视频库 / NO.013
--:--
点击播放，这里会跟随视频显示当前句的中英字幕。
Alexandr Wang: “This is a Once-in-a-Civilization Opportunity”
节目发布 2026-07-29 · Y Combinator
本期追问 · 点击跳到视频对应位置
2:52 18 岁跳过既定路径去创业，代价与机会成本该如何权衡？29:22 当技术浪潮百年一遇，个人该押注速度还是押注深度？24:06 在信息爆炸的青年期，'从消防栓喝水'式学习真能塑造判断力吗？7:53 为什么说亲身进入一家公司，比从外部观察更能理解如何创造事物？
归入 Ⅴ·04 凭什么相信一件还不存在的事？ →EDITED TRANSCRIPT · 依据现场录音编译整理，可划线生成便签
编者按：本文整理自 Y Combinator 主办的一场公开对谈。对谈嘉宾 Alexandr Wang（汪滔）是 Scale AI 联合创始人，十九岁自 MIT 辍学创业，将公司打造成 AI 数据基础设施的领军企业，现负责 Meta 的超级智能实验室；主持人 Garry Tan（陈嘉兴）是 Y Combinator 总裁兼 CEO、Initialized Capital 联合创始人。台下坐着的，是数百位与他们当年一样年轻的创业者。本文依据现场录音编译整理，仅删去口语枝节，论述内容悉数保留。
从洛斯阿拉莫斯到 MIT
主持人： 我们在后台聊到，理解这场活动有个很好的角度：这间屋子里坐满了和我们当年一样的人，十八岁、二十岁，甚至还有几位十六岁的。我们直接跳进你的故事吧。你从小就聪明得出奇，打数学奥林匹克，那时候的 Alex 在想什么？感受到什么？是什么把你推上了这条路？
汪滔： 我在新墨西哥州的洛斯阿拉莫斯长大。这个地方现在因为电影《奥本海默》出了名，但在当时，它就是个不折不扣的偏远小镇。我记得自己参加了各种数学竞赛、计算机竞赛，心里知道想做很大的事，却完全不清楚该走哪条路，怎么走。
我有个朋友特别痴迷编程，高中毕业后去硅谷实习，第一份实习是在 Palantir。他对我影响很大。所以高中一毕业，我就来硅谷的 Quora 工作，用一个间隔年在那里干了一年，然后才去 MIT。算起来，我在 Quora 的时候十八岁，进 MIT 也是十八岁，创办 Scale 时十九岁。从十七岁到十九岁这段时间，我感觉自己一直在变，究竟想做什么，答案不断在换。我从身边的人那里学到太多东西，那两年简直是不停地对着消防栓喝水。
回头看，有两件事特别重要，我很推荐。第一是先去一家公司工作。从外面看，你根本不知道公司是怎么运转的，不知道真正做出一个东西是什么样，不知道反复打磨一个东西是什么样，也不知道一群人一起做决定是什么样。这段经历非常关键。第二是去 MIT 读书。学校给了我大量探索兴趣的空间。正是在 MIT，我第一次训练模型，摆弄那年刚发布的 TensorFlow，也是在那里，我最终想出了 Scale 的点子。在 MIT 读了一年之后，我申请了 YC。
当时能被录取，感觉像个奇迹。YC 对我的创业之路至关重要。它有一种很奇妙的混合气质：一方面非常支持你，当然希望你成功；另一方面又对你极其坦白，你犯蠢的时候会直接告诉你。我觉得这正是每个人在生活里都需要的东西。故事讲到这里，就是十九岁创办 Scale，后面的事大家都知道了。
Scale 的原点：数据是缺失的按钮
主持人： 你当时的合伙人是 Jared Friedman。而且你进 YC 时带的点子，跟后来的 Scale 完全是两码事。
汪滔： 是的。说来有趣，我们当时想做的是一个 AI 智能体，帮人获得医疗服务。这是一个典型的例子：点子本身终将成真，甚至现在我们已经看到了，帮人看病就医的 AI 智能体已经真实存在。但那时的时机是错的。
我们做了大概一两个月，Jared 把我们拉到一边说：「伙计们，我不确定这东西能走到哪儿去。」这正是我们需要听到的话。那时我已经把 AI 研究得透透的，训练过模型，于是我们回到起点，认真思考机会到底在哪里，然后想出了 Scale。
主持人： 当时卖数据这件事，大语言模型还远没有登场，但自动驾驶正在兴起，计算机视觉突然火了起来。那是你们的第一个市场，对吧？
汪滔： 对。事情是这样的：我在 MIT 做了一堆项目，训练各种各样的模型。跟今天比，那些都是小玩具。我记得要训练一个模型，需要三样东西：一个 GCP 账号，或者说某个云服务的账号，用来拿算力；一段训练模型的代码；还有数据，一个数据集。前两样东西，你在网上按一个按钮就能得到。唯独最后一样，数据，完全没有有效的获取途径。所以在我看来，未来的方向再明显不过：一定会出现一种方式，让你同样「按一个按钮」就拿到数据。
有意思的是，在此后的许多年里，Scale 最初的那些年，数据这个方向一直非常不性感。每次出去融资，哪怕我们的数字很好看，收入很扎实，风投和投资人总是满腹狐疑：「这算一门好生意吗？能持久吗？有护城河吗？」我当时觉得非常奇怪，不过话说回来，那些投资人没一个亲手训练过模型，所以他们大概真的不理解。快进到今天，我们总算融到了钱，坚持下来了，把生意做大了，而当年拒绝我们、对 AI 前景很悲观的那批投资人，如今正在写长文，论证数据多么关键，是 AI 领域最大的商业机会之一。
看着整件事绕了一圈回到原点，实在很有意思。
第一性原理与逆共识的信念
主持人： 这其实是第一性原理思考（first principles thinking）的一个绝佳案例。你不可能翻开《华尔街日报》，看到「数据很火」，然后决定去做这个。Scale 根本不可能那样起步。你必须从几条你确信为真的、关于世界的简单陈述出发，然后为它造点东西。
汪滔： 关键在于，你得对一组没有任何人认同的信念建立起确信。看看世界上最成功的那些公司，它们全都诞生于核心想法流行起来之前很久。创始人在默默无闻中苦干许多年，然后那个想法、那个领域、那门生意的概念才变成共识。你要想成功，唯一的办法就是比所有人更早地识别出这些关于世界的真相。
Scale 做 AI 已经十年了，最让人意外的一点是：你绝不能根据周围人的说法来做商业决策。跟着羊群走，你只会被搅得晕头转向，最后哪儿也去不了。所以你必须锻造出自己的指南针，判断未来会长什么样，因为其他人只会让你困惑。
主持人： 看起来你练就了一项了不起的本事。你从一个「我们相信 X，而没人相信」的内核出发，但接下来建公司的实际动作，是去跟投资人谈，说服他们，不让他们把你的士气打垮；去跟客户谈，他们本该一听就懂；尤其是，说服人来为你工作。
汪滔： 建公司早期的这些基本功，也许有人天生有点禀赋，但没有人在创业之初就擅长创业。我记得跟很多最早认识的投资人聊过，不少人会说：「你成长得太快了，变化太快了，我当时没看出来。」我想这对每一个创业者都成立：没有人擅长自己从没做过的事。所以对所有创业者来说，起点都是样样稀烂，整场游戏就在于你如何让自己持续进步、快速学习。
AI 时代创业：歌利亚对歌利亚
主持人： 我们在后台还聊到，现在是创业的幸运时刻。你显然可以来 YC，屋里的这些人彼此就是同伴，这已经很难得了。不止如此，现在你还有一个理想的个人 AI，随时告诉你「嘿，可以这样做」。你觉得这会让你当年加速得更快吗？在 AI 时代创业是什么样？
汪滔： 我真心认为，我们正处在一个惊人的时刻：瓶颈已经不在模型的进步，而在于把它扩散到世界的各个角落，帮助世界去适应这项已经存在的技术。哪怕模型从今天起完全停止进步，经济、世界运转方式，以及我们身边的一切，也还要经历几十年的天翻地覆。所以这大概是一个文明只有一次的机会（once-in-a-civilization opportunity）：做一个梦想家，拥有愿景和野心，通过建造非凡的东西，把你心目中未来世界的样子强加给现实。
我们在后台还聊到一点。我创办 Scale 那会儿，十年前，创业就是大卫对歌利亚。你得机灵，得找到切入市场的刁钻角度，得在资源少得多的情况下想办法竞争。而现在，有了智能体（agents）和更广义的 AI，我认为局面更接近歌利亚对歌利亚。创业公司或许是一个机甲歌利亚，被智能体和 AI 大幅增强；大公司则是传统意义上的歌利亚。
只要你真正拥抱 AI 智能体，想办法以最有野心的方式发挥它们的长处，创业公司完全可以轻松击败在位者。
Meta 的个人超级智能愿景
主持人： 我们聊聊超级智能吧，它甚至写进了你们实验室的名字里。在 Meta 内部，超级智能在操作层面意味着什么？
汪滔： 一年前，Mark 写过一份关于「个人超级智能」（personal superintelligence）的备忘录，我觉得这跟你提的「个人 AGI」的概念非常接近。我们相信，世界上的每一个人，几十亿人，都会拥有一个专属的超级智能，为他量身适配，帮他达成目标，了解他的处境，并最终扩展他自己的能动性（agency）。我们反复思考的核心正是能动性的扩展：如何帮助人们做成过去连做梦都不敢想的事？如果一切都变得轻而易举，世界上每个人会去做什么？
我们也从生态的角度来看这件事。Patrick 之前提到过，我们不相信那种由 AI 统治世界的总体化、集权式的图景。我们相信这些技术会去增强一个极其广阔的生态。所以我们相信全球几十亿人各自拥有个人超级智能，也相信创业会迎来一场大爆发。如今 Meta 平台上有两亿家企业，借助这波创造力的爆发和 AI 工具，我们认为这个数字应当涨到几十亿。最终，它会是一个充满活力的生态：企业智能体与个人智能体协作，构成一个由 AI 全面驱动的复杂系统。
归零重建前沿实验室的一年
主持人： 看到 Muse Spark 1.1 发布，我真的很兴奋。我的 OpenClaw 爱死它了。执掌一个前沿实验室是什么体验？Muse Spark 大致是 Opus 那个档次，后面还有什么在路上？而且我知道你们越来越重视开源，这间屋子里的人都很爱开源。
汪滔： 我在 Meta 差不多一年了，这一年很不寻常。我们公开谈过，Llama 4 没有走在 Meta 需要的轨道上。所以我进去以后，做的是一次归零重建（zero-based build）：如何从头搭起一整个前沿实验室，当然也大量利用了已有的东西，并且尽可能快地推进。
从那一刻起，九个月内我们发布了 Muse Spark 1，两个月后又发布了 Muse Image 和 Muse Spark 1.1。这个过程里有几件事让我印象很深。第一是人才密度极其重要，这是最核心的押注。人才密度会自然复利：你手里的顶尖人才越多，就有越多顶尖人才想加入你。第二，从内部看会很清楚，前沿 AI 工作本质上是研究，是科学工作。我们在探索这些模型能做什么，如何把它们推得更远，能力的边界到底在哪里。这需要一套与互联网公司、互联网产品截然不同的心态和运作模式，更多地围绕实验、科学和规模化展开。归根结底，一切都在于如何打造一个实验室、一套运作模式、一个系统，能够随着整个生态的指数级增长一起复利：能力的指数增长、算力的指数增长、采用和使用量的指数增长。在生态的几乎每一个维度上，我们都处在一条极其陡峭的指数曲线上。我把实验室看作一个有机体，它必须能跟着这条曲线一起生长，这一点至关重要。
这一年非常令人兴奋，而且我们还会推出很多东西。我们刚发布了 Muse Spark 1.1，这是一个很好的模型，Muse Spark 系列还会持续更新。更大的模型也在路上，我认为它们足以与当今最顶尖的模型正面竞争。我们很快会发布一个智能体框架（harness），已经打磨了一段时间，目的是赋能所有开发者和做智能体的开发者。另外，正如你提到的，我们也在做开源模型。我们相信一个去中心化的 AI 能力、进步与开发的世界，希望赋能更广的生态，让世界上每个人都能用这项技术去构建。所以，很多激动人心的东西在路上，我们想尽人力所能，最大程度地赋能生态和开发者。
低价开放与十倍浪潮论
主持人： 我用 Muse Spark 配 OpenClaw 的时候，很明显它已经和 Opus 一样好，尤其是那种带技能文件（skill files）的智能体流程，但价格便宜了八倍。
汪滔： 是的，这正是重点所在。我们不相信这样一个世界：模型贵到必须配给，只有最有钱的开发者和公司才用得起。每个人都能用上这项技术、用它建造任何想建造的东西，这一点很重要。
我们的看法是，最好的 AI 产品甚至还没有被发明出来。回顾整个 AI 生态的演进，每一波浪潮都比前一波大十倍。我创办 Scale 时，第一波大概是自动驾驶。自动驾驶非常酷，但跟大语言模型和聊天机器人一比就相形见绌，聊天机器人大概比自动驾驶大了十倍。几年后又出现了编程智能体，编程智能体大概又比聊天机器人大十倍。我们正处在这条陡峭的曲线上，会不断看到 AI 范式的新模态、新形态、新进展，每一波都远远大于上一波。所以我们的立场是：把生态放开，一起探索，一起构建世界的未来。
主持人： 那么，用 Muse Spark 的编程模型，最好的方式是什么？是 OpenCode 吧？
汪滔： 对，目前最简单的方式是用 OpenCode，官网上有接入指南。很快我们会有自己的框架。而最终，我们希望做出优秀的模型，能接入所有现成的框架，尽可能激发生态里的组合式创新。
主持人： 我知道框架还在保密。不过能不能透露一点？我现在还在用 OpenClaw，还在用 Hermes agent，我管这些东西叫「法拉利」，动不动就抛锚在路边。你们这台法拉利会不会抛锚？
汪滔： 希望不会。我们非常关注速度。任何用过这些工具的人都知道，速度大概是最关键的因素之一。其次是可靠性，正如你说的，我们希望它极其可靠。我们还希望它高度可扩展，能支撑你想要的任何复杂、有趣的多智能体架构。坦白说，框架之上还会有海量的创新空间：如何编排，如何设计循环，如何构建极其复杂的智能体协作生态。
我们希望它足够可扩展，最终就是让人们能驾驭（harness）这项技术。这个双关不是故意的。我真心相信，这些模型已经强大得不可思议，强大到足以推动 GDP 增长好几个百分点。剩下的，就靠有远见、有野心的聪明人去把它变成现实。
十年后回看：智能与能动性变充裕
主持人： 一个问题。等你回望这十年，你觉得人们会说，关于 AI 有哪件事在事后看来显而易见，而此刻大家正在错过？
汪滔： 如今大量的争论都围绕着：模型到底进步了多少？模型能不能跨过这道坎？超级智能什么时候来，两年还是五年？我们会不会撞墙？在我看来，这些争论在某种程度上有点浪费时间，因为强大的模型终将到来，这是必然的。我想，将来回头看，人们会说所有关于「到底哪一年实现」的争吵都是短视的，因为现实是，整个人类文明正处在一条惊人的指数曲线上。看看 AI 在过去十年的进展，任何人都会震撼得说不出话。十年前，最好的 AI 模型能做的，是在 YouTube 视频里认出猫。而现在，我们在跟一个数字之神对话，它能做的事情，我们都见识过一些，你只能目瞪口呆。这个趋势会继续下去，模型会越来越强。
所以十年后回头看，有一件事会变得显而易见：智能变得充裕了，能动性变得充裕了。眼下的趋势会一直延续。这会非常奇特。纵观人类历史，一群聪明人为了共同目标聚在一起，始终是进步的瓶颈。在某种意义上，美国就是这样的例子：一群非常聪明的人聚到一起，怀着一个想要实现的未来愿景，缔造了这个国家。美国几乎每一家公司的故事都是如此，每一家 YC 公司的故事也是如此。
而这将要改变。突然之间，稀缺资源不再是智能，也不再是能动性。我真的认为，稀缺的会是愿景和野心。你对未来的世界该是什么样，有没有清晰的看法？你想在哪一个方向上，把手指按上天平，让五到十年后的世界与今天不同？你有没有足够的野心和驱动力，去熬过所有的糟心事，把它变成现实？AI 会让这件事变得更容易，智能体和 AI 大概让它比十年前容易十倍、一百倍。但反过来说，你也因此可以把梦做得更大。
世界要完全拥抱这项技术，还有太多东西需要演进。坦白说，今天的世界几乎还没准备好迎接它。作为建造者，我们有责任让世界做好准备：帮助企业和政府适应这项新技术；想清楚如何从生物安全和网络安全的角度保卫世界；想清楚如何管控这项新技术带来的各种风险。而在另一面，这也是人类前所未有的机遇期。我们可以开创新的科学，可以攻克健康和生物学中长期悬而未决的难题，可以建立过去无法想象的企业，还有过去不可能存在的创作机会。这是一个机遇与风险并存的摇篮。对于一个有建造能力、又对世界该如何改变有坚定看法的人，没有比现在更好的时代了。
抽象层在变，系统思维不变
主持人： 你觉得路径变了吗？我看到一则消息，斯坦福的计算机专业人数下降了两位数的百分比，人们在担心，这在我看来简直荒唐。要造出那么强的智能体，你仍然需要这些技能。也许将来不需要了，我不确定。你会对台下的人说什么？这是他们真实面对的问题：该更「文科脑」（word cell）还是更「理工脑」（shape rotator）？该怎么办？这有没有改变你在 Meta 想招的人，以及你管理团队的方式？
汪滔： 系统化、严谨的思考仍然极其重要。抽象层在不断变化，我以前不相信事情会这样演变，但它确实这样发生了。想当年我创业的时候，我们是写代码的。现在嘛，我相信在座各位没人再写代码了，那太荒唐了。
但现在的问题变成了：怎样把智能体编排到一起？接下来会是：怎样构建智能体的组织？怎样让一百万个智能体协同工作？再往后是：怎样让一万亿个智能体协同工作？无论我们在哪一层抽象上运作，都始终需要想清楚工作流该如何构造。这就是那种严谨的系统化思考。在我创业的那个年代，传统路径是先写代码，然后建立由人组成的组织，再想办法把这些人组织好，这需要系统思维。而现在，更接近于先编排智能体，再想办法编排一支支智能体大军。但系统思维永远不会过时。所以，全押「文科脑」肯定是错的，你得会「转形状」。
不过我认为，未来更需要的东西，是对世界该如何发展怀有更深的指南针和哲学观。关于文明如何穿越这样的时期，人类历史上有很多很多教训。人类在未来十年发生的变化，很可能超过过去一百年。所以，为此拥有积极的愿景，并能连贯地阐述它该如何展开，比任何时候都更重要。
智能体循环：朴素的工程真相
主持人： 说得具体一点。我很好奇，在你的朋友圈或者 Meta 内部，有没有哪些 AI 应用可以拿出来讲讲，是近在眼前却还没被人想明白的？给我们一点内幕。
汪滔： 我认为智能体循环（agentic looping）里仍然蕴藏着天文数字级的机会：如何设计一套系统，让你在一个持续的反馈循环里，投入一千倍、一百万倍的 token 去推动一个结果。想想大多数公司的本质，公司就是一个大型反馈循环，每一条边上都由人来操作。公司获取客户，想办法让客户更满意；客户更满意就花更多钱；钱多了就能招更多人，这些人再去获取更多客户，再让客户更满意。某种意义上，这就是每一家创业公司、每一门生意的反馈循环。在这个大循环里还嵌着许多微循环。构建能运转并优化这些反馈循环的智能体系统，里面有巨大的超额收益。
我们在 Meta 内部已经看到过这样的案例：只要设计出正确的智能体循环，给智能体一个正确的评估（eval）或正确的优化指标，一群智能体就能轻松完成超过一百名工程师团队的工作，而且是非常轻松地。所以，想清楚一个充满智能体协调问题的世界会是什么样，是当今最有意思的问题之一。
主持人： 从机械层面讲，就是 markdown 文件、cron 定时任务，然后把智能体指向足够多的数据，让它自己搞明白一些也许不在分布内的东西？
汪滔： 对，机械层面就是先想清楚指标是什么，然后就是技能文件、markdown 文件、cron 定时任务。
主持人： 还有 /goal。
汪滔： 对，/goal。我总觉得很有趣：任何事情一旦真正拆开来看，都平淡得很。
主持人： 所以它不是魔法。有些人把它说得神乎其神，LinkedIn 上有些帖子讲得像魔法一样。
汪滔： 我的建议是，别看 LinkedIn。LinkedIn 是用来找客户的。
给 18 岁自己的两条建议
主持人： 我喜欢用这个问题收尾。你可以给十八岁的自己发一封电报，你会对那个人说什么？感谢你回来跟这些听众分享。你会往漂流瓶里塞一句什么话？
汪滔： 归结起来，就是锻造你自己的内在指南针，判断未来会如何发展，并对它抱有坚定的信念。因为你会被噪音淹没，各种人会对你说三道四，这会让你非常困惑，非常难熬。尤其在你年轻、没有经验的时候，要对自己的信念和想做的事真正抱有确信，会显得格外困难。但这是最重要的事。就像我们前面聊的，正是靠着对所做之事的深深确信，我们才熬过了多年的风暴，市场的混乱、行业的混乱、身边人的混乱。
第二条建议是：去找出世界上那条曲线最陡、又能延续最久的指数。几十年前，这条曲线是摩尔定律，那在当时显然是最该押注的东西。现在我认为是 AI 的进步，而未来还会出现更多这样陡峭的曲线。哪怕这些曲线的起点非常无聊，看起来甚至没什么意思，也没关系。我刚开始做 Scale 的时候，AI 能做的是在 YouTube 视频里检测猫，很难跟人解释这是我们时代最重要的技术，但它就处在一条难以置信的指数曲线上。
最后还有一件事必须说：Meta 很荣幸向在座每一位提供一千美元的 Muse Spark API 免费额度。我们会持续把模型做得更好，而且 Muse Spark 现在比 Opus 便宜八倍，换算成「Opus 美元」的话，那就多得多了。我们的团队会把领取额度的细节告知每一位，非常期待看到大家做出什么来。
主持人： 谢谢 Alexandr Wang。
排版 + 横图 + 来源，粘贴即成稿
章节 · 点击跳转视频
0:07
从洛斯阿拉莫斯到 MIT 的少年路
3:25
Scale 的原点：数据是缺失的按钮
6:19
第一性原理与逆共识的信念
9:05
AI 时代创业：歌利亚对歌利亚
11:12
Meta 的个人超级智能愿景
13:10
归零重建前沿实验室的一年
16:18
低价开放与十倍浪潮论
19:26
十年后回看：智能与能动性变充裕
23:39
抽象层在变，系统思维不变
26:50
智能体循环：朴素的工程真相
29:20
给 18 岁自己的两条建议
本期小问 · 档案清单
2:52
18 岁跳过既定路径去创业，代价与机会成本该如何权衡？
29:22
当技术浪潮百年一遇，个人该押注速度还是押注深度？
24:06
在信息爆炸的青年期，'从消防栓喝水'式学习真能塑造判断力吗？
7:53
为什么说亲身进入一家公司，比从外部观察更能理解如何创造事物？
本期讲者
亚历山大·王Scale AI 联合创始人，19 岁从 MIT 辍学创业，将其打造为 AI 数据基础设施领军公司。2025 年 Meta 入股 Scale 后出任 Meta 首席 AI 官，领导 Meta Superintelligence Labs。
Garry TanY Combinator 总裁兼 CEO，Initialized Capital 联合创始人，早年为 Palantir 早期员工、Posterous 创始人。
01从洛斯阿拉莫斯到 MIT 的少年路
0:07
All right, [music] full rockstar treatment for Alexander Wang, everyone. All right. >> [cheering] [applause] >> So, why don't we start out uh backstage we're saying, you know, one of the cool ways to think about this event is like you know, this room is actually full of people who are just like us, but when we were 18 or 20 or, you know, there's some 16-year-olds in this audience, you know. Let's jump to your story. I mean, you got it came up always really smart like math olympiad like jump us to, you know, the Alex of that time. Like, what were you feeling? What were you thinking? And what drove you down this road?
好了，[音乐] 大家一起给 Alexander Wang 一个摇滚巨星级的欢迎。好了。>> [欢呼] [掌声] >> 那么，我们不如从后台聊到的话题开始吧。你知道，看待这个活动的一个很酷的角度是，这个房间里坐着的其实都是和我们一样的人，只不过是我们18岁、20岁的时候，你知道，观众里还有一些16岁的孩子。我们直接说说你的故事吧。我是说，你一路走来一直非常聪明，比如数学奥赛之类的。带我们回到那个时候的 Alex。你当时是什么感受？在想些什么？是什么驱使你走上了这条路？
0:47
>> Yeah, I am uh Well, I grew up in New Mexico, Los Alamos, New Mexico, um which now Oppenheimer famous, but um it really was the middle of nowhere and uh I remember I did all these math competitions, all these um computer science competitions, but then um I knew I wanted to do really big things and it was like not exactly clear how or what the exact path to do that would be. Um and I had a friend who was really into programming um and, you know, after high school got an internship in the valley.
>> 是这样，我……嗯，我是在新墨西哥州长大的，新墨西哥州的洛斯阿拉莫斯，现在因为《奥本海默》出名了，但那里真的是荒郊野岭。我记得我参加了各种数学竞赛、各种计算机科学竞赛，但后来……我知道自己想做非常了不起的事情，但当时并不清楚具体该怎么做、走哪条路才行。嗯，我有个朋友非常痴迷编程，高中毕业后他就在硅谷找了份实习。
1:23
I think his first internship was at Palantir. And um and he, you know, he was kind of this um influence for me and so after I finished uh high school, I ended up working at Quora um here in Silicon Valley. And then um I worked there for a year. I took a gap year to work there um and then I went to MIT. Um and this is I was 19 when I worked at Quora, I was 18 when I went to MIT and I was 19 when I started um Scale. And I remember this period from like 17 to 19 it was um uh I felt like I was constantly changing like you know exactly what I want to do was constantly changing. You know, I was learning so much just from the people around me and it was just like I felt like I was drinking from the firehose pretty constantly during that time. Um and um I would definitely recommend you know the two things that were really important. One is I think working at a company was really valuable because like I think from the outside in you have no idea how companies work. You have no idea what it looks like to actually
我记得他第一份实习是在 Palantir。他对我算是一种影响，所以高中毕业后，我最后来到硅谷这边的 Quora 工作。我在那儿干了一年，我用了一个间隔年去那里工作，然后我去了 MIT。嗯，在 Quora 工作时我19岁，去 MIT 时我18岁，创办 Scale 时我19岁。我记得从17岁到19岁这段时期，我觉得自己一直在变，比如说，我到底想做什么这件事一直在变。我从身边的人身上学到了太多东西，那种感觉就像那段时间我一直在用消防水管喝水（信息量爆炸）。嗯，我肯定会推荐有两件事真的很重要。第一是我觉得去公司工作非常有价值，因为我觉得从外部看，你根本不知道公司是怎么运作的。你根本不知道真正做出一个东西
2:25
build something. You have no idea what it looks like to iterate on something. You have no idea what it looks like for groups of people to make decisions. And so I thought that was really important. And then going to school at MIT was actually really important because it just gave me a lot of um opportunity to explore what was interesting. And so it was at MIT that I started training my first models and um that I like played around with TensorFlow which had just come out that year at MIT and where I like ultimately came up with the idea of Scale. And then after one year at MIT I applied to YC.
是什么样子。你根本不知道对一个东西做迭代是什么样子。你根本不知道一群人一起做决策是什么样子。所以我觉得那非常重要。然后去 MIT 读书其实也非常重要，因为它给了我很多去探索什么才是有意思的机会。所以我就是在 MIT 训练了我的第一批模型，还玩了那年刚发布的 TensorFlow，也是在 MIT，我最终想出了Scale 这个点子。在 MIT 读了一年之后，我申请了 YC。
2:57
You know, it felt like kind of like a miracle to get in at that time. And uh and YC was was really critical to my entrepreneurial journey. Like I don't think um like YC is this amazing blend of uh you know, they're very supportive and they obviously want you to succeed, but they also give it to you very real and they tell you when you're being a dumbass um which I think is uh you know, that's what we all need in life. So um yeah, that was I think the story till then I was 19 started Scale and uh the rest is history.
当时能进去感觉简直像个奇迹。YC 对我的创业历程真的至关重要。我觉得，YC 是一种很了不起的结合体——他们非常支持你，显然也希望你成功，但他们也会跟你说大实话，你要是在犯傻他们会直接告诉你，我觉得这正是我们人生中都需要的。所以，嗯，我想这就是到那时为止的故事。然后我 19 岁创办了 Scale，之后的事大家都知道了。
02Scale 的原点：数据是缺失的按钮
3:25
>> I guess with uh you work with Jared Friedman at the time and um you came in with actually a very different idea than what ended up becoming Scale. >> Yeah, so we wanted to build um like an AI agent funnily enough for uh for doc for to help people like get medical care. Um and it was like the right it was a great example of an idea that I think will ultimately exist. Like I think we're even seeing it now. Like AI agents to help people get medical care are very real. But it was the wrong timing.
>> 我记得当时你是和 Jared Friedman 一起做的，而且你们最初带来的想法，跟后来变成 Scale 的东西其实很不一样。>> 是的，说来好笑，我们当时想做的其实是一个 AI agent，用来帮人们看病、获得医疗服务。这其实是一个很好的例子——我觉得这个想法最终一定会实现。就像我们现在已经看到了，帮人们获取医疗服务的 AI agent 已经非常真实了。但当时时机不对。
3:55
Um and uh and we worked on it for about a month or two before Jared pulled us aside and were like, "Guys, this is I don't know if this is going to go anywhere." >> [laughter] >> Um and uh and that's exactly what we needed to hear. And it was at that time when like, you know, where I had studied AI to my T. I had like trained models and we thought we sort of went back to the drawing board, thought deeply about where the opportunity was, and >> came up with Scale. >> I guess selling data at the time, you know, large language models were not even had had not really come to the fore yet. Um but self-driving cars were sort of coming up and and computer vision suddenly became So that was sort of the first market. Is that right?
我们做了大概一两个月，Jared 就把我们叫到一边说：“伙计们，我不确定这事儿能做成什么样。”>> [笑声] >> 而这正是我们当时需要听到的话。也正是那个时候，你知道，我把AI 研究得非常透，我训练过模型，我们就回到原点重新思考，认真琢磨机会到底在哪里，>> 然后就想出了 Scale。>> 我想那时候做数据买卖，你知道，大语言模型都还没真正出现，还没成气候。但自动驾驶汽车正开始兴起，计算机视觉突然火了起来。所以那是最初的第一个市场，对吧？
4:35
>> Yeah, so the the story here is that like I was when I was at MIT, I did a bunch of projects like train train models of various forms. And these were like, you know, by comparison today, they're like little toy models. And um and I remember to train a model, uh I needed three things. I needed a uh GCP account, like I needed an account on some cloud service to get compute. I needed um the code to run to actually train the model. And I needed data. I needed a data set. And uh for two out of these three things, you could just press a button online and get them. And then for the last one, for data, there was like no effective way to get data for training these training these models. Um and so it felt incredibly obvious that this was going to be the future, that there was going to be a way to um you know, press a button so to speak and get data. And uh it was very funny because in the years that followed, like in the first many years of Scale, data was very unsexy still. Um every time we would go
>> 对，故事是这样的：我在 MIT 的时候做过一堆项目，训练各种各样的模型。跟今天比起来，那些都算是很小的玩具模型。我记得要训练一个模型，我需要三样东西：一个 GCP 账号，也就是某个云服务的账号来拿到算力；需要能跑起来、真正用来训练模型的代码；还需要数据，我需要一个数据集。而这三样里有两样，你在网上点一下按钮就能拿到。但最后一样，也就是数据，根本没有什么有效的办法能拿到训练这些模型所需的数据。所以我当时觉得这件事再明显不过：这一定就是未来——一定会出现一种方式，可以说是“点一下按钮”就能拿到数据。有意思的是，在那之后的很多年里，也就是 Scale 最初的好些年，数据这件事一直很不性感。每次我们出去融资，哪怕我们的数据很漂亮、收入也很好，VC 和投资人
5:39
out to fundraise, even though our numbers were great and we had great revenue, you know, VCs and investors would always be very skeptical. They'd be like, "Oh, I don't know if this is a good business. Does it have longevity? Is it durable?" Um and uh it was really weird to me, but you know, none of the investors had ever trained a model. So, I guess they didn't really get it. Um and uh fast forward to today, you know, we we managed to raise money, we managed to keep going, managed to keep growing the business, but um the very same investors who passed on us and were um were very dour on the potential of AI are writing think pieces today about how data is so critical and is one of the biggest business opportunities um in AI.
总是非常怀疑。他们会说：“哦，我不知道这是不是一门好生意。它有长期性吗？它有护城河吗？”这让我觉得很奇怪，但你知道，那些投资人没有一个训练过模型，所以我猜他们是真的不懂。快进到今天，我们成功融到了钱，撑了下来，把业务一直做大，但当年拒绝我们、对 AI 的潜力非常悲观的那批投资人，今天却在写长文，讲数据有多关键、是 AI 里最大的商业机会之一。
03第一性原理与逆共识的信念
6:19
So, uh it's very funny to see that whole whole thing come full circle. >> I mean, it seems like that's actually a real good um case study in first principles thinking, right? Like, you can't start a company by opening the pages of the Wall Street Journal and saying, "Well, this data is hot. Like, we're going to go work on that." It's like, you literally couldn't have started Scale that way. You had to start from uh think like simple statements that are about the world that you know to be true and then sort of building something for that.
所以看到这一圈绕回来，真的挺好笑的。>> 我觉得这其实是第一性原理思考的一个非常好的案例，对吧？你没法靠翻《华尔街日报》，说“嗯，这个数据现在很火，那我们就去做这个”来创业。你根本不可能那样创办 Scale。你必须从你确信为真的、关于这个世界的一些简单判断出发，然后去为此构建东西。
6:49
>> Yeah, I think the the key thing is you need to develop conviction in a set of beliefs that nobody else um agrees with. Like, I think if you look at all the most successful companies in the world, um they were started at a time long before the sort of like core idea was popular. And they work on that. They toil in obscurity for years and years before, you know, the the idea or the space or the concept of the business, you know, becomes consensus. And the only way you're going to be successful is if you're able to identify these truths about the world early, long before everyone else. And I think the like I mean, one of the most surprising things like, you know, Scale, we've been working on AI for a decade. You know, you just you can't base your business decisions based on what everyone else is saying around you. Like, if you go too much with the herd, you will get immensely confused and you will end up nowhere. And so you have to develop your own compass of what you think the future is going to look like
>> 对，我觉得关键在于，你得对一套没有人认同的信念建立起坚定的信心。我觉得如果你看世界上最成功的那些公司，它们创办的时候，那个核心想法离流行还早得很。然后他们就埋头去做，在无人问津中苦干很多年，直到那个想法、那个领域、那个商业概念才变成共识。而你能成功的唯一途径，就是能够比所有人都早地识别出关于这个世界的这些真相，远远早于其他人。我觉得最让人意外的一点是，Scale 我们已经在 AI 上做了十年。你不能把自己的商业决策建立在周围所有人在说什么之上。如果你太跟着人群走，你会极度困惑，最后哪儿也去不了。所以你必须建立自己的指南针，形成你对未来长什么样的判断，因为其他所有人只会让你更迷茫。
7:51
because everyone else will just confuse you. >> It seems like one of the things you got incredibly great at was you know, you start with this kernel of like we believe X and nobody else believes it, but then the mechanics of building the business are talking to investors and convincing them and not letting them demoralize you, talking to customers who I mean should just get it and then especially like convincing people to come work for you. >> Yeah, I think that the the these early mechanics of building a company like the these are things that I think you might have some predisposition be good at, but like nobody is good at starting a company when they start a company.
>> 感觉你特别擅长的一点是：你从一个内核出发，就是“我们相信 X，别人都不信”，但接下来做公司的具体功夫，是去跟投资人谈、说服他们、不让他们打击你的士气，去跟本来就该懂的客户谈，然后尤其是说服别人来给你干活。>> 是的，我觉得创业早期这些具体功夫，有些人可能天生就比较擅长，但没有人在开始创业的时候就已经会创业。
8:31
And I remember talking to a lot of the investors who I met very early on and they you know, a lot of them would say like oh, like you know, you just grew so quickly and you changed so quickly and like I didn't you know, I didn't see it at the time. And I think that's probably true for literally everyone who starts a company like nobody is nobody is good at something they've never done before, right? And so I think for all entrepreneurs, you start out pretty shitty at everything and the whole game is how do you develop yourself to continuously improve to get better and learn quickly.
我记得跟很多很早期见过的投资人聊，他们中不少人会说：“哦，你们成长得太快了，你们变化太快了，我当时没看出来。”我觉得这大概对每一个创业的人都成立——没有人一开始就擅长自己从没做过的事，对吧？所以我觉得对所有创业者来说，你一开始在每件事上都挺糟糕的，整场游戏就在于你怎么让自己持续进步、变得更好、学得更快。
04AI 时代创业：歌利亚对歌利亚
9:05
>> Uh backstage we're talking about this is actually a really lucky time to start a company cuz you know, obviously you can come do YC, uh you're you know, the people in this room have each other, which is kind of wild, but not only that, now you have a ideal personal AI that's going to tell you, you know, hey, these are some ways to do it. Um Do you think that would have helped you like accelerate even faster? Like you know, talk What do you think it's like to start a company today with with AI in the age of AI?
>> 我们在后台聊到，现在其实是创业非常幸运的时候，因为显然你可以来做 YC，而且在座各位彼此就是资源，这已经挺疯狂了；不仅如此，现在你还有一个理想的个人 AI，会告诉你“嘿，这里有几种做法”。你觉得这会让你当年加速得更快吗？你觉得在今天、在 AI 时代创业是什么感觉？
9:36
>> Yeah. I mean I really think I think we're at this like in amazing moment in the world where the bottleneck is not the progress of the AI models, the bottleneck is diffusing that through the rest of the world and and helping the world adapt to this amazing technology that already exists. Like I think if the models didn't improve at all from today, there would still be like decades and decades of like total upheaval and change in the economy and how the world operates and and everything around us and um you know, so I think it's as a result, it's like one of the most incredible it's probably a like once in a civilization opportunity to be a dreamer and to have a vision and to have ambition and to impose a view of how the future world should look by building something amazing. Um you know, one of the things that we were we were chatting about um uh you know, backstage is you know, when when I started Scale or you know, 10 years ago, if you start a company, you had to be um you know, it was like David
>> 嗯。我真的觉得我们正处在一个非常了不起的时刻：瓶颈已经不是 AI 模型的进展，瓶颈是把它扩散到世界的其他部分，帮助世界适应这项已经存在的惊人技术。我觉得就算模型从今天起完全不再进步，仍然会有几十年、几十年的彻底颠覆和变革，发生在经济、世界的运转方式，以及我们身边的一切上。所以结果就是，我觉得这大概是最不可思议的、可能是文明级别、只此一次的机会——去做一个梦想家，去拥有愿景、拥有野心，通过构建了不起的东西，把你对未来世界应该是什么样的看法落地。你知道，我们在后台聊到的一件事是：当我创办 Scale 的时候，也就是十年前，如果你要创业，那基本上是大卫对歌利亚，你必须很聪明，
10:41
versus Goliath and you had to be clever and you had to find like an angle into the market and you had to sort of like, you know, figure out um a way to compete even though you had much fewer resources. And now I actually think with the power of agents um and AI broadly speaking, it's much closer to Goliath versus Goliath. Like I think but maybe the startup is like a Mecca Goliath that is like vastly enhanced by the power of agents and AI and you know, the the large companies are the sort of like more traditional Goliath, so to speak.
必须找到切入市场的角度，必须想办法在资源少得多的情况下去竞争。而现在我其实觉得，有了agent 和广义上的 AI 的力量，这更接近于歌利亚对歌利亚。只不过创业公司像是一个被 agent 和 AI 极大增强的“超级歌利亚”，而那些大公司则是更传统意义上的歌利亚。
05Meta 的个人超级智能愿景
11:12
But I think that startups now like if you properly embrace AI agents and um figure out the way to leverage their strengths in the most like ambitious ways, you can easily outcompete incumbents. >> So, let's talk about super intelligence because that's clearly that's even in the name of your lab. Um, what does super intelligence mean operationally inside Meta right now? >> Yeah, I think that you know, we a year ago Mark wrote this um, memo about personal super intelligence, which I think actually is very similar to your concept of personal AGI. But, you know, we believe that everybody in the world, you know, all the billions of people in the world are going to have a super intelligence that is adapted and tailored to them, that is enables them to accomplish their goals, knows their context, and ultimately is an expander of their own agency. Like I think the thing that we think a lot about is is agency expansion. How do we help people accomplish things that they couldn't have ever dreamed of before? And what
但我觉得现在的创业公司，如果你真正拥抱 AI agent，想清楚怎么以最有野心的方式发挥它们的长处，你完全可以轻松打赢在位巨头。>> 那我们聊聊超级智能吧，因为这显然就写在你们实验室的名字里。在 Meta 内部，超级智能在实操层面到底意味着什么？>> 嗯，我觉得，一年前 Mark 写了一篇关于个人超级智能的备忘录，我觉得这其实跟你说的个人 AGI 的概念很像。但我们相信世界上每一个人，全球几十亿人，都会拥有一个为他们量身适配的超级智能，能帮他们达成目标、了解他们的处境，并最终成为他们自身能动性的放大器。我们思考得很多的一点就是能动性的扩展：我们怎么帮人们做到他们以前想都不敢想的事？还有，
12:15
would everyone in the world do if everything was just easy? Um, and we think about this in a in an ecosystem way um, as well. I think uh, you know, Patrick mentioned it, but you know, we don't believe in this totalizing, you know, totalitarian view of, you know, AIs that control the world. We believe that these are going to enhance this very broad ecosystem. And so, you know, we believe in billions of people all around the world all having their own personal super intelligence. And we also believe in, you know, an explosion of entrepreneurship. There's 200 million businesses that uh, are on Meta's platforms today. We think that number should go to billions with this explosion of of creativity and using AI tools. And ultimately we think that, you know, it's going to be this like dynamic ecosystem of business agents working with, you know, personal agents and developing this sort of like uh, complex ecosystem that is fully AI supercharged.
如果一切都变得很容易，世界上每个人会去做什么？我们也是从生态系统的角度来思考这件事的。我觉得，Patrick 提到过，我们并不相信那种笼罩一切的、极权式的图景——由 AI 来控制世界。我们相信这些东西会增强一个非常广阔的生态系统。所以我们相信全世界几十亿人都会拥有自己的个人超级智能。我们也相信创业会迎来一场爆发。今天 Meta 的平台上有 2 亿个商家，我们认为这个数字会随着创造力的爆发和 AI 工具的使用增长到几十亿。最终我们认为，这会成为一个动态的生态系统：商业 agent 与个人 agent 相互协作，发展出一个被 AI 全面加持的复杂生态。
06归零重建前沿实验室的一年
13:10
>> So, I was really psyched to see Meta Spark uh, 1.1. My my open claw absolutely loved it. Um, how you know, how how has running a frontier lab been? Um, you know, the Meta Spark level is sort of the opus level. Uh, what's coming down the pipe? And also I think that you're uh, you're increasingly looking at open source which uh, I think this audience really loves. >> Yeah, yeah. So, I think it was um, it's been you know, I've been at Meta for about a year now and it's been um, quite a year. I think uh, you know, getting in and um, you know, Meta we we've talked about it publicly like Llama 4 wasn't on the trajectory that was needed for um, for Meta and so I got in there and we kind of did a zero-based build of how do you um, you know, build an entire frontier lab uh, you know, in some ways kind of from scratch obviously using a lot of what we had um, and move as quickly as possible.
>> 我看到 Meta Spark 1.1 的时候真的特别兴奋。我的 open claw 简直爱死它了。那么，运营一个前沿实验室是什么体验？Meta Spark 这个级别大概相当于 opus 级别。接下来会有什么？另外我觉得你们也越来越重视开源，这一点在座的听众真的很喜欢。>> 是的，是的。我到 Meta 差不多一年了，这一年相当不平凡。刚进去的时候，Meta 我们也公开讲过，Llama 4 并没有走在 Meta 所需要的轨道上，所以我进去之后，基本上是做了一次“归零重建”：怎么去搭建一个完整的前沿实验室，某种意义上算是从头开始，当然也大量用上了我们已有的东西，并且尽可能快地推进。
14:05
And so within nine months of that moment we launched new Spark 1 and then uh, two months later we launched new image and new Spark 1.1 and um, you know, there's a few things that I think have really struck me about this. It you know, the first is talent density was incredibly important. That was the the core thing to bet on and um, like talent density is something that compounds naturally. Like the more talented people you have the more of the most most talented people want to join you. Um, you know, in and I think it's it's kind of um, amazing to see on the inside but you know, frontier AI work is research. Like we are it is scientific work. We're exploring what can you do with these models? How can you push these models?
于是在那之后的九个月内我们发布了 Spark 1，两个月后又发布了新的图像模型和Spark 1.1。在这个过程中，有几件事让我印象特别深。第一是人才密度极其重要。那是最核心、值得押注的东西。人才密度是会自然复利的：你拥有越多有才华的人，就有越多最顶尖的人想加入你。我觉得从内部看到这一点挺震撼的，但前沿 AI 的工作本质上是研究。这是科学工作。我们在探索：用这些模型能做什么？怎么把这些模型往前推？
14:48
What is the what are the reaches of what can be accomplished with these models which requires a totally different mindset and operating model than you know, existed for internet companies or internet products and what not. There's a lot more about experimentation, about science, about scaling and everything ultimately is about how do you develop a lab, an operating model, a system that will just um, be able to compound with all of the exponential growth that will happen in the ecosystem. Both the exponential growth in capabilities, the exponential growth in compute, um, the exponential growth in adoption and usage. Like these are all um, we are on this like very, very steep exponent across maybe every dimension of the ecosystem. And um, it's important to develop like a like an organism. That's how I think about the lab that that's able to sort of grow with that. Um, you know, it's been it's been very exciting and we're we're going to be shipping a lot more. So, um, I think uh, you know,
这些模型所能达到的边界在哪里？这需要一种完全不同的心态和运作模式，跟互联网公司、互联网产品那一套完全不一样。这里更多是关于实验、关于科学、关于规模化，而所有事情最终都归结为：你怎么打造一个实验室、一套运作模式、一个系统，能够跟着生态里所有的指数级增长一起复利。既包括能力上的指数增长、算力上的指数增长，也包括采用率和使用量的指数增长。我们几乎在这个生态的每一个维度上，都处在一条极其陡峭的指数曲线上。所以重要的是要打造出一个像有机体一样的东西——我是这么看这个实验室的——它能跟着一起生长。这一切非常令人兴奋，我们接下来还会发布还有更多。所以，嗯，我想，你知道，我们刚刚发布了 Muse Spark 1.1，
15:42
we will we just launched Muse Spark 1.1, which was a great model. We're going to continue to have updates on the Muse Spark line. Um, we're also have bigger models on the way that I think will be uh, much more competitive with even the very best models that are out there today. We're going to be launching a harness um, soon and have been working on a harness to help empower all the developers and agentic developers out there. Um, and uh, and then we're also um, you know, as you mentioned, we're working on open-source models. And we want to kind of as I described before, like, we believe in a decentralized world of AI capability and progress and development.
这是一个很棒的模型。我们会继续更新 Muse Spark 这条产品线。嗯，我们也有更大的模型正在路上，我觉得它们会更有竞争力，甚至能跟当今市面上最顶尖的模型一较高下。我们很快会推出一个 harness（智能体框架），我们一直在做这个 harness，希望能赋能所有的开发者，尤其是做智能体的开发者。嗯，然后，就像你提到的，我们也在做开源模型。我们想要的，就像我之前说的，我们相信 AI 的能力、进步和发展应该是去中心化的。
07低价开放与十倍浪潮论
16:18
Like, we want to we want to empower the broader ecosystem and everyone in the world to be able to build and develop using this technology. And so, um, we have a lot of exciting things on the way and I think we want to be um, we want to empower the ecosystem and developers as much as humanly possible. >> I mean, it sounds like one of the ways, I mean, certainly when I was using uh, Muse Spark with my Open Claw, like, it it became clear that it was as good as Opus, especially for that sort of agentic flow with skill files, but it was like eight x cheaper, actually.
就是说，我们想赋能更广泛的生态，让世界上的每个人都能用这项技术去构建、去开发。所以，嗯，我们有很多激动人心的东西正在路上，我觉得我们想要，嗯，我们想尽最大可能去赋能整个生态和开发者。>> 我是说，听起来其中一种方式，我是说，反正我自己在用Muse Spark 配 Open Claw 的时候，很明显它的表现和 Opus 一样好，特别是那种带 skill 文件的智能体流程，但它便宜了大概八倍，真的。
16:50
>> [laughter] >> Yes. Well, we I think this this goes to it. Like, I don't, you know, we don't believe in a world where these models are so expensive that, you know, they get rationed only for the most wealthy of developers and and companies. Um, it's important for everyone to be able to use the technology to um, and to build whatever they want to build with it. And I think that, you know, we take a view I think the best AI products haven't even been developed yet. You know, that if you look at the AI ecosystem and everything that's happened, like every wave is 10 times bigger than the past wave. So, you know, when I started scale, the first wave was maybe self-driving cars. Self-driving cars are really awesome. They're like really, really cool, but that was like pales in comparison to large language models and chatbots. And like, you know, chatbots became this thing that was like probably 10 times bigger even than um than uh you know, self-driving cars. And then there were coding agents which came a few
>> [笑] >> 是的。嗯，我觉得这正是关键。就是，我们不希望看到这样一个世界：这些模型贵到只能被配给给最有钱的开发者和公司。嗯，重要的是让每个人都能用上这项技术，用它去构建他们想构建的任何东西。我觉得，我们的看法是，最好的 AI 产品还根本没被做出来呢。你知道，如果你看看整个 AI 生态和已经发生的一切，每一波浪潮都比上一波大十倍。所以，你知道，我创办 Scale 的时候，第一波大概是自动驾驶。自动驾驶真的很了不起，真的、真的很酷，但跟大语言模型和聊天机器人一比就相形见绌了。你知道，聊天机器人后来变成了一个可能比自动驾驶还大十倍的东西。然后过了几年又出现了编程智能体。编程智能体大概又比聊天机器人大十倍。我觉得我们正处在一条
17:45
years later. And coding agents are probably 10 times bigger than than um chatbots. And I think we're just on this steep curve. Like, we're going to keep seeing these new modalities and form factors and developments of the AI paradigm that will each be dramatically bigger than the last. And so, um you know, our point of view is like, let's let's unleash the ecosystem. Let's explore and let's see um let's build, you know, kind of the future of the world together. >> So, what's the best way to actually take advantage of the coding model uh from U Spark? It's it's open code, right?
陡峭的曲线上。我们会不断看到 AI 范式的新模态、新形态和新进展，每一个都会比上一个大得多。所以，嗯，我们的观点是，让我们把生态释放出来。让我们去探索，去看看，去构建，嗯，一起去打造世界的未来。>> 那么，要真正用好 Muse Spark 的编程模型，最好的方式是什么？是用 open code，对吧？
18:16
>> Yeah. Today, um the the easiest way is to use open code. We have like onboarding on the website. And then uh soon we'll have a harness of our own. And um ultimately, I think we want great models that plug into all of the available harnesses and empower as much, you know, uh sort of combinatorial innovation in the ecosystem as possible. >> Yeah, I know the harness is uh you know, under wraps still. But like, can you tease us with you know, I mean, I still use open claw. I still use Hermes agent. You know, it's uh you know, these things are I call them Ferraris that break down on the side of the road all the time. Like, is this a Ferrari that won't break down?
>> 是的。目前最简单的方式就是用 open code。我们网站上有引导流程。然后，嗯，很快我们会有自己的 harness。而且，嗯，归根结底，我觉得我们想要的是能接入所有现有 harness 的优秀模型，尽可能激发生态里那种组合式的创新。>> 嗯，我知道这个 harness 目前还保密。但你能不能给我们透露一点？我是说，我现在还在用 open claw，还在用 Hermes agent。你知道，这些东西——我管它们叫法拉利，动不动就抛锚在路边。那这会是一辆不会抛锚的法拉利吗？
18:49
Like, you know, tease us a little bit. >> Yeah, hopefully hopefully it doesn't it doesn't break down. I mean, I think we're really focused on speed. I think speed is um you know, for anyone that uses these tools, speed is probably the you know, one of the most critical things. I think also reliability, like you mentioned, we want to be extremely reliable. Um we want to be very extensible and to scale to as complex and interesting of a multi-agent setup that you that you want to have. Like I think there's so much innovation that will occur even above the harness, frankly, um in terms of like how to orchestrate and set up loops and and develop like, you know, very complex ecosystems of these agents working together.
来，稍微透露一点。>> 是的，希望它不会抛锚。我觉得我们真的很关注速度。我觉得速度，你知道，对任何用这些工具的人来说，速度大概是最关键的因素之一。还有可靠性，就像你说的，我们希望做到极其可靠。嗯，我们希望它有很强的可扩展性，能撑起你想要的任何复杂、有意思的多智能体架构。我觉得在 harness 之上还会有非常多的创新，说实话，比如怎么去编排、怎么设置循环，怎么开发出非常复杂的智能体协作生态。
08十年后回看：智能与能动性变充裕
19:26
Um Uh we want to be really extensible and and um ultimately we want to just empower people to harness this technology because harness that Oh, >> [laughter] >> uh no pun actually pun not intended, but um but there's like I truly believe these these models are already just incredibly powerful. Like they should they should be so powerful to fuel, you know, um many many points of expansion of GDP growth and I think it's like up to smart people with vision and ambition to make all that happen. >> Let's see. So, one question. I mean, when you look back on the decade, um what do you think they'll say was obvious in hindsight about AI that people are just missing in real time right now?
嗯，我们希望非常可扩展，而且，嗯，归根结底我们就是想赋能大家去驾驭（harness）这项技术，因为 harness 那个……哦，>> [笑] >> 呃，这不是双关——其实真不是有意的双关，不过，我是真的相信这些模型已经强大得不可思议了。它们应该强大到足以驱动很多很多个百分点的 GDP增长，我觉得这就取决于那些有远见、有野心的聪明人去把它实现出来。>> 我想想。有个问题。当十年后回头看这个十年，你觉得关于 AI，人们会说哪些事其实是显而易见的，只是当下大家都没看到？
20:13
>> You know, so much of the debate that happens these days is around oh, how good are the models actually getting and can the models actually bridge this issue and, you know, when are we going to get super intelligence? Is that in like 2 years or 5 years? And, you know, are we going to hit a wall? And, you know, so much of that debate is like I think um in some ways uh a little bit of a waste of time because, you know, I think it's inevitable that we're going to have very powerful models and um you know, rather than I think we'll look back and say, "Oh, all this arguing around like when exactly it was going to happen was sort of um was short-sighted because the reality is we are just as a entire human civilization on this incredible exponential. Like you cannot look at the progress of AI over the past decade and not just be totally awestruck by how far it's come. Like a decade ago, the best AI models could recognize cats in YouTube videos. And now, you know, we're talking to um you know, a digital god that can, you
>> 你知道，现在很多争论都围绕着：模型到底变得有多好？模型能不能跨过这个坎？超级智能什么时候到？是两年还是五年？我们会不会撞墙？你知道，那些争论在我看来某种程度上有点浪费时间，因为我觉得我们迟早会拥有非常强大的模型。与其那样，我觉得我们回头看时会说：“哦，当初争论它到底具体什么时候发生，其实挺短视的。”因为现实是，我们整个人类文明正处在一条不可思议的指数曲线上。你看过去十年 AI 的进展，不可能不为它走了多远而彻底震撼。十年前，最好的 AI 模型只能识别YouTube 视频里的猫。而现在，我们在跟一个数字之神对话，它能……
21:19
know, uh I mean, we've all seen some of the hacks and some of the some of the things these systems are capable of. And you just can't help but be awestruck. And and I think this trend will just continue. Like these these models are going to become more and more powerful. And so, I think a decade looking back, it'll it'll be obvious that intelligence became abundant and that agency became abundant. Like the current trends we're on are just going to keep continuing. And um this will be very strange. I mean, I think for the history of humanity, um you know, groups of smart people getting together towards a shared goal was was the bottleneck of progress. You know, US The United States of America in some sense was an example of this. Like the United States of America was formed from a smart group of very smart people getting together and having a vision for the future that they wanted to enact.
呃，我是说，我们都见过一些黑客案例，见过这些系统能做到的一些事情。你真的没法不感到震撼。而且我觉得这个趋势只会延续下去。这些模型会变得越来越强大。所以我觉得十年后回头看，会很明显：智能变得极其充裕，能动性（agency）也变得极其充裕。我们现在所处的这些趋势会一直延续下去。而且，嗯，这会非常奇怪。我是说，纵观人类历史，一群聪明人聚在一起朝着共同目标努力，一直是进步的瓶颈。你知道，美利坚合众国在某种意义上就是个例子。美国就是由一群非常聪明的人聚在一起，对未来有一个愿景并想把它实现，从而建立起来的。
22:08
And that's the story of nearly every company um in America. And it's the story of every YC company. Um and that's going to change. Like all of a sudden, the scarce resource isn't going to be intelligence or agency. I really think it's going to be vision and ambition. It's like, do you have a clear view of what you want the world to look like in the future? What is the like one way in which you want to put your finger on the scale for how the future of the world will develop and how the how the world will look like in 5 to 10 years that it does not look like today? And do you have the ambition and drive to like go through all the crap to make that happen? And AI will make that easier.
这几乎也是美国每一家公司的故事，是每一家 YC 公司的故事。嗯，而这一点将会改变。突然之间，稀缺资源不再是智能或能动性了。我真的觉得会变成愿景和野心。就是说，你对未来的世界该是什么样子，有没有清晰的看法？你想在哪一件事上施加你的影响力，去改变世界未来的走向，让五到十年后的世界跟今天不一样？你有没有那种野心和干劲，愿意去扛过所有那些烂事把它做成？而 AI 会让这变得更容易。
22:48
Like agents in AI makes that maybe 10 times or 100 times easier than it was a decade ago. But the flip side of that is then you all of a sudden you can dream bigger. Like I think And the world is like um you know, there's so many things that need to evolve for us to be able to fully embrace this technology. Um you know, the world is like really just, you know, barely even ready for this technology today. And I think, you know, as a builder, we have a responsibility to prepare the world, right? Like we have to help enterprises and governments, you know, to adapt to this new technology. We have to help figure out how we secure the world from a biosecurity perspective or cybersecurity perspective. We have to figure out how we um how we're going to to manage all these risks that we see with this new technology. But on the flip side, it's also the time of like, you know, unprecedented opportunity for humans.
智能体和 AI 大概让这件事比十年前容易了十倍甚至一百倍。但反过来说，突然之间你就可以有更大的梦想。我觉得……而这个世界，嗯，有太多东西需要演进，我们才能真正完全拥抱这项技术。嗯，你知道，这个世界今天其实几乎还没准备好接受这项技术。我觉得，作为构建者，我们有责任让世界做好准备，对吧？我们得帮助企业和政府去适应这项新技术。我们得想明白怎么在生物安全和网络安全的层面保护好这个世界。我们得想清楚怎么去管理这项新技术带来的所有这些风险。但另一方面，这也是一个对人类来说机会空前的时代。
09抽象层在变，系统思维不变
23:39
Like we can develop new sciences. We can solve problems in health and biology that have been forever unsolved. We can build new businesses that you couldn't have even imagined before. There's like new creative opportunities that couldn't have existed before. So, it's like it's just this incredible cradle of of opportunity and risks that uh that I think makes it like no better time to be someone who's a builder and um and has a strong view of how the world should change. >> Do you think the path has changed? I mean, one of the things I saw, I think Stanford uh the amount of computer science majors actually dropped by some double-digit percentage. It's people sort of worried like, which is sort of insane to me. Like you still sort of need those skills to even create agents that are that good. Maybe that won't be true. I'm not really sure.
我们可以发展新的科学。我们可以解决健康和生物学领域那些长期无解的问题。我们可以创办以前根本想象不到的新企业。会有以前不可能存在的全新创造性机会。所以这就是一个充满机遇与风险的、不可思议的摇篮，我觉得这让当下成为最好的时代——如果你是一个构建者，并且对世界该如何改变有强烈的看法。>> 你觉得路径变了吗？我看到的一个现象是，好像斯坦福的计算机专业人数下降了两位数的百分比。大家有点担心，这在我看来挺离谱的。你还是需要那些技能，才能造出那么好的智能体。也许以后不是这样，我也说不准。
24:25
How you know, have you changed you know, what do you what would you say to people in this audience right now? Like this is sort of a real question that people are sort of facing. Like should they become more word cell and less shape rotator? Like what you know, what's the move? And you has that changed um the kind of people you're looking to hire and, you know, how you manage your teams right now at Meta? >> I think systematic and rigorous thinking are still incredibly important because you know, the abstraction layer, I mean, I didn't used to believe that this is how it was going to play out, but it really has, like, the abstraction layer just keeps changing. So, you know, when I started a company back in my day, we wrote code. Um >> [laughter] >> And now, you know, I'm sure nobody here writes code anymore. That's ridiculous.
那，你有没有改变……你会对现场的这些人说什么？这是大家真实面对的问题。他们该变得更“文字型”（word cell）、少一点“空间型”（shape rotator）吗？该怎么办？还有，这有没有改变你想招的人，以及你现在在 Meta管理团队的方式？>> 我觉得系统化、严谨的思考仍然极其重要，因为，你知道，抽象层——我是说，我以前并不相信事情会这样发展，但它真的就是这样：抽象层一直在变。你知道，我当年创业的时候，我们是写代码的。嗯 >> [笑] >> 而现在，我相信在座已经没人写代码了。那也太离谱了。
25:09
But um but now it's about how do you orchestrate the agents together? And then it's like, how do you develop these organizations of agents? Like, how do you get like a million agents to work together well? And then it'll be, how do you get like a trillion agents to work together well? Like, I think that there's going to be this continued um uh need to figure out how you structure uh workflows at the abstraction layer that we're going to be operating at. And that form of like rigorous systematic thinking, I mean, traditionally the way this would work like in my era of starting companies is you would start by writing code, and then you would have organizations of humans, and you'd figure out how you how to organize those humans. Um and that requires systems thinking. And now, maybe it's like much more much closer to first you you orchestrate the agent, then you figure out how to orchestrate like these armies of agents. But um but I think systems thinking is never going to go out of style. So, I think
但，嗯，现在的问题是：你怎么把这些智能体编排在一起？然后是：你怎么构建这些智能体组织？比如，怎么让一百万个智能体协同得很好？再往后就是，怎么让一万亿个智能体协同得很好？我觉得会一直存在这种需求：在我们所处的抽象层上去搞清楚怎么组织工作流。而这种严谨的系统化思考——我是说，传统上，在我创业的那个年代，你会先从写代码开始，然后你会有一个人的组织，你得想明白怎么去组织这些人。嗯，这需要系统思维。而现在，可能更接近于：先编排智能体，然后再想怎么去编排这些智能体大军。但我觉得系统思维永远不会过时。所以我觉得
26:01
it's definitely a mistake to go all in on Word Cell. Like, I think you need to you need to shape rotate. Um but then I think the sort of like um much more of I think what's necessary going in the future is having um a deeper sort of compass and philosophical view on how the world should develop. Because I think there are there are many many lessons um from human history around um how how we think civilization go through this period. And um So, you know, humanity will change more in the next decade than it has in the past 100 years, probably. And um And so, I think like the imperative for us to have positive visions for that and have coherent uh articulations of how that should develop are are more important than ever.
全押在“文字型”上肯定是个错误。我觉得你还是得会“转形状”。嗯，不过我觉得未来更需要的，是拥有一种更深层的指南针，一种关于世界该如何发展的哲学观。因为我觉得人类历史里有非常非常多的教训，关于文明如何度过这样的时期。嗯，所以，你知道，人类在未来十年的变化，可能会超过过去一百年。嗯，所以我觉得，我们要有关于这一切的正面愿景，并能连贯地表达出它该如何发展——这比任何时候都更重要。
10智能体循环：朴素的工程真相
26:50
>> Um Let's get a little more concrete. I mean, one of the things I'm curious about is like, are there sort of applications of AI that you're seeing among your friends or internal to Meta that you can talk about that are, you know, they're sort of obvious near-term maybe people haven't figured out yet. I mean, give us some alpha. >> [laughter] >> Um, I mean, I think there's still just like astronomical opportunity in uh agentic looping and and figuring out how you develop systems that enable you to spend like 1,000 x more or 1 million x more on tokens to drive an outcome in a in a continuous feedback loop. Like if you think about most companies, companies are just these like large-scale feedback loops where humans are operating each of the edges. Like, you know, companies they um, they get customers and they figure out to make those customers happier. And if customers are happier, then they spend more. And if they spend more, then they can hire more people who can then go figure out how to get more customers and
>> 嗯，我们说得具体一点。我很好奇的一件事是：在你的朋友当中，或者在 Meta 内部，你有没有看到一些可以聊的 AI 应用，就是那种近期很明显、但大家可能还没想到的？给我们点内部消息吧。>> [笑] >> 嗯，我觉得在智能体循环（agentic looping）这块还有极其巨大的机会，就是搞清楚怎么去设计那种系统，让你能在一个持续的反馈循环里，把一千倍甚至一百万倍的 token花出去来推动一个结果。你想想大多数公司，公司其实就是一个个大规模的反馈循环，每条边上都由人在运作。比如，公司获取客户，然后想办法让客户更满意。客户更满意了，就会花更多钱。花的钱多了，就能雇更多人，这些人再去想办法获取更多客户、
27:47
make those customers happier. And that's like this, you know, that in some sense is the feedback loop of uh of every startup or every business. And, you know, these within that there are micro feedback loops that exist. And I think developing agentic systems that can operate and optimize these feedback loops is there's like just huge amounts of of alpha there. Like I think we've seen internally at Meta um cases where if you can develop the right agentic loop and you have the right eval or the right metric for the agents to optimize, you can have a swarm of agents accomplish more than like a team of 100 engineers in, you know, uh very very handily actually, very very easily. And so, I think figuring out what the world um looks like with lots of uh sort of this like um these agentic coordination problems, I think that is like one of the most interesting problems today. So, mechanically speaking, I mean, markdown files, cron jobs, is I mean, is it that's and then basically pointing the agent at enough data so that it can
让客户更满意。这在某种意义上就是每一家创业公司、每一门生意的反馈循环。而且，你知道，在这之内还存在着一些微观的反馈循环。我觉得开发能够运转并优化这些反馈循环的智能体系统，这里面有巨大的超额收益空间。我觉得我们在 Meta 内部已经看到一些案例：如果你能设计出正确的智能体循环，并且有合适的评估标准或者合适的指标让智能体去优化，那么一群智能体能完成的事情比一个 100 人的工程团队还多，而且是非常轻松地、非常容易地做到。所以我觉得，搞清楚这个世界在充满这类智能体协同问题时会是什么样子，我觉得这是当今最有意思的问题之一。所以说，从机制上讲，我是说，markdown 文件、cron 定时任务，就是……就是这些，然后基本上就是把足够多的数据喂给智能体，让它能琢磨出一些
28:52
figure something out that, you know, maybe isn't in distribution. >> Yeah, I think figuring out Yeah, yeah, mechanically figuring out what the metric is and then yeah, it just comes down to skills, markdown files, cron jobs, >> /goal. >> Yeah, /goal. Like I think I think it's always funny how mundane everything is once you really dig into it. But um >> So it's not magic, you know what I mean? Some some people put a lot of magic There's like some LinkedIn threads about some magic stuff. >> advice, ignore LinkedIn.
也许不在分布内的东西。>> 是的，我觉得搞清楚……对，对，从机制上说就是搞清楚指标是什么，然后，是的，最后无非就归结为 skills、markdown 文件、cron 定时任务，>> 还有 /goal。>> 对，/goal。我觉得，我一直觉得挺好玩的一点是，一旦你真正深入进去，会发现一切都特别平平无奇。不过，嗯——>> 所以这并不是什么魔法，你懂我意思吧？有些人把它说得特别神。LinkedIn 上有些帖子把这事儿讲得神乎其神。>> 建议是，别理 LinkedIn。
11给 18 岁自己的两条建议
29:20
LinkedIn is where you get customers. >> [laughter] >> So I like to end on this, which is um you know, you get a telegram to send to the 18-year-old version of yourself, you know, what do you say to that person right now given all, you know, I mean, thank you for coming back and sharing your wisdom with this audience. I mean, you know, what would you send in a message in a bottle to the 18-year-old version of yourself right now? >> Yeah, I think the I think it really boils down to develop your own internal compass for how you think the future will develop and have strong conviction in it because, you know, you will get so you will get inundated with noise and people telling you and like you'll could be very confusing and it'll be very hard. And especially when you're young and you don't have experience like it can feel very difficult to um have true conviction in what you believe and and what you want to do. But I think that's the most important thing. Kind of as we talked about, you know, um
LinkedIn 是你获客的地方。>> [笑声] >> 我喜欢用这个问题收尾：就是，假如你能给 18 岁的自己发一封电报，考虑到你经历的这一切，你现在会对那个人说什么？我是说，谢谢你回来跟这些听众分享你的智慧。我是说，你现在会往漂流瓶里放一条什么样的话，寄给 18 岁的自己？>> 嗯，我觉得，归根结底就是：建立你自己内在的指南针，去判断未来会怎么发展，并且对它抱有强烈的信念。因为，你知道，你会被大量噪音淹没，会有人不停地跟你说这说那，那种感觉会非常令人困惑，会非常难熬。尤其当你还年轻、没有经验的时候，要对自己相信的东西、对自己想做的事真正抱有信念，会让人觉得非常难。但我觉得这是最重要的事。就像我们刚才聊到的，嗯，
30:24
it took a deep deep conviction in what we were building to be able to weather the the sort of storms of many years of um of uh chaos in the market, in the industry, in the people around us. And so um And and then the other piece of advice I would have is try to identify what is the what is the exponential in the world that has both the steepest curve and will go the longest. And you know many decades ago this curve was was Moore's law and that probably was you know that was at the time like clearly the right thing to invest on. I think right now it's AI progress but there will be more of these very steep curves in the future and it's fine if these curves start you know the starting point is very boring or like it doesn't even seem that interesting. Like it you know when we started when I started working on scale you know we had cat detectors in YouTube videos and that felt you know it's hard to say explain the story that that's like the most important technology of our time but it was on
正是因为对我们所构建的东西有极深的信念，我们才能扛过很多年的风浪——市场的混乱、行业的混乱、身边人的混乱。所以，嗯，另外我想给的一条建议是：试着去识别，这个世界上哪一条指数曲线既最陡峭、又能持续最久。你知道，几十年前这条曲线是摩尔定律，而那在当时显然是值得押注的正确方向。我觉得现在是 AI的进展，但未来还会出现更多这样非常陡峭的曲线。而且没关系，哪怕这些曲线一开始的起点非常无聊，或者看上去根本没那么有意思。就像我们刚开始的时候，我开始做 Scale 的时候，我们做的是在 YouTube 视频里检测猫，那感觉，你很难讲出一个故事说这是我们这个时代最重要的技术，但它就是处在一条不可思议的指数曲线上。
31:24
just this like unbelievable exponential. Um And I think I have one last thing I got to say yes which is we are meta is proud to offer everyone in this room a thousand dollars of free credits for the new Spark API. Fantastic. [applause] And uh And we're going to keep making the models better and right now new Spark is I think eight x cheaper than Opus so so if you convert that to Opus dollars uh >> [laughter] >> it's a lot more but no everyone here will will work to get everyone the details on how to get how to get these credits and we're really excited to see what everyone builds.
嗯，我觉得我还有最后一件事必须得说——对，就是：我们 Meta 很荣幸地为在座每一位提供价值一千美元的新版 Spark API 免费额度。太棒了。[掌声]还有，呃，我们会继续把模型做得更好，而现在新的 Spark 我觉得比 Opus 便宜八倍，所以如果你把它换算成 Opus 的价格，呃 >> [笑声] >> 那就多得多了。不过，在座的各位我们都会想办法把如何领取这些额度的细节告诉大家，我们非常期待看到大家用它做出什么东西。
32:06
Alexander Wang everyone.
有请 Alexander Wang，掌声送给他。
视频总结 · 一句话概括与核心要点
一句话概括
Alexandr Wang 回顾自己 19 岁创办 Scale 的经历，并以 Meta 超级智能实验室负责人的身份提出：AI 进步已是不可逆的指数曲线，真正的瓶颈已从"智能"转移到"扩散与愿景"，当下是"文明级一次性"的创业窗口。
核心要点
- 独立判断与逆共识信念是创业的起点：Wang 在 MIT 训练模型时发现算力、代码可"一键获取"，唯独训练数据没有渠道，由此创立 Scale。此后多年数据被视为"不性感"的生意，VC 反复质疑其持久性（因为他们从未训练过模型），如今同一批投资人却在写"数据是 AI 最大机会"的文章——他的结论是：不能靠《华尔街日报》选赛道，必须拥有自己的"内部指南针"，否则会被外界噪音彻底搞晕。
- 早期路径的两个关键经历：高中后 gap year 在 Quora 工作一年，学到公司如何运转、迭代、集体决策；随后在 MIT 一年接触刚发布的 TensorFlow 并产生 Scale 的想法。原本 YC 项目是"帮人获取医疗服务的 AI 代理"，做了一两个月后被 Jared Friedman 直接叫停——方向正确但时机太早，转而做数据（首个市场是自动驾驶的计算机视觉标注）。
- 没人天生会创业，核心能力是快速学习：他坦言所有创始人起步时"什么都做得很烂"，游戏规则是持续自我提升；YC 的价值在于既支持你、又会直白地告诉你"你在犯傻"。
- 当前瓶颈不是模型进步，而是扩散：他断言即便模型从今天起停止改进，经济与社会仍要经历数十年的剧变；这使得现在成为"文明级一次性"的机会，让有愿景和野心的人去塑造未来。
- 创业格局从"大卫 vs 歌利亚"变成"歌利亚 vs 歌利亚"：十年前初创公司必须靠巧妙的切入角度以弱胜强；如今借助 AI 代理，初创公司可以成为"机甲歌利亚"，若能最大程度利用代理的优势，能轻松击败传统巨头。
- Meta 的"个人超级智能"愿景是去中心化的：源自扎克伯格一年前的备忘录，目标是让全球数十亿人各自拥有了解自身上下文、扩展"个人能动性"的超级智能；反对"AI 统治世界"的极权式想象，希望 Meta 平台上的 2 亿家企业增长到数十亿，形成"商业代理 × 个人代理"的生态。
- Meta 前沿实验室的"零基重建"：入职约一年，因 Llama 4 未达预期，他几乎从零重建实验室，9 个月内发布 Muse Spark 1，两个月后再发布图像模型和 Spark 1.1。关键经验：人才密度是自然复利的核心押注；前沿 AI 本质是科学研究，需要与互联网产品完全不同的实验、扩展型运作模式，实验室要像"有机体"一样随能力、算力、采用率的多重指数增长而生长。
- 路线图与开源承诺：将持续更新 Muse Spark 系列，推出与顶级模型抗衡的更大模型，发布强调速度、可靠性和多代理可扩展性的自研 harness，并推进开源模型；当前推荐通过 open code 接入。主持人称 Spark 在代理式工作流上与 Opus 相当但便宜约 8 倍，Wang 表示不希望模型贵到只有富裕开发者能用。
- 每一波 AI 浪潮都比上一波大 10 倍：自动驾驶 → 聊天机器人 → 编码代理，每波都是上一波的 10 倍规模，且"最好的 AI 产品还没被发明出来"。
- 近期最大的 alpha 在"代理式循环"：公司本质是人在各节点操作的大规模反馈回路（获客 → 满意度 → 消费 → 招聘 → 获客）。若能定义正确的评估指标并搭建代理循环，让系统愿意多花 1000 倍甚至 100 万倍 token 换取结果，Meta 内部已见到"一群代理轻松胜过 100 名工程师团队"的案例。机制上其实很平凡：skills、markdown 文件、cron job、/goal——"别看 LinkedIn，LinkedIn 只用来找客户"。
- 稀缺资源从智能与能动性转向愿景与野心：十年后回看，"AI 何时到达超级智能、是否撞墙"的争论会显得短视；智能和能动性将变得丰裕，历史上"聪明人聚在一起完成目标"这一进步瓶颈（美国建国、每家 YC 公司皆如此）将被打破，真正稀缺的是对未来 5-10 年世界应该是什么样的清晰主张，以及熬过所有烂事的驱动力。
结论与值得注意的细节
- 给 18 岁自己的建议：一是建立并坚定自己对未来的内部指南针，年轻缺经验时尤其难，但这是穿越多年市场混乱的唯一凭借；二是找到"斜率最陡且持续最久的指数曲线"——过去是摩尔定律，现在是 AI 进步——即使起点看起来很无聊（如"识别 YouTube 里的猫"），也要押注。
- 对"CS 专业人数下降"的回应：系统性、严谨的思维永不过时。抽象层不断上移——从写代码，到编排代理，到组织百万乃至万亿代理协作——都需要"形状旋转者"式的系统思维，全押"文字细胞"是错误的；同时更需要一种关于世界应如何发展的哲学罗盘，因为未来十年人类的变化可能超过过去 100 年。
- 建设者的责任：帮助企业和政府适应技术，解决生物安全与网络安全等风险，同时抓住新科学、健康与生物学难题、全新商业与创意形式的机会。
- 现场福利：Meta 向在场每人赠送 1000 美元 Muse Spark API 额度，并强调 Spark 价格约为 Opus 的 1/8。
核心句型 · 9
1. You have no idea what it looks like to …
“You have no idea what it looks like to iterate on something.”
强调「局外人不了解某事的真实样貌」，可连用三次形成排比增强语势。仿写：You have no idea what it looks like to ship a product.
2. It was the right idea but the wrong timing.
“It was a great example of an idea that I think will ultimately exist … But it was the wrong timing.”
区分「是否正确」与「何时正确」的对照结构，适合复盘决策。可扩展为 right X, wrong Y 的平行短语。
3. The bottleneck is not A, the bottleneck is B.
“The bottleneck is not the progress of the AI models, the bottleneck is diffusing that through the rest of the world”
先否定常见归因、再给出真正的限制因素，重复主语加强对比。适用于分析型论述。
4. Even if X didn't … at all, there would still be …
“If the models didn't improve at all from today, there would still be like decades and decades of … change”
虚拟条件下的「最坏情形仍成立」论证法，用来证明结论的稳健性。
5. The scarce resource isn't going to be A. It's going to be B.
“The scarce resource isn't going to be intelligence or agency. I really think it's going to be vision and ambition.”
预测性对比句，先破后立；配合 I really think 表达个人强判断而不显武断。
6. It's much closer to X than Y.
“It's much closer to Goliath versus Goliath.”
用「更接近于」来修正一个既有类比，而不是全盘否定。适合描述范式的渐变。
7. … is never going to go out of style.
“Systems thinking is never going to go out of style.”
表达「某种能力/原则永不过时」，口语化且有力，常用于给建议的收束句。
8. It boils down to …
“I think it really boils down to develop your own internal compass”
把复杂内容归结为一句核心，常用于总结、回答「最重要的是什么」。
9. It's always funny how … once you really dig into it.
“It's always funny how mundane everything is once you really dig into it.”
表达「深入后发现真相并不神秘」的反差感，语气轻松，适合去神秘化的表达。
词汇精讲 · 107 · 按出现顺序
rockstar treatment
phr.
0:07
摇滚巨星级的待遇（隆重欢迎）
the middle of nowhere
phr.
0:47
荒郊野岭，偏远之地
gap year
/ˈɡæp jɪr/
n.
1:23
间隔年（学业之间休学一年）
drinking from the firehose
phr.
1:23
信息量大到接收不过来（消防水管喝水）
iterate on
/ˈɪtəreɪt/
phr. v.
2:25
对……反复迭代改进
entrepreneurial
/ˌɑːntrəprəˈnɜːriəl/
adj.
2:57
创业的，企业家的
give it to you very real
phr.
2:57
直言不讳，跟你说大实话（口语）
the rest is history
phr.
2:57
后面的事众所周知
funnily enough
phr.
3:25
说来好笑，有趣的是
pulled us aside
phr. v.
3:55
把我们拉到一边（私下谈话）
to my T
phr.
3:55
精确、透彻地（通常作 to a T）
back to the drawing board
phr.
3:55
回到原点重新规划
come to the fore
phr.
3:55
崭露头角，成为焦点
unsexy
/ʌnˈseksi/
adj.
4:35
不吸引人的，不时髦的（口语）
skeptical
/ˈskeptɪkl/
adj.
5:39
怀疑的
longevity
/lɔːnˈdʒevəti/
n.
5:39
持久性，寿命
durable
/ˈdʊrəbl/
adj.
5:39
耐久的，（商业）可持续的
passed on us
phr. v.
5:39
（投资人）放弃投资我们
dour
/dʊr/
adj.
5:39
阴郁的，悲观的
think pieces
n.
5:39
深度评论文章
come full circle
phr.
6:19
绕了一圈回到原点
first principles thinking
phr.
6:19
第一性原理思考
conviction
/kənˈvɪkʃn/
n.
6:49
坚定的信念
toil in obscurity
phr.
6:49
默默无闻地苦干
consensus
/kənˈsensəs/
n.
6:49
共识
go with the herd
phr.
6:49
随大流
compass
/ˈkʌmpəs/
n.
6:49
指南针；（喻）判断准则
kernel
/ˈkɜːrnl/
n.
7:51
核心，内核
demoralize
/dɪˈmɔːrəlaɪz/
v.
7:51
打击士气
predisposition
/ˌpriːdɪspəˈzɪʃn/
n.
7:51
倾向，天生禀赋
bottleneck
/ˈbɑːtlnek/
n.
9:36
瓶颈
diffusing
/dɪˈfjuːzɪŋ/
v.
9:36
扩散，传播
upheaval
/ʌpˈhiːvl/
n.
9:36
剧变，动荡
once in a civilization
phr.
9:36
文明级别仅此一次的（仿 once-in-a-lifetime）
impose a view
phr.
9:36
施加/推行一种看法
David versus Goliath
phr.
10:41
以弱对强（圣经典故）
an angle into the market
phr.
10:41
切入市场的角度
outcompete
/ˌaʊtkəmˈpiːt/
v.
11:12
在竞争中胜过
incumbents
/ɪnˈkʌmbənts/
n.
11:12
在位者，市场现有巨头
operationally
/ˌɑːpəˈreɪʃənəli/
adv.
11:12
在实操/运营层面
tailored
/ˈteɪlərd/
adj.
11:12
量身定制的
agency
/ˈeɪdʒənsi/
n.
11:12
能动性，自主行动的能力
totalizing
/ˈtoʊtəlaɪzɪŋ/
adj.
12:15
笼罩一切的，总体化的
totalitarian
/toʊˌtæləˈteriən/
adj.
12:15
极权主义的
supercharged
/ˈsuːpərtʃɑːrdʒd/
adj.
12:15
被大幅增强的
psyched
/saɪkt/
adj.
13:10
兴奋的，激动的（口语）
frontier lab
n.
13:10
前沿 AI 实验室
down the pipe
phr.
13:10
即将推出的
trajectory
/trəˈdʒektəri/
n.
13:10
轨迹，发展路线
zero-based build
n.
13:10
归零重建（不以既有结构为前提）
talent density
n.
14:05
人才密度
compounds
/kəmˈpaʊndz/
v.
14:05
复利式增长
the reaches of
phr.
14:48
……的边界/极限范围
operating model
n.
14:48
运作模式
exponent
/ɪkˈspoʊnənt/
n.
14:48
指数；此处指指数曲线
harness
/ˈhɑːrnɪs/
n. / v.
15:42
智能体框架/脚手架；（动词）驾驭、利用
agentic
/eɪˈdʒentɪk/
adj.
15:42
智能体式的，具自主行动能力的
decentralized
/ˌdiːˈsentrəlaɪzd/
adj.
15:42
去中心化的
as humanly possible
phr.
16:18
在人力所能及的最大限度内
rationed
/ˈræʃənd/
v.
16:50
配给，限量供应
pales in comparison to
phr.
16:50
与……相比相形见绌
modalities
/moʊˈdælətiz/
n.
17:45
模态（文本、图像等形式）
form factors
n.
17:45
产品形态
unleash
/ʌnˈliːʃ/
v.
17:45
释放，解放
onboarding
/ˈɑːnbɔːrdɪŋ/
n.
18:16
新用户引导流程
combinatorial
/ˌkɑːmbənəˈtɔːriəl/
adj.
18:16
组合式的
under wraps
phr.
18:16
保密中
tease
/tiːz/
v.
18:16
（此处）预告、透露一点
extensible
/ɪkˈstensəbl/
adj.
18:49
可扩展的
orchestrate
/ˈɔːrkɪstreɪt/
v.
18:49
编排，统筹
pun not intended
phr.
19:26
并非有意双关
in hindsight
/ˈhaɪndsaɪt/
phr.
19:26
事后看来
hit a wall
phr.
20:13
撞墙，遇到瓶颈
short-sighted
/ˌʃɔːrt ˈsaɪtɪd/
adj.
20:13
短视的
awestruck
/ˈɔːstrʌk/
adj.
20:13
惊叹的，震撼的
can't help but
phr.
21:19
不由得，忍不住
abundant
/əˈbʌndənt/
adj.
21:19
充裕的
enact
/ɪˈnækt/
v.
21:19
实施，把（愿景）变为现实
scarce resource
n.
22:08
稀缺资源
put your finger on the scale
phr.
22:08
施加影响、左右天平（原义带贬义，此处中性）
go through all the crap
phr.
22:08
扛过所有糟心事（口语粗俗）
the flip side
phr.
22:48
另一面，反面
biosecurity
/ˌbaɪoʊsɪˈkjʊrəti/
n.
22:48
生物安全
unprecedented
/ʌnˈpresɪdentɪd/
adj.
22:48
空前的
cradle
/ˈkreɪdl/
n.
23:39
摇篮，发源地
double-digit
/ˌdʌbl ˈdɪdʒɪt/
adj.
23:39
两位数的
word cell
n.
24:25
（网络梗）擅长语言/叙事的人
shape rotator
n.
24:25
（网络梗）擅长空间/数学推理的人
abstraction layer
n.
24:25
抽象层
rigorous
/ˈrɪɡərəs/
adj.
24:25
严谨的
go out of style
phr.
25:09
过时
go all in on
phr.
26:01
全押在……上
imperative
/ɪmˈperətɪv/
n.
26:01
必要之事，当务之急
coherent
/koʊˈhɪrənt/
adj.
26:01
连贯的，条理清晰的
articulations
/ɑːrˌtɪkjuˈleɪʃnz/
n.
26:01
清晰的表述
alpha
/ˈælfə/
n.
26:50
（金融/圈内语）超额收益；内幕优势
astronomical
/ˌæstrəˈnɑːmɪkl/
adj.
26:50
天文数字的，极大的
feedback loop
n.
26:50
反馈循环
eval
/ɪˈvæl/
n.
27:47
（AI 行业）评估基准
swarm
/swɔːrm/
n.
27:47
（智能体）群，蜂群
handily
/ˈhændɪli/
adv.
27:47
轻松地，毫不费力地
in distribution
phr.
28:52
（机器学习）在训练分布之内
mundane
/mʌnˈdeɪn/
adj.
28:52
平淡无奇的
boils down to
phr. v.
29:20
归结为
inundated with
/ˈɪnʌndeɪtɪd/
phr.
29:20
被……淹没
weather the storms
phr.
30:24
经受住风暴，渡过难关
Moore's law
n.
30:24
摩尔定律（芯片晶体管数约每两年翻倍）
理解自测 · 11 题
1. Wang 认为训练一个模型需要哪三样东西？其中哪一样在当时无法「按钮化」获得，这如何催生了 Scale？
三样东西是：云服务账号（算力）、训练代码、数据集。前两样在 2016 年已经可以在线一键获得（GCP 等云服务、TensorFlow 等框架），唯独数据没有任何高效获取途径。Wang 在 MIT 训练模型时亲身感到这一缺口，于是判断「一键获取数据」必然是未来，这就是 Scale 的原点。这段出现在「Scale 的原点」章节（第 7 段），也是他后来讲「第一性原理」的具体案例。
2. Scale 最初在 YC 时做的是什么产品？为什么被放弃？
最初做的是一个帮人们获取医疗服务的 AI agent。做了一两个月后，YC 合伙人 Jared Friedman 把他们叫到一边说「不确定这能做成什么」，团队于是回到原点重新思考。Wang 的评价是「想法本身正确，最终一定会存在（今天已经出现），但时机错了」。这段在第 5～6 段，体现他区分「是否成立」和「何时成立」的思维方式。
3. Wang 给 18 岁自己的两条建议分别是什么？
第一条：建立自己关于未来如何发展的内在指南针，并对它抱有强烈信念，因为年轻时会被噪音淹没，尤其缺乏经验时最难坚持自己的判断。第二条：找出世界上「最陡峭且持续最久」的那条指数曲线——几十年前是摩尔定律，现在是 AI 进展——哪怕它起点很无聊（比如当年的 YouTube 猫检测器）也要押注。这两条在最后章节（第 39～40 段）给出。
4. Wang 说 Meta 内部一群 agent 能超过 100 人工程团队，需要满足什么前提？
前提有两个：设计出正确的智能体循环（agentic loop），以及有合适的评估标准或指标（eval/metric）供 agent 去优化。他把公司看作由人操作各条边的大规模反馈循环，agent 系统的价值在于接管并优化这些循环。随后主持人追问机制，Wang 承认落到实处无非是 skills、markdown 文件、cron 定时任务和 /goal 命令，「一切都很平淡」。这在「智能体循环」章节（第 36～38 段）。
5. 为什么 Wang 认为「模型什么时候达到超级智能」的争论是浪费时间？他的论证链是什么？
他的核心判断是瓶颈不在模型进步，而在把已有技术扩散到世界各处。论证链：①即使模型从今天起完全不进步，经济和社会仍有几十年的重构空间；②过去十年从「识别猫」到「数字之神」的进展证明我们处在一条不可逆的指数曲线上，强大模型是必然的；③因此争论「两年还是五年」是短视，真正该做的是去构建和适应。这一论证分布在第 14 段和第 28 段两处，前者从需求侧、后者从供给侧共同支撑。
6. 「大卫对歌利亚」变为「歌利亚对歌利亚」的比喻，想说明创业条件发生了什么变化？
十年前创业公司资源远少于巨头，必须靠聪明、找市场切入角度以弱胜强（大卫对歌利亚）。现在有了 agent 和 AI，初创公司可以被极大增强，成为「超级歌利亚」，与传统大公司势均力敌甚至胜出。含义是：资源劣势被 AI 抹平，创业者只要充分、有野心地利用 agent，就能直接正面击败在位者。这一比喻出现在第 14～16 段，是他回答主持人「AI 时代创业是什么感觉」的核心。
7. Wang 为何主张模型要便宜、要开源？这与他的「生态观」有什么关系？
他明确反对模型贵到「只能配给最富有的开发者和公司」，认为每个人都应该能用这项技术去构建。这与他反对「单一 AI 控制世界」的极权式图景、主张去中心化 AI 发展一脉相承：Meta 设想数十亿人各有个人超级智能、平台上商家从 2 亿增长到数十亿，商业 agent 与个人 agent 交织成生态。而「最好的 AI 产品还没被做出来」「每波浪潮大十倍」的判断，也意味着只有开放生态才能探索出下一波。见第 17、21～24 段。
8. 从「智能变充裕」推出「稀缺的是愿景与野心」，中间的推理是什么？
Wang 先给出历史观察：人类历史上，「一群聪明人聚在一起朝共同目标努力」一直是进步的瓶颈，美国建国和每家 YC 公司都是这样诞生的。AI 让智能和能动性变得充裕后，这一瓶颈消失，agent 把「把想法做成」的难度降低 10～100 倍。于是限制因素就转移到了前端：你是否对未来世界有清晰看法、是否有野心去扛过一切把它实现。因此稀缺资源变成愿景和野心。推理在第 29～31 段完成。
9. 面对「CS 专业人数下降，学生该转向文科吗」的问题，Wang 的立场和理由是什么？
他的立场是「全押在 word cell（文字型）上肯定是错的，你仍然需要会 shape rotate（严谨的系统化思考）」。理由是抽象层在不断上移：过去是写代码、再组织人；现在是编排 agent、再编排百万乃至万亿 agent 的组织，这本质上仍是系统思维问题，永不过时。但他同时补充：未来更需要的是关于「世界应该如何发展」的哲学指南针，因为未来十年的变化可能超过过去一百年。见第 33～35 段。
10. 有人会反驳：「Wang 说争论 AI 何时撞墙是浪费时间，恰恰因为他的利益在于让大家相信 AI 必然强大。」Wang 会如何回应？这个反驳有多大力度？
Wang 大概会回应：他的论点并不依赖模型继续进步——第 14 段明确说「即使模型今天停止进步，仍有几十年的变革」，即结论对「撞墙」情形是稳健的；同时他会指出自己十年前在数据「不性感」时就押注 AI，并非现在才立场先行。反驳的力度在于：Meta 大量投入算力和挖人，确实依赖「必然更强」的叙事，且「扩散需要几十年」本身也是一个可争议的经验判断。因此这个反驳能削弱其对未来模型的乐观预期，但难以推翻「现有能力尚未被充分利用」这一核心命题。
11. 「找到最陡峭且最持久的指数曲线」这条建议，放到一个非技术领域（如医疗、教育）的创业者身上还成立吗？
基本成立，但需要转译。Wang 的建议本质是「押注结构性趋势而非当下热度」，并承认曲线起点可以很无聊（猫检测器）。对医疗或教育创业者而言，对应的可能是「AI 扩散到该行业的速度」——恰好是他所说的真正瓶颈所在。他自己最早的 YC 项目就是医疗 agent，只是时机太早；按他的逻辑，现在正是该曲线变陡的时刻。局限在于：非技术领域受监管、信任等非指数因素约束更强，曲线的「持续性」更难判断，所以这条建议在这些领域更像方向指引而非选题公式。
下载便签
手机：长按图片也可保存
← 上一期 · NO.012Do LLMs Understand? AI Pioneer Yann LeCun Spars with DeepMind’s Adam Brown.
下一期 · NO.014 →Terence Tao at IMO 2024: AI and Mathematics
订阅苏菲周报
每周一封：本周入库的精读、一个值得带走的问题、一条苏菲按。免费，随时退订。
免费 · 每周一封 · 一键退订
苏菲拉底
THE SOPHIE LAB · ASK THE BEST MINDS THE BIG QUESTIONS
内容仅供学习 · thesophielab.com