 AI 完整接入所有会议录音、核心代码库、Slack 聊天记录与往来邮件。当每个员工都能依托这种全知视角与 AI 协作时，小团队所爆发出的生产力令人惊叹。尽管这在大企业严苛的隐私法规与敏感数据合规面前难以照搬，但它所释放的效率正对现有规则提出深刻挑战。 - **高管表率的分形效应**：CEO 亲自使用 AI 改造自身工作流并设定刚性标准，能以极快速度将自动化文化渗透至组织的最末梢。 - **数据开放度与生产力权衡**：打破会议、邮件和沟通工具的数据孤岛，是当前极少数初创团队拉开人效差距的关键驱动力。 ### [38:09](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2289s) 从 Slack 单频道全自动运转到人机界面的抽象鸿沟 科里森顺着这种工作形态分享了一个生动案例：由 Stripe 与 Paradigm 孵化的高性能区块链项目 Tempo。该团队仅有二三十人，他们在 Slack 中搭建了一套由大模型驱动的自动化编排工具，几乎承接了公司的全部日常运转。团队成员只需在 Slack 中抛出一句指令，Agent 便会自行调取并解析 Google Docs 文档，将其拆解为 Linear 任务列表，紧接着编写 Pull Request 提交代码、执行自动化部署，最后调用日志分析工具验证上线结果。科里森感叹，亲眼看着一个组织在单个 Slack 频道中完成全部端到端工作“极其震撼”，但他坦言完全不知道该如何将这种范式平移到 Stripe 这样体量庞大的公司。 奥特曼对此深表赞同，并指出业内目前存在巨大的认知悬置（Overhang）：绝大多数人至今无法理解“你只要对 AI 提出要求，事情大概率就会被办妥”这一现实，甚至他自己有时也仍保留着某种惯性的不信任感。对于大企业而言，核心瓶颈在于缺乏一层关键的抽象架构——目前尚无人摸索出在超大规模组织中，成千上万的人类与 AI 应当如何顺畅交互；相比之下，微型团队的全部优势恰恰在于它们大多由纯粹的 AI 自主协同构成，根本无需耗费精力去解决复杂的人际协作界面。 ```mermaid flowchart TD accTitle: Tempo 团队 Slack 全自主研发与运维流程 accDescr: 展示单个 Slack 指令如何触发 AI Agent 跨工具协同完成从需求到上线验证的全流程 A["Slack 频道任务指令"] --> B["Agent 读取并解析 Google Docs"] B --> C["自动拆解生成 Linear 任务"] C --> D["编写代码并提交 GitHub PR"] D --> E["自动执行生产环境部署"] E --> F["调用日志分析工具验证运行状态"] ``` ### [40:05](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2405s) 开源 AI 的未来与大模型驱动科学发现的拐点 在谈及开源 AI 是否有未来时，奥特曼给出了明确肯定的回答。他分析指出，虽然眼下主流市场的核心诉求高度集中于更聪明、更迅捷、成本更低的前沿闭源智能，但开源生态的实际需求同样极其庞大，并且随着技术发展，这种需求在相对比例上还会进一步放大。 随后，对话转向了奥特曼最为振奋的议题——AI 与基础科学研究的结合。科里森提到 OpenAI 作为一家非营利基金会，近期向由他联合创立的生物医学研究机构 Arc Institute（弧光研究所）提供了一笔资金资助。奥特曼直言，加速科学发现将是 AI 对人类生活质量带来的“最重要贡献”。他坚信人类福祉的提升本质上遵循一条第一性原理：加深对科学规律的理解，据此制造出工具与产品，并分发给全人类。从近几个月的模型演进，特别是达到当前 5.5 水平后，顶尖科学家们已开始反馈模型不仅能激发更优质的假说，甚至能自主完成一些虽小但至关重要的科学发现。未来随着自动化实验室与机器人的加入，“将过去需要耗费十年才能取得的科研成果压缩至一年内完成”的复利效应将彻底重塑文明进程。 - **开源与前沿双轨并行**：尽管前沿能力仍在定义行业天花板，开源生态在长尾场景与自主可控方面的需求将持续上升。 - **科研周期的数量级压缩**：大模型结合实验自动化，核心价值在于大幅缩短基础科学从假说构想到实验验证的迭代周期。 ### [42:48](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2568s) OpenAI 基金会的使命与 Arc Institute 的前沿攻坚 奥特曼强调，加速科学研发将是未来 OpenAI 基金会的核心战略重心之一，该机构将注入充沛的资金、工程专家与算力技术，甚至有望成长为“全球规模最大的公益基金会”。除基础科研之外，基金会的另一大重点在于提升社会的“AI 韧性（AI Resilience）”，帮助全人类平稳度过这场技术大变革。谈及对 Arc Institute 的资助，奥特曼赞誉其代表了当前“AI + 生物学”探索的最高水准，原本他预计生物医药的突破需要更长时间，但 Arc 的卓越进展让他深感治愈疾病、改善人类健康的里程碑并不遥远。 科里森风趣地将接下来的环节比作播客中途插入的“赞助商广告时间”，向在场听众解释为何会场楼下会设立一个 Arc Institute 的展台。他介绍道，Arc Institute 是一个成立于四年前的研究机构，其终极使命是针对人类的“复杂疾病（Complex Diseases）”研发出首批根治性疗法。所谓复杂疾病，是指由遗传背景与环境因素交织共同诱发的疾病形态，绝大多数癌症、自身免疫性疾病以及神经退行性疾病均属于此类范畴。借助计算智能与前沿生物技术的交汇，跨学科攻克这些顽疾正迎来历史性的突破窗口。 ### [44:32](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2672s) Arc Institute 的医学使命与现场“插播广告” Patrick Collison 谈及创立 Arc Institute 的初心时指出，纵观人类历史，人类其实“从未治愈过任何一种复杂疾病（complex disease），一种都没有”。人类能够治愈许多传染病，也能对由单一基因突变引起的单基因遗传病进行筛查与干预，但面对多因素交织的复杂病症始终束手无策。Arc Institute 正是带着攻克这一难题的目标应运而生，并将阿尔茨海默病（Alzheimer's）作为首个主攻靶标。研究所成立仅四年，结合 CRISPR 基因工程技术与人工智能的最新飞跃，早期研究已经展现出令人振奋的势头。 话锋一转，Patrick 在台上大方做起招聘宣传：Arc 正在物色一位全职 CTO。去年，OpenAI 联合创始人兼总裁 Greg Brockman 曾在 Arc 度过了一段学术休假（sabbatical），并协助团队训练出了迄今为止规模最大的生物基座模型 Evo 2。Patrick 笑着对台下观众打趣，这位未来的 CTO 可能就在观众席中，或者在大家的朋友圈里，欢迎大家移步楼下的 Arc 展位咨询。面对这段突如其来的现场宣传，Sam Altman 调侃道：“我觉得你直接坦白这是一段‘插播广告’（interstitial ad），可比暗戳戳地植入好多了。” - Arc Institute 聚焦于复杂疾病机制攻坚，首要突破方向为阿尔茨海默病。 - 核心技术结合了 CRISPR 等新一代基因工程技术与前沿 AI 算法。 - Greg Brockman 曾以休假学者身份在此协助训练了生物学基座模型 Evo 2。 ### [46:16](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2776s) 帕洛阿尔托厨房里的早期投资：为什么看好“满脸青春痘的少年” Patrick 将话题拉回 Stripe 自身，提到 Sam 是当年 Stripe 的第二位投资人（或者说与 Paul Graham 并列第一）。那是发生在 Paul Graham 帕洛阿尔托厨房里的一幕，两张投资支票几乎在同一时刻递出。Patrick 回忆道，当时自己和弟弟 John 还只是两个“满脸青春痘的少年”（pimply teenagers），大言不惭地说要建立一个现代金融服务机构，听上去极其荒唐可笑，他好奇 Sam 当时究竟看中了什么。 Sam 坦言，这并非事先准备好的客套吹捧，而是他当时见过无数创业者后得出的直观判断：“你们兄弟俩是我见过所有年龄段中最令人印象深刻的创始人之一，更不用说是十几岁的青涩少年了。”Sam 补充说，当时不仅在网络上和 Paul Graham 口中了解过他们，更被打动的是他们是在为自己解决真实的痛点——Patrick 和 John 当时开发了一款离线下载维基百科全部内容的 iPhone 应用，结果在收取费用时被繁复的支付接入折磨得苦不堪言。Sam 的投资哲学十分纯粹：找到极其聪明的顶尖创业者，再匹配一个必将爆发的时代趋势（商业全面线上化与创业潮爆发），那就毫不犹豫直接开出支票。 - 投资场景发生在 Paul Graham 位于帕洛阿尔托的家中厨房，支票几乎与 Y Combinator 同时交付。 - Stripe 的原始产品需求脱胎于创始人此前开发离线维基百科 iPhone 应用时所遭遇的支付接入困境。 - 投资逻辑底色：顶尖且解决切身痛点的创始人 + 商业线上化大趋势。 ### [48:02](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2882s) 给 Stripe 的航向建议：警惕“颠覆一切”的妄想症 面对 Patrick 询问“作为核心客户，OpenAI 对 Stripe 在 AI 时代的产品功能与战略发展有何建议与吐槽”，Sam 反客为主，直言自己现在反倒想更多地向 Stripe 的商业模式取经。Sam 回忆，当年许多自以为精明的投资人曾放言 Stripe 会沦为管道类的大路货（commodity），认为企业一旦做大就会自建支付；但现实证明，只要把产品做到极致，成为生态中靠谱、理性的长期伙伴，客户就会坚定不移地选择留下。 谈及各行各业在 AI 浪潮下的恐慌与躁动，Sam 提醒大家不要陷入极端妄想。在他看来，企业借助 AI 提质增效是必然趋势，但那种“所有现有公司都将化为乌有、社会经济系统将被彻底重写”的论调完全是“陷入了轻度妄想”（a little bit delusional）。哪怕未来支付场景演进为智能体（Agents）之间相互结账，资金流转与价值结算的底层需求依旧稳固。Sam 给 Stripe 的建议非常清晰：在内部全力采用 AI 以打造更强大的产品，但千万不要预设既有的社会经济秩序会在一夜之间全盘颠覆。 ### [50:24](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3024s) 视野越过 AI：从数据中心底座、防御性生物科技到全新人机协议 展望未来十年的技术格局，Patrick 询问除 AI 之外还有哪些领域最让 Sam 兴奋。Sam 坦言，自己目前最为痴迷的是支撑 AI 运行的底层物理世界——从数据中心基础设施、前沿能源技术到实体机器人系统，构成了他日常思考最多的技术栈。此外，脑机接口（BMI）终于迎来突破性进展；而在生物科技领域，由于合成生物学门槛降低带来的潜在生物安全威胁，他认为“防御性生物技术”（defensive biotech）的建设已经变得极其迫切。 Sam 还特别抨击了当下极其扭曲的交互界面体验。他指出，我们手握具有划时代意义的智能技术，却依然被禁锢在几十年历史的旧设备与传统操作系统中。让强大的大模型像真人一样在屏幕上笨拙地“点击按钮、操作窗口”，实际上是非常低效且割裂的妥协方案。这套专为人类感官设计的 GUI 并不适合 AI 运行，行业亟需重新构想人机交互逻辑，甚至制定一套全新的原生互联网通信协议。 ```mermaid flowchart TD accTitle: Sam Altman 关注的前沿技术图谱 accDescr: 展示 Sam Altman 在模型层之外关注的物理底座、应用前沿与交互革新三大板块 subgraph PhysicalLayer ["底层物理支撑与基建"] A["数据中心物理架构"] B["清洁与聚变能源"] C["具身实体机器人"] end subgraph FrontierBio ["生命与人机前沿"] D["脑机接口 (BMI)"] E["防御性生物科技 (Defensive Biotech)"] end subgraph InteractionParadigm ["交互与协议革命"] F["摆脱传统 GUI 模拟操作"] G["原生 AI 交互系统与全新互联网协议"] end PhysicalLayer --> FrontierBio FrontierBio --> InteractionParadigm ``` - 核心技术栈聚焦：数据中心物理层创新、大规模清洁能源供应以及具身机器人。 - 生物科技态度：兼具期待与警惕，呼吁加速部署应对未知风险的防御性生物技术。 - 人机交互痛点：批评现有大模型调用计算机时模拟人工点击界面的倒退现象，呼吁建立 AI 原生通信协议与全新操作系统。 ### [52:09](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3129s) 闪电快问：核聚变商用、高超音速飞行与被忽视的材料科学 在闪电问答环节，Patrick 抛出了一连串硬科技时间表与冷门领域问题。当被问及“全球首座可实现商业盈利的核聚变反应堆何时出现”时，Sam 敏锐地指出这取决于数据中心庞大用电需求将电价推高到何种程度，随后他给出了一个极为大胆的预言：“我猜就在未来 5 年之内。”而谈及 4 马赫以上的高超音速商业客运飞行，他坦承自己并非专家，保守估计至少需要 10 年以上。 随后，Patrick 请他指出一个当前未被舆论充分重视、但极具社会冲击力且会被 AI 深度加速的科学领域。Sam 毫不犹豫地选择了“材料科学”（Materials Science）。Sam 表示，材料学听起来缺乏科技光环、并不酷炫，但公众严重低估了人类现实世界对先进材料与催化剂的依赖程度。材料研发从本质上说是一个“与 AI 能力极其契合的问题”（a beautifully AI-shaped problem），随着 AI 在分子模拟与合成路径上的突破，该领域将迎来爆发式进展并深刻重塑人类生产生活。 - 商业化核聚变预测：在算力能耗推高电力价值的催化下，有望在未来 5 年内实现盈利。 - 高超音速商业航行预测：由于工程复杂性与监管门槛，预计仍需 10 年以上。 - 被低估的变革领域：材料科学（如高效工业催化剂研发），属于天然适合 AI 发挥范式突破的学科。 ### [53:51](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3231s) 拒绝象牙塔垄断：“迭代式部署”与技术民主化的分水岭 在对谈的尾声，Patrick 提出了一个具有历史决定论色彩的宏大问题：鉴于全球资本与算力正以前所未有的确定性涌向 AI，某种形式的超级智能似乎不可避免，那么 Sam 个人的躬身入局，相对于其他假想历史（counterfactual），究竟希望给世界带来怎样的根本不同？ Sam 给出的答案是深植于骨子里的“技术民主化”（democratization）与“个人能动性”（personal agency）。他回顾道，OpenAI 历史上最具争议的抉择正是其开创的“迭代式部署”（iterative deployment）策略。在当年决定公开发布 ChatGPT 时，安全圈内许多主流声音强烈反对，认为这极其危险、属于不可逆的“信息危害”（infohazard），主张将大模型锁在象牙塔里，仅由少数安全研究者代为掌管和分配研究果实。 这种傲慢与权力高度集中的设想让 Sam 感到极其不安。他坚定地认为，与其把技术垄断在极少数精英手中，不如将其尽早交给全人类去探索与测试。尽管开放的过程注定伴随着混乱与未知，也必须设置合理的安全护栏，但唯有让全世界普通人尽早拥有支配智能工具的能动性，才能真正防止权力的极致垄断，探索最广阔的未来可能。 ### [55:34](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3334s) 技术的民主化：推动 AI 走向大众与对谈结语 Sam Altman 总结了他对技术普惠与赋能大众的坚定信念。他指出，只要把强大的工具赋予世界，全世界的人们就会在它的基础之上，为全人类创造出一份更为巨大的馈赠。回顾过往，Altman 提到在 OpenAI 涉足并改变行业格局之前，当时从事前沿 AI 研究的人群中曾存在一种心照不宣的大致共识——那就是将这项技术严密地“封锁起来”（keep it locked up）。尽管在今天看来，把技术开放给公众使用似乎是理所当然的事情，但在当时并非如此。 Altman 直言，如果当时真的选择将技术深锁高阁，后果将会极其糟糕。他表示自己始终坚信创业精神（entrepreneurship）与技术创新的力量，坚信人性大体是善良的（"people are mostly good"），而且绝大多数人在获得优秀工具后，都会创造出令人惊叹的成果。回顾自己的心路历程，Altman 认为自己迄今乃至未来“最大的单一贡献”（single biggest contribution），正是坚定不移地推动 AI 成为一项民主化普及的技术（democratized technology），确保普通人都能亲手使用它并在其之上构建未来。 对谈在此刻推向高潮，Patrick Collison 向 Sam Altman 致以真挚的谢意，全场响起热烈的掌声，两人的炉边对话在对技术开放与人类创造力的展望中圆满收尾。 - **行业初期的封闭共识**：在 OpenAI 改变范式前，主流研究圈的倾向是将高阶技术严加封锁，而 Altman 坚信封闭路线会带来糟糕的结局。 - **对人性和创新的信任**：Altman 强调人性本善，技术扩散能够释放大众的创业精神，促使人们利用工具做出卓越贡献。 - **自身的核心贡献**：Altman 将推动 AI 技术的民主化、让所有人都能接触并基于此进行二次开发，视为自己最重要的一项贡献。 ## Notable quotes > [00:01](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1s) 一旦人们确认某件事在物理上是可行的，后续的推进就会爆发出难以想象的干劲。 — Sam Altman > [05:53](https://www.youtube.com/watch?v=5eouRdDYM2c&t=353s) 在互联网上给自己挑一件 20 美元以内的礼物，任意东西都行。 — Patrick Collison > [10:29](https://www.youtube.com/watch?v=5eouRdDYM2c&t=629s) 我们是不是陷入了某种集体精神病（collective psychosis）？我们是不是只是在互相煽动、把彼此逼进了某种盲目的狂热之中？ — Sam Altman > [18:19](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1099s) 祝那帮人好运，我觉得他自己其实也没当真。 — Sam Altman > [29:33](https://www.youtube.com/watch?v=5eouRdDYM2c&t=1773s) 点子王的大复仇（the revenge of the idea guys）。 — Sam Altman > [35:04](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2104s) 我们要用 AI 把所有糟糕的流程变好。 — Tobi Lütke > [44:32](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2672s) 从未治愈过任何一种复杂疾病（complex disease），一种都没有。 — Patrick Collison > [44:32](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2672s) 我觉得你直接坦白这是一段‘插播广告’（interstitial ad），可比暗戳戳地植入好多了。 — Sam Altman > [46:16](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2776s) 你们兄弟俩是我见过所有年龄段中最令人印象深刻的创始人之一，更不用说是十几岁的青涩少年了。 — Sam Altman > [48:02](https://www.youtube.com/watch?v=5eouRdDYM2c&t=2882s) 所有现有公司都将化为乌有、社会经济系统将被彻底重写（完全是）陷入了轻度妄想（a little bit delusional）。 — Sam Altman > [52:09](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3129s) 材料研发从本质上说是一个‘与 AI 能力极其契合的问题’（a beautifully AI-shaped problem）。 — Sam Altman > [55:34](https://www.youtube.com/watch?v=5eouRdDYM2c&t=3334s) 坚信人性大体是善良的（people are mostly good），推动 AI 成为一项民主化普及的技术（democratized technology）是自己迄今乃至未来最大的单一贡献（single biggest contribution）。 — Sam Altman ## Mentioned Sam Altman（萨姆·奥特曼）, Patrick Collison（帕特里克·科里森）, John Collison（约翰·科里森）, Greg Brockman（格雷格·布罗克曼）, Paul Graham（保罗·格雷厄姆）, Tobi Lütke（托比·吕特克）, Elon Musk（埃隆·马斯克）, Jensen Huang（黄仁勋）, OpenAI, Stripe, DeepMind, Y Combinator, Shopify, Arc Institute（弧光研究所）, Paradigm, Gumroad, Codex, ChatGPT, GPT-3, GPT-3.5, GPT-4, GPT-5.2, GPT-5.5, GPT-5.6, OpenClaw, Stripe Link CLI, Evo 2, Tempo, Slack, Linear, GitHub, CRISPR ## Video description OpenAI cofounder and CEO Sam Altman joins Stripe CEO Patrick Collison for a fireside chat at Stripe Sessions. More information: https://stripesessions.com Register for Stripe Sessions 2027: https://register.stripesessions.com/2027?promocode=s27youtube Opening keynote: https://youtu.be/lIsHZfRl2zw Developer keynote: https://youtube.com/live/-vRY2dtD7iQ Indexing the economy keynote: https://youtu.be/5wGqWRv1Z1s Sam Altman fireside: https://youtu.be/5eouRdDYM2c Nat Friedman and Daniel Gross fireside: https://youtu.be/l9wzs_QIyp0 [Source](https://www.youtube.com/watch?v=5eouRdDYM2c)

[0:01] Sam Altman: 一旦人们确认某件事在物理上是可行的，后续的推进就会爆发出难以想象的干劲。 [5:53] Patrick Collison: 在互联网上给自己挑一件 20 美元以内的礼物，任意东西都行。 [10:29] Sam Altman: 我们是不是陷入了某种集体精神病（collective psychosis）？我们是不是只是在互相煽动、把彼此逼进了某种盲目的狂热之中？ [18:19] Sam Altman: 祝那帮人好运，我觉得他自己其实也没当真。 [29:33] Sam Altman: 点子王的大复仇（the revenge of the idea guys）。 [35:04] Tobi Lütke: 我们要用 AI 把所有糟糕的流程变好。 [44:32] Patrick Collison: 从未治愈过任何一种复杂疾病（complex disease），一种都没有。 [44:32] Sam Altman: 我觉得你直接坦白这是一段‘插播广告’（interstitial ad），可比暗戳戳地植入好多了。 [46:16] Sam Altman: 你们兄弟俩是我见过所有年龄段中最令人印象深刻的创始人之一，更不用说是十几岁的青涩少年了。 [48:02] Sam Altman: 所有现有公司都将化为乌有、社会经济系统将被彻底重写（完全是）陷入了轻度妄想（a little bit delusional）。 [52:09] Sam Altman: 材料研发从本质上说是一个‘与 AI 能力极其契合的问题’（a beautifully AI-shaped problem）。 [55:34] Sam Altman: 坚信人性大体是善良的（people are mostly good），推动 AI 成为一项民主化普及的技术（democratized technology）是自己迄今乃至未来最大的单一贡献（single biggest contribution）。

[0:01] Sam Altman: Once you know something's possible, it's much easier to go do it with vigor. [3:32] Sam Altman: The degree to which most people will realize they can like sit back and watch an AI do most of their like drudgery is going to surprise people. [10:05] Sam Altman: Are we engaging in like collective psychosis? Like, have we just totally whipped each other into this frenzy? [12:15] Sam Altman: I think it's not going to be a natural fit for my management style. [18:03] Sam Altman: we're not going to build a Dyson sphere and then just like cover it with data centers, but maybe we do. [21:05] Sam Altman: Economists are the people who have called eight of the last three recessions. [27:07] Sam Altman: All of a sudden, it's like the revenge of the idea guys, which is actually awesome for the world... [31:53] Sam Altman: You have to live as if stuff's just going to keep going in an understandable way for a long time. [50:24] Sam Altman: defensive biotech is about to become very important unfortunately. [55:37] Sam Altman: people are mostly good and mostly do amazing things with tools.

[0:01] I think we have some Codex fans in the [0:03] audience. [0:04] >> Love to hear that. [0:06] >> Um, [0:08] how's the week going? [0:09] >> So fun. Um, it's a busy week, but I'm [0:13] happy to be here. This is a unexpected [0:15] surprise. [0:15] >> Well, thank you for joining us. We [0:16] appreciate it. Um so we opened uh this [0:20] morning by saying that it is uh we've [0:22] kind of arbitrarily decided uh that the [0:24] singularity started on January 1st and [0:27] thus today is day 119. Uh what do you [0:31] think of that? [0:35] It does feel like we are somehow in the [0:37] takeoff. Day 119 feels like a reasonable [0:40] enough guess. Uh yeah, I won't fight it. [0:45] Did you did you feel like did did you [0:47] feel this? So we've started to see a [0:49] bunch of our metrics inflect as of um I [0:53] mean late last year beginning of this [0:55] year I mean things were kind of doing [0:57] well but somehow they really just the [1:00] the shapes of the curves changed. They [1:02] really went parabolic. Is that matched [1:04] in what you guys see? Was there some [1:06] trajectory change? Like why are we [1:08] seeing this? [1:11] I do think the models got really good [1:13] especially for coding but really good in [1:15] general starting [1:18] late last year very early this year and [1:22] at least in my own experience of using [1:26] this technology and seeing what other [1:27] people are doing with it. It and also [1:30] this sense that every week is now a [1:33] little bit different than the week [1:34] before like a lot happens very fast. um [1:39] it seemed to all correlate with the [1:41] models hitting some threshold. [1:43] >> And why did they like why did coding [1:45] models suddenly start to click over the [1:48] last you know couple months? Like was [1:49] there a was there a research trick? Was [1:51] it just got enough code data in the [1:53] pre-trip? Like what why why did why did [1:55] it suddenly start to work? [1:56] >> Yeah, it's a great question. We we [1:58] wondered a lot about like why kind of [2:00] several people cross that threshold at [2:02] the same time. Um, [2:06] I I I'm sure it's a number of factors, [2:08] but model intelligence, just the raw [2:10] kind of reasoning horsepower, um, enough [2:13] of a feedback loop of people using it [2:15] for code to figure out where it was good [2:17] and where you needed to improve it. Uh, [2:19] enough data. I think it was like all of [2:21] these things. And then also there was [2:23] like a like many other endeavors, once [2:26] you know something's possible, [2:29] it's much easier to go do it with vigor. [2:32] And so it seems like codeex is kind of [2:34] having a moment right now. [2:37] >> Yeah, it really it crossed some [2:40] subjective threshold for me with the [2:43] latest app updates and 5.5. And this is [2:46] also like it's quite hard to say why [2:48] right now and not a little bit sooner or [2:50] why not the next model. But one of the [2:53] things I have learned about the history [2:55] of all of the things we put out is [2:58] it is very hard to say why this [3:02] particular thing was the thing that [3:03] worked. And this goes back to chatg like [3:05] why was GPT 3.5 the thing that got over [3:08] the threshold where most people went [3:10] from saying not that impressive to like [3:11] going to change the world and why not [3:13] one model earlier or later. I really [3:15] can't explain it. um you just kind of [3:18] feel it and the I've had two inflection [3:21] points with codecs. One was kind of with [3:24] GPT 5.2 [3:26] and then a really big one in the last [3:28] few weeks where it's like okay this is [3:30] going to be the primary interface to a [3:31] computer for me. [3:32] >> And is everyone using it for coding or [3:35] are you starting to see usage diffuse [3:37] into other domains? [3:41] >> I think the most adamant users are still [3:44] using it for coding. Um, but there [3:46] there's been just this like tidal wave [3:48] of people coming into Codeex recently [3:50] and one of and I'm like really trying to [3:52] understand like what has happened that [3:55] this is [3:57] causing this and the the depth of what [4:01] people are using it for or starting to [4:02] use it for has surprised me. So [4:04] certainly our ambition is for it not [4:06] just to be about coding but to be about [4:09] all the work you do in front of a [4:10] computer. Uh, and [4:14] you know, I would say we're maybe like [4:16] 10% of the way there for the non-coding [4:18] stuff, but now that we see what's [4:19] happening, now we have a real user base [4:21] sort of using it in in these other ways, [4:23] I think we'll get good at it very fast. [4:25] >> What what do you think will be the next [4:27] domain that subjectively feels like it [4:29] has this big unlock after coding? Like [4:31] it's going to be spreadsheets, will it [4:34] be uh, you know, performance reviews? [4:36] What's what's it going to be? So I think [4:38] there will be a lot of first of all I do [4:40] think coding is a little bit special and [4:43] these models are a great fit for coding. [4:45] The world needs so much more code than [4:47] currently gets written. Um there may be [4:49] no other domain that is quite like [4:51] coding. Uh but there will be a lot of [4:52] others that I think are close. But the [4:55] the next kind of coding like thing that [4:57] I think will happen is not any specific [5:00] domain but the realization of how much [5:02] time people waste trying to use a [5:04] computer and the idea that you can do a [5:08] huge percentage of your day in a very [5:11] different way. And you know, maybe maybe [5:14] you don't realize like how much time you [5:16] spend clicking between messaging apps [5:19] and copying and pasting stuff and like, [5:21] you know, responding to very boring [5:22] things that you could clearly automate [5:23] once. Um, but the degree to which most [5:26] people will realize they can like sit [5:27] back and watch an AI do most of their [5:31] like drudgery [5:32] um is going to surprise people. And [5:37] in my own experience trying to work that [5:39] way actually gives me much more like [5:41] enjoyment of work. I didn't realize how [5:43] much like the little stuff drags me [5:45] down, gets me like out of a sort of [5:46] happy flow state or whatever. Um, so the [5:50] subjective like quality of life [5:51] improvement is huge. [5:53] >> Are you an OpenClaw user? [5:55] >> I am. Um, I [5:57] >> you open users here. [6:00] [cheering and applause] [6:00] >> We have some uh good news for you all [6:02] coming. Um I [6:04] >> you can just tell them. [6:08] >> Um OpenClaw has been one of my biggest [6:12] like this is magic AGI moments ever in [6:16] the field. Uh I I remember the first [6:19] time someone told me about it. They were [6:20] trying to explain it and I was like okay [6:22] like those all that all sounds cool but [6:24] you know I can like make a lot of that [6:25] work. And then it was a real reminder [6:27] how when when the models cross some [6:29] threshold and also the product designer [6:31] gets a handful of critical ideas really [6:34] right, it's like a much more magical [6:36] experience than it than it sounds like. [6:38] I find um I I've been sort of a um an [6:44] attempting openclaw evangelist uh and [6:47] you know trying to describe it's um you [6:49] know that that experience you just [6:51] recounted uh to others and I find it a [6:53] difficult experience to communicate. I [6:56] mean it sounds kind of prosaic right [6:57] it's a it's a a stateful chat GPT [7:00] session that can also make some use of [7:02] tools and so forth. Um, how like what do [7:04] you use your open cloud for? [7:07] >> If we scrolled your message thread, what [7:09] do we see? So, this is like a very [7:13] embarrassing thing to admit. The thing I [7:15] always try first, [7:16] >> perfect place. [laughter] [7:18] Um, the thing I always try first with a [7:20] new like a new kind of AI system of any [7:22] sort is to I'm like a home automation [7:25] nerd and so to like try to build a [7:27] better home automation interface system [7:28] because it's all it never works. Like it [7:30] never is good. Um, and OpenClaw was the [7:34] first time that I was able to get like a [7:35] setup that I was happy with. Um, [7:39] I also built a messaging app that I had [7:41] always wanted to work. I've since [7:43] switched it to something I built with [7:44] codecs, but OpenCloud was the first time [7:46] I was able to like I'm sure like you it [7:48] like feel like drowning in messages and [7:50] it's like this very unpleasant task to [7:52] wake up in the morning have to go [7:53] through all this stuff. So, I was like, [7:54] "All right, I'm finally going to be able [7:55] to automate this." And that was like [7:59] again should have been doable with [8:01] previous systems. Hard to explain like [8:04] what it's like when it all actually just [8:05] works and you trust that it's going to [8:07] work. Um I was uh testing the uh the new [8:10] link uh uh CLI that we uh that we just [8:13] launched today uh in preparation for [8:15] launch and I asked my agent to uh you [8:18] know it it means you can uh you can [8:20] easily get a singleuse card they could [8:23] use on any business. Um, and so, uh, I [8:25] asked my agent to go and buy itself a [8:28] gift, just anything on the internet, uh, [8:30] for under $20. Um, and it, uh, it it [8:32] chose to buy itself, uh, an HTTP design [8:36] from, uh, from Gumroad. [8:38] Wow. [8:40] Yeah. There's all of the there's all [8:42] this stuff that like feels no matter in [8:45] no matter how much how convinced you are [8:48] intellectually that this is like not a [8:49] real thing wanting a real gift for [8:51] itself. And no matter how much you're [8:52] convinced like, okay, this is like a [8:54] weird emergent behavior and I'm not [8:56] supposed to read into this. Um there are [8:59] these things that uh [9:04] feel a little strange. I [9:07] um we're going to have a party for GPT [9:09] 5.5 and I wasn't quite sure like we're [9:12] going to, you know, invite people who [9:13] are big users and whatever and I wasn't [9:14] quite sure what to do. And so on like [9:16] kind of whim, I was like, um, I'm going [9:18] to ask 5.5 like xx high what it would [9:21] like for a party for itself [9:24] this morning. Um, and [9:28] I did and it was this sort of like [9:30] beautiful set of things including like, [9:35] you know, I like here's what I would [9:37] want for like the flow of the party. [9:39] Here's what I would not want. You know, [9:40] you should do it on May 5th. That would [9:42] be funny. Um, I would like, you know, [9:44] like only a short little toast and I [9:46] want it not to be by me but by the [9:47] people that built it. I would like like [9:49] a s, you know, big central suggestion [9:51] for 5.6 and I would like you to like [9:53] feed them all into me and I'll like, you [9:55] know, make sure we work on that. And it [9:57] >> and now there's real moral pressure on [9:58] you to [9:58] >> Well, we're going to do it. Um, but it [10:02] was a strange thing. [10:05] >> So, I want to ask you about OpenAI [10:06] itself. um you know a lot of crazy open [10:10] eyes I mean it's it's now [10:12] >> I am aware [laughter] [10:14] >> it's an 11year-old organization now [10:17] >> somehow feels like so much longer than [10:19] that but yes I I get just over 10 I [10:21] guess yeah okay 10 years um a long 10 [10:23] years um I can't like remember [10:26] pre-openai life that well at this point [10:28] it feels like it's been so long but yes [10:29] you you have a singularity looking [10:31] forwards but also a singularity looking [10:32] backwards um so what's Um, what's the [10:37] craziest OpenAI story that's never been [10:39] told? [10:48] I mean it sound it sounds so prosaic [10:51] relative to like the crazy drama that's [10:54] happened but [10:58] there was this period [11:01] after we had finished training oh there [11:06] I guess there's two that are kind of [11:07] similar but the one that just came to [11:08] mind was after we had finished training [11:10] GPT4 there were about eight months [11:12] before we released [11:14] And so there was this 8-month period [11:15] inside of OpenAI where we were all using [11:18] this thing. We kind of knew that it was [11:23] dramatically better and different and [11:25] going to unlock a bunch of things in the [11:27] world. Um, [11:31] and no one outside the company or almost [11:33] no one outside the company knew about [11:34] it. And and there was this we like [11:38] walked the hall sometimes. like, you [11:39] know, are we engaging in like collective [11:41] psychosis? Like are like do we just have [11:45] we like gotten like, you know, totally [11:47] whipped each other into this frenzy? And [11:49] and there was there wasn't [11:52] there was like no feedback to keep us in [11:55] check or sane from the outside world. [11:58] And [12:00] it doesn't sound that weird relative to [12:03] like, you know, crazy board drama or [12:06] Elon trial or something like that, but [12:09] living through it was an unbelievably [12:12] strange time. [12:15] Um, what's the um what's the Sam [12:18] Almanagement [12:20] style? If I'm if I'm working for you [12:23] directly or maybe indirectly I'm you [12:25] know leading some product or or [12:26] something just what does that look like? [12:31] Um [12:35] I'm definitely not a hands-on manager. I [12:37] I'm very much the [12:40] of the of this style that you get great [12:43] people um kind of give them like a very [12:47] high level thing to point at. um and try [12:51] to let stuff just happen. Um [13:00] I I think there have been [13:03] kind of two main phases of open eye and [13:05] we're heading into a third and [13:08] the first one was like we were we were [13:10] only a research company. We were trying [13:11] to figure out how we were going to build [13:14] AGI at a time when it sounded completely [13:16] completely crazy. uh and we really had [13:18] no idea what to do. And then there was a [13:20] second phase where in addition to [13:21] continuing to do that, we had to figure [13:23] out how to build a product company. Um [13:26] now we have to in addition to both of [13:29] those two things, figure out how to [13:31] build this like mega mega scale token [13:34] factory for the world. Like like I I [13:35] think of what we're doing is building [13:38] sort of a new utility. Um and people are [13:43] going to want to use a lot of tokens, a [13:44] lot of intelligence in all sorts of [13:46] ways. We need to make that as smart, as [13:48] cheap, as abundant, as easy to use as [13:50] possible. Um, and that will require, I [13:52] think, pretty deep full stack [13:54] integration and a massive, massive [13:56] infrastructure buildout. Um, [13:59] the thing that I didn't really [14:00] appreciate between the phase one phase [14:03] two shift was how much my management [14:05] style had to change. You know, running a [14:07] research lab and running a like a [14:10] product company are two extremely [14:12] different things. And I suspect this [14:14] third phase is going to be very [14:16] different yet again. And so I've been [14:18] reflecting on, you know, if we're going [14:19] to really go do this, how I'll have to [14:22] change. Um, [14:24] and I think it's not going to be like a [14:26] natural fit for my management style. Uh, [14:28] so I either have to find someone or a [14:31] few people great to hire or I have to [14:32] figure out how to do things in a [14:33] different way or uh I have to like build [14:35] an AI that can manage this new thing. [14:38] When I interviewed Jensen here uh two [14:40] years ago, he told me and you know his [14:44] several thousand closest friends uh [14:45] about his 60 direct reports. Do you have [14:48] any super weird not that that I [14:50] shouldn't call his practices weird, but [14:52] uh do you have any unusual practices [14:53] like that? [14:55] >> I think the closest thing that I have to [14:58] anything like that is I I probably talk [15:00] to like via Slack or whatever or text [15:03] like a few hundred people at the company [15:05] a day. [15:06] um very quick like you know one two [15:10] messages whatever not done by an agent [15:12] like I actually do it and the the [15:16] context I get from that [15:18] sometimes is like very helpful in these [15:20] diffuse ways. [15:21] >> I find there's an interesting watershed [15:23] uh of pre-Slack organizations post Slack [15:27] organizations and they're truly quite [15:30] different. [15:30] >> Totally. Um, I [15:33] like many other people, I hate Slack, [15:35] but I can't imagine [15:37] having to like still communicate via [15:39] email or whatever we used to. [15:41] >> That's roughly where Stripe is. [15:42] >> Yeah. Yeah. Um, [15:45] so, um, okay. I want to talk about, um, [15:48] I mean, you kind of just elliptically [15:50] referenced it. Um there's uh a view that [15:54] the AI labs are going to uh progress up [15:58] the stack gobbling up the value chain [16:02] voraciously all these things uh that are [16:05] certainly within the software sector but [16:07] perhaps even other sectors and that [16:09] there'll be this incredible positive [16:11] feedback loop and runaway and um kind of [16:15] hegemonic force that we should all be [16:17] getting very concerned about. Um [16:20] what's your view? I think some of them [16:22] do want that. Um we don't I one of the [16:26] things I've always admired about Stripe [16:28] is it is like very clear that Stripe is [16:31] aligned with its customers. You know we [16:35] make more revenue. We charge our [16:36] customers more. Um thank you by the way. [16:38] The partnership with Stripe when Chatbt [16:41] launched was extremely critical and I [16:43] don't think anyone else could have [16:44] scaled that quickly but we scaled. We [16:46] pay you more money. it's like very [16:47] aligned and we're all happy and you just [16:50] provide like a layer of infrastructure [16:52] for the internet. Internet gets bigger, [16:53] you're happy, your users are happy. It's [16:54] clear what the alignment is. Um [16:58] I don't know exactly how to do this yet, [17:00] but I would like to get to a model for [17:02] OpenAI that is similar. Like I would [17:05] like us to be an infrastructure [17:06] provider. I'd be happy for us to be like [17:08] a forever low margin as long as we can [17:10] be huge and growing fast business. And I [17:12] would like us to supply kind of [17:17] an intelligence meter I don't know what [17:19] quite to call it that companies can buy [17:22] that they can you know use to automate [17:25] things accelerate things inside their [17:26] company they can use to build products [17:28] people can buy it people can take it [17:30] with them and we find ways to really [17:32] align ourselves with the success of all [17:34] the entire gigantic distributed [17:38] um economic engine of the world I [17:41] believe that will work um I believe that [17:44] switching costs of AI it's going to be [17:46] hard to have like huge margins in AI [17:48] anyway. It's like you've seen recently [17:50] how easy it is, many people have seen to [17:52] switch from our competitor's coding [17:54] product to ours. Um, this is actually a [17:56] consequence of AI getting smarter. It [17:57] gets easier to do things like this. It [17:59] gets easier to just say like, "Hey, [18:00] agent go do this thing for me." Um, but [18:03] if we can provide a utility and people [18:04] build on top of that utility, uh, and [18:08] you know, we think of ourselves as that [18:10] kind of a company, I think that can be [18:11] quite powerful and very aligned. Well, [18:13] we're we're happy to share lots of tips [18:15] and tricks for being a a low margin [18:17] business. Um, [18:19] but um [18:21] so many people um [18:25] um have been [18:27] um either implicitly critical or in [18:31] certain cases explicitly critical uh of [18:34] OpenAI for um for procuring so much [18:38] compute. Uh and I think not the codeex [18:42] users, [laughter] [18:44] >> right? Exactly. So um you I think were [18:48] quite noteworthy for as early as I [18:51] exactly but you know on the order of two [18:52] or three years ago um stipulating what [18:55] at the time sounded like preposterous [18:57] figures uh with respect to the magnitude [18:59] of the buildout that would be required. [19:01] And obviously the preposterousness of [19:03] those figures now looks um less uh less [19:07] uh tenuous by uh by the day. Um [19:11] thoughts on compute capex the buildout. [19:14] Um [19:16] yeah it's going to take a lot of money. [19:18] Uh it I think this will be clearly at [19:22] this point the most expensive [19:23] infrastructure project that the world [19:26] has ever undertaken. Um the revenue is [19:30] ramping to meet it. So people feel [19:32] better about that. Al also the [19:34] efficiency gains that we've all been [19:36] finding are incredible. So we're going [19:38] to get way more out of each GPU than I [19:41] thought we were going to. Uh but as has [19:44] often been remarked like the demand goes [19:47] up more than linearly as you drop the [19:48] price of each kind of unit of [19:50] intelligence, particularly if you can [19:52] drop the price and the sort of speed [19:53] with which you get it back. Uh so [19:59] this this question now of like what is [20:01] enough [20:02] I don't have a good answer to. I don't [20:04] in some sense I think demand for [20:06] intelligence at a low enough price um is [20:09] effectively uncapped. Now, [20:14] I was going to say we're not, but maybe [20:16] we are. Like, you know, we're not going [20:17] to build a Dyson sphere and then just [20:18] like cover it with data centers, but [20:20] maybe we do. [20:23] Space data entries. [20:26] Good luck with that. [laughter] [20:28] Um, I don't even think he's that serious [20:30] about it. [20:39] [applause] [20:44] Um, [20:45] I don't myself think that we're in a [20:47] capex bubble. I'm not an expert in this. [20:50] This is not, you know, Stripes business, [20:51] but just the figures I see relative to [20:54] the magnitude of demand, it looks [20:55] reasonable to me. If we were in a capex [20:59] bubble in the future, how would we tell? [21:05] People love to proclaim bubbles. [21:09] I [21:10] And I I can't articulate why, but [21:12] intellectually I kind of get it. Like [21:14] it's it does feel fun and it feels [21:16] smart. Journalists in particular love to [21:19] um you know talk about bubbles. So there [21:22] there's like ample desire to write about [21:24] this when anything looks a little bit [21:26] silly. And clearly sometimes it's right [21:28] like there there clearly are bubbles. [21:31] But how [21:34] you can discern between the amount of [21:37] times someone calls a bubble and the [21:39] amount of time you're actually in a [21:40] bubble, I have never figured out how to [21:42] do. In my previous career, I was an [21:44] investor. So, I was like quite [21:45] interested in trying to see if I could [21:47] come up with some sort of framework for [21:48] this and figure out when you're supposed [21:49] to deploy capital or not. And I never [21:52] was able to figure it out. Um the like I [21:56] went back and I read what smart people [21:58] had read had said at different points in [22:00] history and I was like oh they called it [22:01] exactly right. But then I read a little [22:03] more and they said it like 10 more times [22:05] than the 10 previous years. I don't [22:07] know. I don't have an answer. Economists [22:09] are the people who have called eight of [22:11] the last three recessions. [22:16] But they're so happy when they're right. [22:18] [laughter] [22:19] So um [22:22] you know a lot of I mean your business [22:25] open AI um depends in a very significant [22:28] way on super talented people. Um and the [22:32] difference as I understand it between uh [22:34] the uh the um you know 20th most [22:38] talented person versus the fifth most [22:39] talented person versus the most talented [22:41] person might be quite large andite quite [22:44] consequential. And then these super [22:47] talented or effective people, you know, [22:51] in they're not in every case super easy [22:54] to work with. [22:54] >> No. Um, [22:57] and [22:59] look, and some of them are wonderful [23:01] people and some of them are the most [23:02] fantastic collaborators and some of them [23:03] are very iconoclastic and strong willed [23:06] and they get easily, you know, whatever [23:07] just like it's a full spectrum of of the [23:09] human condition. Um, but I guess I'm [23:11] curious in a domain that's so sensitive [23:13] to this efficacy and skill and talent [23:15] and so forth kind of intersected with [23:18] all the foibless of humans as they [23:20] exist. How do you think about this? Like [23:22] do you guys tolerate primadanas? Do you [23:26] tolerate them more than you used to, [23:27] less than you used to? Do you try to [23:29] manage them in a special way? How do you [23:31] think about managing elite skill here? [23:35] Someone was working on this book, [23:36] OpenAI, and said to me like, I I think I [23:40] I think I figured out the thing that you [23:41] you sort of were really great at and [23:43] kind of did uniquely well in making [23:46] OpenAI happen. And I was like, I would [23:48] love to hear. I have no idea what the [23:49] next sentence is going to be. Like, I [23:50] could not predict the next sentence. Um, [23:56] and they said, you know, you figured out [23:58] how to get [24:00] a lot of people who all thought they [24:02] were the only capable or most capable [24:06] person uh, and everything had to go [24:08] their way to work together long enough [24:10] to like figure out the breakthroughs and [24:12] that, you know, that was the that was [24:14] the magic of OpenAI and Okay. So, so [24:17] what's the trick? Uh, a a lot of pain. [24:20] Um [24:23] I I think we [24:26] we had very even when people didn't like [24:29] each other and even when people thought [24:31] they were much smarter than other people [24:33] or had a better approach than other [24:35] people um we had a few deeply shared [24:39] convictions. uh like we we did kind of [24:42] collectively believe in scale and [24:45] concentrating resources and that we were [24:46] going to do this one thing and that we [24:51] thought getting this right was important [24:54] enough that people were going to like [24:55] put aside various personal conflicts. [24:58] Um, [25:01] one of the most unusual things about [25:04] OpenAI is at the time we trained GPT3, [25:06] the vast majority of our compute at [25:10] at the whole organization was going into [25:12] this one single research program and we [25:16] would talk to, you know, people that we [25:18] were like trying to recruit from Deep [25:19] Mind at the time and they would say, [25:22] "That's insane. It's going to create [25:23] this terrible culture. [25:48] We love music. [25:54] Um anyway, they would say like [25:58] >> we're we're talking about managing [26:00] unusual personalities. [26:01] >> Yeah. [26:04] Um [applause] [26:10] [applause] [26:13] they they would say you have to like [26:15] divide your compute equally otherwise [26:16] you'll have this like very toxic [26:18] competitive culture and you know this [26:19] thing and that thing and and we would [26:22] just take the approach that like we're [26:25] going to bet with conviction on this. [26:26] It's not going to feel totally equal but [26:28] this is the right thing and we do think [26:30] [music] we know the thing. We do think [26:31] we know the direction. we we we really [26:33] want to go in and they would say, "Well, [26:35] you know, you might be wrong. Like, we [26:37] have to do those other things. You have [26:38] to have this research program." And [26:42] and having a culture where we said, [26:43] "We're going to [26:45] we're going to have conviction and do [26:47] this uh and ignore the distractions um [26:50] was great. [26:52] We hope this is a memorable sessions for [26:54] all of you. Um, [26:57] [applause] [27:07] so John and I have been doing this thing [27:10] at Stripe for, you know, quite a while [27:11] now. We started out in 2010. Uh, you and [27:14] Greg started out in 2015. And to your [27:17] point, you've um ventured through many [27:21] trials and tribulations and [27:23] stratospheric successes and you know all [27:26] the rest. Um that was a nice little [27:28] gloss over but go ahead. [laughter] Um [27:32] thoughts on I mean it's it's not easy to [27:35] work successfully with a co-founder for [27:37] now more than a decade uh and for things [27:40] even after a decade uh to seemingly work [27:43] as well as they did uh you know from the [27:45] from the beginning. just thoughts on [27:47] thoughts on that partnership. Why has it [27:49] worked? Um, and how have you guys made [27:51] it such a success? [27:53] >> Obviously, you and John knew each other [27:55] for longer than Greg and I did. Um, but [27:58] Greg and I did know each other for a [27:59] long time before OpenAI and I think [28:02] having the shared history really helps. [28:05] One of the things that I had observed [28:08] at Y cominator was that one of the [28:11] biggest predictors of success was had [28:14] the co-founders known each other for a [28:15] long time or at least relative to their [28:18] lives for a long time and the teams that [28:20] came together like you know 7 days [28:22] before applying to YC on a co-founder [28:24] matching site or whatever that was that [28:27] didn't work too often. It's not [28:29] impossible. I think there one or two [28:30] cases where it did work but but it was [28:31] rare. Uh so we had known each other for [28:34] a while and we had kind of had a sense [28:36] of sort of shared values and history and [28:40] ecosystem and kind of we were clear on [28:43] what we wanted to do. Um and [28:51] and I think we had this like deep mutual [28:53] respect and complimentary skill set that [28:56] has [28:58] just worked really well. Um [29:02] the [29:04] I'm extremely grateful. I think having [29:05] to go through a any startup experience [29:09] but particularly like a intense one [29:11] without a co-founder you have like a [29:13] deep connection trust to is is really [29:16] hard. I've watched people do it but it's [29:18] very hard. So um I am you know extremely [29:21] grateful that we've gotten to do this [29:22] together. [29:26] >> [applause] [29:33] >> on a adjacent topic um you know we've [29:37] talking about open AI but then there's [29:39] this entire ecosystem um of companies [29:42] and startups and uh and enterprises that [29:45] are you know building on the platform. [29:46] It's obviously an interesting moment in [29:48] startups um given on the one hand the [29:53] ability to build products and generate [29:56] revenue at seemingly unprecedented rates [29:59] and certainly we see this in the stripe [30:00] data like the number of businesses [30:02] reaching thresholds um you know [30:04] meaningful thresholds is far faster than [30:06] ever has been before um you are one of [30:09] the most prolific and successful startup [30:11] investors ever. You of course ran Y [30:13] Combinator. Have the traits that make [30:16] founders successful changed in this era [30:20] or is it sort of the same thing it's [30:22] always been? [30:24] Uh [30:26] there was a time when we used to make [30:28] fun of the idea guy. There were these [30:30] people wanted to start a company and [30:32] they'd say like I have the best idea. [30:36] I'm not going to tell you what it is. I [30:37] have the best idea. I just need a coder [30:39] to like build it for me and then I'm [30:41] gonna be in great shape. And we would [30:44] like make fun of these people. They [30:45] weren't that successful. Um the [30:49] and it was like kind of always [30:51] personally annoying to me cuz it's like [30:53] it it would be like saying like I have a [30:55] great idea for [30:57] a song and I just need that guy with the [30:59] guitar to like make it for me. Um [31:03] and [31:06] and so it kind of it didn't it didn't [31:08] work like you know NYC had a version of [31:10] this which is teams without nontechnical [31:12] founders are difficult to get work all [31:15] of a sudden it's like the revenge of the [31:16] idea guys which is actually awesome for [31:19] the world like I'm happy I'm here for it [31:21] for sure um but [31:24] for a long time I think the most [31:26] important ingredient that [31:29] I looked for YC looked for that kind of [31:31] this part of our industry looked for on [31:33] a founding team with technical talent [31:36] and that's still very important but now [31:39] people who just really deeply understand [31:41] their users and can't code at all I like [31:43] want to fund those people and that's a [31:45] that's a big turnaround [31:49] [applause] [31:53] >> how does one think about startup [31:55] investing these days because on the one [31:57] hand you have a couple of years [32:01] potentially to AGI or ASI or the [32:03] Singularity, who knows what. And then [32:05] you have investing time horizons or [32:08] funds with 10ear time horizons. How does [32:12] that all fit together? Does it? Uh [32:18] I think to do anything at this point on [32:20] 10ear time horizon requires a real [32:23] suspension of disbelief. Uh and yet [32:26] that's probably the right way to live [32:27] your life. Like I don't think it works [32:29] to say there's this singularity in three [32:32] years or five years, whatever. We can't [32:34] see past it and so we're going to do [32:35] nothing or we're just gonna like give up [32:37] or we're going to go crazy or whatever. [32:38] You like you have to live as if stuff's [32:40] just going to keep going in an [32:42] understandable way for a long time. [32:45] >> How how far ahead does OpenAI plan? [32:50] Uh [32:52] I mean we signed like 20 year power and [32:54] land agreements and for the product [32:59] um [33:01] I think we have like a clear vision of [33:03] what things can look like in [33:06] two years and then it gets much hazier [33:09] after that. [33:11] So there was in the relatively recent [33:13] past uh a narrative that um you know GPT [33:18] rappers and companies of that ilk were [33:21] um were I guess as the porative suggests [33:24] undifferentiated and flimsy and be swept [33:26] away by a rising tide of model [33:28] improvements. Whereas now it feels like [33:30] that narrative in some sense is flipping [33:32] somewhat for now instead of talking [33:34] about rappers we talk about harnesses [33:37] and harnesses are seen as having this [33:38] you know significant heft and uh and [33:40] importance and I guess I'm curious for [33:43] your view on this and how you view uh [33:46] businesses for which AI is a critical [33:48] enabling component and their prospect of [33:51] durability. [33:55] I I have kind of had [33:58] uh the same view all the way through [34:00] which is you as a business want to be on [34:03] the side of hoping that AI gets smarter. [34:07] Um and whether [34:09] so so like in the early model days if [34:11] you were the GPT rapper um and you were [34:14] like patching some kind of weakness in [34:16] the current model that was clearly going [34:17] to get better with the next model. If [34:19] the next model was much better you were [34:20] kind of sad. If you were doing something [34:22] that got better like you know you're [34:23] making [34:26] any of the wonderful like services that [34:28] people were building with the models [34:30] that benefited from intelligence you you [34:31] would be happier. I think the same thing [34:33] in the world of harness is like I I kind [34:36] of think the right way to think about [34:37] this is like data center [34:40] model harness like that whole thing is [34:42] just this one cluster out of which comes [34:44] this like very usable intelligence. Um [34:46] but there are so many things to go build [34:48] where you're just like happier for that [34:50] whole cluster to get better and better [34:51] and better. Um, and then if you're kind [34:54] of secretly hoping it doesn't because [34:55] you're patching some weakness in that, [34:57] probably like the next model crank turn [35:00] somewhere in that stack is just going to [35:01] solve it. When you look at the [35:04] organizations that are making the most [35:05] effective use of AI today, like if you [35:08] um I mean you meet OpenAI customers [35:10] constantly, large and small, if you [35:13] think about the top three that have [35:15] impressed you the most or the one that's [35:17] impressed you the most, what [35:19] specifically are they doing that's [35:21] different? like everyone here knows yes [35:23] AI big deal we should make enthusiastic [35:26] use of it etc. But what specifically [35:28] differentiates those which in your [35:30] opinion are employing it most [35:32] effectively? [35:33] >> Uh a a few different directions there. [35:36] um friend of both of ours, uh Toby Lucky [35:39] of Shopify was the first CEO I knew that [35:41] just said like we are going to be all in [35:44] on AI and the way we run our company and [35:47] he got himself got his hands dirty just [35:49] building like AI automation of [35:51] everything and made his team do it. [35:54] >> Um and you know said we're just going to [35:56] figure out how we take all these things [35:58] that are bad and make them good with AI. [36:00] And it was not like you know a token [36:03] leaderboard. It was not some other kind [36:04] of like gamified [36:07] hackable thing. It was just like the CEO [36:09] of the company said, "We are now going [36:11] to put AI into everything we do and I'm [36:14] going to like not be happy with you, I [36:17] guess, or we're not, you know, we're not [36:19] going to allow it if you're not doing [36:20] that." So, that energy has now been done [36:22] by other people. But when um when the [36:25] CEO of a company just says like we're [36:28] going to automate ourselves, accelerate [36:29] ourselves, however they phrase it [36:30] internally, um and then really holds [36:32] [clears throat] people to it and ideally [36:34] does it themselves. Um [36:37] that has worked very well. Uh, I think [36:40] we're going to try this new experiment [36:41] where we start [36:43] sending like an FD to work with like [36:47] hands-on just whatever the CEO of a [36:49] company needs work with them to automate [36:50] their job. Um, as much of it as they can [36:53] and that I think will have like if you [36:57] just do it for the leader of a company [36:58] there's like a nice fractal effect [36:59] throughout the company. So that works [37:01] and we'll try to help companies do that. [37:04] Um, [37:06] a second thing is being like [37:09] uncomfortably permissive with data [37:11] access. Um, there's huge reasons not to [37:14] do this and I'm not this is like I'm [37:16] stopping short of recommendation. You [37:17] just ask what I've seen from the most [37:20] effective companies. [37:21] um this is easier for small startups [37:24] than companies that have a lot of [37:25] sensitive data and a lot of um [37:30] you know like process and compliance in [37:33] place but saying like you know what we [37:35] are going to record our meetings we are [37:37] going to like let this AI have access to [37:39] our codebase we are going to let this [37:41] have access to every Slack message every [37:42] email everything and every employee at [37:44] the company is going to get to use it [37:45] that way like it is amazing watching [37:49] these two or three person startups and [37:52] AI doing everything work. Um, [37:57] and I don't know how the world is going [37:58] to decide the trade-offs on data privacy [38:01] versus AI efficiency. [38:03] And I think there's like some regulation [38:05] that's going to have to change for that. [38:06] Uh, but it's so powerful. [38:09] Um, Tempo is a new blockchain uh that [38:12] Stripe incubated with Paradigm uh which [38:15] obviously we're partners uh with uh with [38:16] you guys on. Um, and uh, it the mainet [38:21] just launched, but the project launched [38:24] uh, back uh, last uh, last summer. So, [38:26] it's a relatively new and uh, and small [38:28] team, a couple dozen people. The Tempo [38:31] team um, set up a um, a harness um, a [38:37] tool in their uh, in their Slack [38:40] installation for orchestrating pretty [38:43] much everything at the company. [38:45] Everything. You can just ask any task. [38:47] Go and read these Google Docs. Uh turn [38:51] those into a bunch of um uh a bunch of [38:53] linear tasks, then go write a pull [38:55] request to, you know, implement them. Uh [38:58] then go deploy them and use our log [39:01] analysis tool to test that the [39:02] deployment actually worked. And the [39:03] agent will happily go and employ tool [39:05] use across all of this. And it's [39:06] extremely trippy watching a whole organ, [39:09] I mean a small organization, but an [39:10] organization of people do everything in [39:13] a single Slack channel. And I don't [39:15] think that would scale to stride. It was [39:16] the first time I had the experience [39:17] you're describing, which is it's clearly [39:19] incredible for them. I don't quite see [39:21] how to transpose it for us, but this is [39:23] really something. It's really something [39:25] to watch. Um the [39:28] and I find that a lot of people just [39:30] can't. This is where there's a big [39:31] overhang. They have not yet been able to [39:33] wrap their heads around the fact that [39:35] you can just kind of ask us anything and [39:37] it'll probably happen. Um I I myself [39:40] still find myself like not trusting [39:43] quite enough that it's going to be [39:44] possible. I don't know exactly how it's [39:46] going to transpose to bigger companies. [39:48] It does feel like we're missing kind of [39:50] like one more abstraction there. Um like [39:53] how humans and AIs are going to [39:54] interface at massive scale. The [39:56] advantage that these smaller companies [39:57] have is like it's just the AIS. They [39:59] don't have to figure out the interface [40:01] with all the all the people. Um but [40:03] we'll figure it out. [40:05] Open source AI. Where's it going? Do [40:08] they have a future? For sure. Uh [40:13] right now people clearly want smarter, [40:17] faster, cheaper frontier intelligence [40:20] and most of the demand is there but [40:22] there is also a lot of demand for open [40:24] source and I expect that to increase [40:27] relatively over time. [40:30] So, I want to talk a little a little bit [40:32] about science um because I know it's [40:33] something you're very excited about. Um [40:36] but it's also uh relevant to uh to [40:40] something I spend some of my time on uh [40:42] which is uh which is the Arc Institute. [40:44] Um and [40:48] OpenAI is in fact a foundation um a [40:51] nonprofit and recently um recently made [40:55] a grant uh to the ARC Institute. uh and [40:58] so maybe we'll get to the details of [41:00] that uh in a second but first you just [41:02] want to speak a little bit to AI as [41:05] applied to science you know what you're [41:06] seeing and just how you think about [41:08] grant making generally uh in the context [41:10] of the uh open eye foundation so [41:15] generally on on the AI and science [41:17] question I I hope that this will be the [41:20] most important contribution um of [41:24] AI of this technology to human quality [41:26] of life over [41:28] And that if we can start to discover new [41:30] science at a much faster rate which can [41:32] be like new materials or cures to [41:34] diseases or any number of other things [41:36] like I believe that to a first order [41:39] approximation life gets better because [41:41] we understand science better and then we [41:43] figure out how to build stuff with it [41:44] and distribute it to people. um [41:47] starting with the models of a few months [41:50] ago but really now with 5.5 the models [41:54] have gotten smart enough that excellent [41:56] scientists are saying I am able to [41:59] figure out better ideas the models are [42:01] able to make some small but important [42:04] discoveries um and the pace of science [42:06] is going to increase eventually we'll [42:08] have you know automated labs and robots [42:10] and you know building who knows what uh [42:14] and we'll be able to do science much [42:16] faster. But if we can start doing like a [42:19] decade of science of what it would have [42:21] taken us in the old world in a year, the [42:24] compounding effect there and what we'll [42:25] be able to do and discover will will [42:27] just be extremely great. So uh I think [42:31] this is going to be incredible and this [42:32] will be one of the big areas of focus of [42:35] the of the OpenAI Foundation is uh [42:38] basically like money and expertise and [42:40] technology to accelerate science and [42:42] trusting that that will [42:44] you know flow to the world in wonderful [42:46] ways and this is going to be a big [42:47] foundation. [42:48] >> Yeah, this will I think it is [42:51] it's one of the biggest maybe it's the [42:53] biggest I think it will be the biggest [42:54] foundation um in the world. Uh [42:58] and yeah, so we're really focused on [43:00] science and then AI resilience um like [43:02] helping the world through this [43:03] transition uh with this new technology [43:06] in it. Um [43:09] but you know, we were thrilled to get to [43:11] support ARC. I think it's clearly the [43:14] best sort of AI and bio effort and if we [43:19] can make even a small contribution to [43:22] with this technology and with the [43:23] capital and the foundation to helping uh [43:28] make people healthier, treat diseases, [43:30] this whole cluster of what we can do as [43:32] we get better at understanding biology. [43:34] Um we will be very thrilled. I thought [43:37] that was going to take longer and [43:38] looking at the incredible work [43:42] the Arc Foundation is doing. Uh, and now [43:44] I think it maybe won't be won't be that [43:45] far off. So, you know, in a podcast, um, [43:49] midway through you might hear a little [43:51] interstitial ad. Um, this is your [43:54] interstitial ad for the Arc Institute. [43:56] Um, there is, um, there's an Arc [43:59] Institute booth downstairs. Uh, and you [44:02] might wonder why there's uh, what what [44:04] it's doing at the um, at the Internet [44:05] Economy Conference. Um and uh the uh the [44:09] context here is so the the ARC Institute [44:11] is an organization we started four years [44:13] ago. Um and its goal is to uh produce [44:17] hopefully the first cure for a complex [44:20] disease in humans. So a complex disease [44:22] is uh is one that involves some genetic [44:24] factors and some environmental factors. [44:26] So you can think of most cancers, most [44:27] autoimmune disease, most [44:28] neurodeenerative disease uh for example [44:30] as being a complex diseases in this kind [44:32] of specific sense. And humanity has [44:35] never cured a complex disease. Not one. [44:37] We've cured lots of infectious diseases. [44:40] We know how to screen for monogenic [44:42] diseases for this one genetic mutation [44:44] that um that undergurs it. We've never [44:46] cured a complex condition. Um so we [44:51] started our institute with this is the [44:52] goal. Alzheimer's uh is the first [44:55] complex disease uh that we're uh that [44:57] we're targeting. Um and our hope is that [45:01] with both new genome engineering [45:02] technologies like crisper and then sort [45:04] of the amazing advances in AI uh that uh [45:07] that we'll be able to make some [45:08] hopefully meaningful progress and it's [45:10] only four years old uh but the early [45:12] results uh are very encouraging. The ARC [45:16] Institute is currently looking for a [45:18] CTO. [45:20] Um uh we had uh one CTO uh do a sbatical [45:25] at ARC uh last year. uh his was it last [45:29] year before? It was last year. Um and [45:30] his name was Greg Brockman. Um and he [45:33] did some great stuff. Uh he helped us [45:35] train Evo 2, which is the largest [45:38] biology foundation model uh ever [45:40] trained. Um but we're looking for a [45:42] full-time CTO. Um uh and uh the um uh we [45:48] thought that well perhaps that person [45:51] might be in this audience and if not [45:53] their friend might be in this audience. [45:55] So, if you know somebody for whom that [45:56] sounds of interest, go check out the arc [45:59] stand downstairs and that's your [46:01] interstitial ad. [46:04] [applause] [46:09] I think it's so much better that you [46:10] label that as the interstitial ad rather [46:12] than just doing it. That was good. [46:15] Um, [46:16] okay. Well, we haven't talked about [46:17] Stripe. So um [46:21] you were the um you were the I think it [46:23] was second investor in Stripe. [46:25] >> Was YC the first? [46:26] >> You and Paul at the same time in the [46:29] same kitchen in fact. Uh so maybe you [46:31] were first in fact. I don't remember in [46:32] which order the checks were handed over. [46:33] >> Oh yes, this was his like Palo Alto [46:35] kitchen. [46:35] >> Yeah, that's right. Um so you know we [46:38] were sort of you know we were two pimply [46:41] teenagers uh proposing building this [46:43] financial services institution. It [46:45] sounded like a bit of a a ludicrous [46:47] proposition. Why? Why did you invest? [46:50] Um, honestly, I I he didn't team me up [46:54] for this. Uh, [46:56] you and John were two of the most I I'd [46:58] seen a lot of founders and you and John [47:00] were two of the most impressive founders [47:01] of any age, but certainly of pimply [47:04] teenager age um that I had ever met. I [47:08] clearly was right about that. Um, and [47:12] I had like known of you and I had heard [47:15] Paul talk about you and I had like known [47:17] of you on the internet. Um, [47:20] and [47:23] I was struck that you were solving a [47:24] problem for yourself. Um, and that it [47:28] was sort of like it fit a trend that I [47:32] thought was going to be big [47:35] in the world. um like I kind of really [47:38] believed that commerce was going to move [47:39] online in a huge way and there were [47:40] going to be lots of startups and both of [47:42] those suggested that this would you know [47:44] could be really big but um I don't know [47:47] I was like a believer that if you can [47:48] find really smart and still am if you [47:50] can find like really smart founders and [47:52] a market that's going to be big you [47:54] should just invest and that's kind of it [47:57] and based on your um well just based on [48:01] your general perspective on the world um [48:02] but then also OpenAI's use of Stripe and [48:04] what you've seen from that like what you [48:06] see OpenAI needing and needing in the [48:09] future and building the business and so [48:11] forth. What's your advice for Stripe on [48:14] navigating this um the AI era before [48:18] that? Just to check my memory, the thing [48:20] you were building the re the reason you [48:21] realized you needed payments was you had [48:23] like built this iPhone app to download [48:25] all of Wikipedia cuz you were going [48:26] somewhere crazy offline and it was like [48:29] hard for you to take payments for it and [48:31] that was like [48:34] hard. Yeah. Okay. I hadn't thought about [48:36] that in a long time. Um [48:43] I I think my my advice more towards like [48:45] Stripe as a company itself. [48:47] >> Yeah. Going forward um like it's it's a [48:49] crazy time in the world. Uh there's a [48:50] lot changing a lot a lot happening. Um [48:53] and again you've seen Stripe from the [48:55] perspective of a customer. So what um [48:57] what what are your complaints and [48:58] feature requests? [49:00] >> [laughter] [49:01] >> Well, it's more like I want to I now [49:02] have now that I'm thinking that we [49:04] should be thinking more like a stripe [49:06] style model. I have a bunch of question. [49:07] I probably have more to learn from you [49:09] than you you do from me here. Um cuz I [49:12] remember like when investors would [49:13] outsmart themselves saying like, "Oh, [49:15] Stripe's going to be a commodity. When [49:16] everybody gets big, they're going to [49:17] just like build their own thing." And it [49:19] turns out that like yes, there are [49:21] multiple payments providers. And yet in [49:23] practice, if you make a great product, [49:25] if you make like a great thing, people [49:27] are still going to need to take money in [49:28] the future and they're probably just [49:30] going to like stick with you if you're [49:31] like a good reasonable ecosystem partner [49:33] and don't do crazy things. Um, and I [49:36] think that's going to keep working. I I [49:38] actually like I think every company [49:39] clearly does need to get more efficient [49:42] and with AI and they will do that. But [49:45] this this whole [49:47] mindset that like every company is going [49:49] to go away and everything is going to be [49:52] completely different and like maybe [49:54] it'll be agents that need to handle [49:57] payments between each other instead of, [49:59] you know, consumers and merchants, but [50:02] like clearly money is going to have to [50:04] move somehow. And my my advice would be [50:07] like adopt AI especially internally. Use [50:11] AI to build better products but don't [50:13] assume that the like entire [50:15] socioeconomic system completely [50:16] reconfigures. I think the world has [50:18] gotten a little bit delusional about [50:20] this. [50:24] Um apart from AI as we look out over the [50:27] next decade just what are the [50:29] technologies and areas that excite you? [50:37] um sort of AI at the kind of model and [50:42] product layer. Um I'm obsessed now with [50:45] data center infrastructure [50:48] like all see I [snorts] think there's so [50:50] much cool stuff to do there. Uh I think [50:52] there will be amazing new technologies [50:54] like at the physical layer. Uh energy [50:57] robots [50:59] those are made like that stack is [51:02] probably what I think about the most. Um [51:05] it does seem like the world is finally [51:06] making a little progress on um brain [51:09] machine interfaces. I'm excited about [51:11] that. Uh I am excited about [51:17] well I'm both afraid of and I'm excited [51:19] about progress in biotech. Um, but you [51:22] know, I'm certainly hopeful that that [51:23] can get much better quickly and and I [51:27] think defensive biotech is about to [51:28] become very important unfortunately. Um, [51:32] I'm excited about new kinds of computer [51:33] interfaces. I think we are in a insane [51:38] area right now where we're stuck with [51:41] these kind of old devices and old [51:43] operating systems and we have this magic [51:46] new enabling technology [51:49] and it feels like Codex is amazing and [51:53] what it but it feels like very broken to [51:55] be like telling this thing to use my [51:56] computer and then it's like clicking [51:57] around and there's all the stuff that [51:59] was made for a person but doesn't really [52:01] make sense for like an AI to go use like [52:03] we can do so much better there. I think [52:04] there's a whole new internet protocol to [52:05] make too. Um, so those things. [52:09] >> When do you think we'll have the world's [52:11] first profitable nuclear fusion reactor? [52:17] >> Depends how far electricity prices get [52:18] pushed by data center demand. Maybe [52:20] sooner than we thought. Um, [52:27] I'll guess in the next 5 years. [52:31] It's a bold prediction. I hope so. [52:35] >> [applause] [52:37] >> Just going to hope we're on a roll here. [52:38] Um, hypersonic commercial air travel. [52:42] Don't follow it as closely. Um, [52:45] I [52:49] hypersonic meaning like Mach 4. Sure. [52:58] A good chunk of time. [snorts] [53:02] I don't know more than 10 years. [53:03] >> Okay. [53:04] >> Maybe a little less but something like [53:05] that. [53:07] >> Um are there any domains of science [53:09] technology that are not in the discourse [53:11] in a significant way that you think will [53:13] be accelerated a lot by AI and will have [53:16] broad kind of social consequence and [53:17] impact? [53:19] >> Yeah. The one that I think just does not [53:20] get enough attention is material [53:22] science. Um, [53:25] it's like a very it's not a cool thing [53:28] and I think people underestimate how [53:30] much of the world is materials like how [53:32] much of what we depend on and how much [53:34] progress AI can make like it's such a [53:36] beautifully AI AI shaped problem just [53:38] getting new catalysts totally. Did I [53:41] expect very rapid progress there and [53:45] that it'll impact all of our lives in a [53:46] very positive way and it gets like very [53:49] little attention? [53:51] Last question. Um [53:54] it feels that in some sense [53:57] at least some version of AI is kind of [53:59] inevitable. There's lots of people [54:01] rushing towards it and building it out [54:03] and creating the data centers and uh and [54:06] all the rest. So there's this kind of [54:07] sense of determinacy. Um, how do you [54:11] hope that your specific involvement [54:14] changes the trajectory for the world [54:16] relative to some other counterfactual? [54:20] Um, [54:23] I believe in democratization and [54:27] personal agency and access and that [54:29] everybody deserves a really great life. [54:31] Um, the most controversial decision we [54:33] made in the history of OpenAI was um, [54:36] what we now call iterative deployment. [54:38] But a lot of the thinking at the time [54:40] that we released chat GBT was this was [54:42] insanely dangerous uh to do. You can't [54:44] do this. Only this like small set of [54:47] people who have been thinking about AI [54:48] safety can know what's coming. It's an [54:49] infohazard to tell the world and it's [54:51] all too dangerous to ever release. We [54:52] have to keep this locked up and we'll, [54:54] you know, in our ivory tower, we will [54:56] discover these wonderful things and [54:58] we'll share the fruits with the world, [54:59] but we'll have the AI and we'll control [55:00] it. And that sat very poorly with me. [55:04] And I thought then and I believe now [55:07] that it is extremely important that we [55:10] avoid that kind of power concentration [55:12] and that we build this for the world and [55:15] the world gets to use it in lots of [55:17] ways. Some of which will be good, not [55:18] all of which will be good. Um but that [55:20] the by enabling people to explore this [55:23] very wide opportunity space in front of [55:25] us. Messy at times though it will be and [55:28] obviously we'll put guard rails on it [55:29] for reasonable safety. um we will give [55:33] the world a gift, but the world will [55:34] give the world will build a much bigger [55:37] gift on top of it for all of us. And [55:39] that if you don't enable people with [55:40] this technology and if you tried to keep [55:42] it locked up, which again now this may [55:44] sound obvious, but this like was the [55:46] sort of rough consensus plan of people [55:49] working on it before we came along. Um [55:52] I think that would have been really bad. [55:54] I am a believer in entrepreneurship and [55:57] innovation and and that people do people [56:01] are mostly good and mostly do amazing [56:02] things with tools. Um so by uh I think [56:07] my single biggest contribution uh has [56:09] been will be whatever pushing for this [56:12] to be a democratized technology that [56:13] people get to use and build on. [56:15] >> Sam Alman thank you very much. [56:17] >> Thank you very much. [applause] [56:21] >> [music]