打造 Muse：ChatGPT 之后增长最快的 AI 产品｜Alexandr Wang、Meta
Muse 源于 Meta Superintelligence Labs 在 2025 年 6 月提出的“个人超级智能”命题：打造强大的 AI，通过帮助人们追求所想，而不只是自动化工作，来“提升人类体验”。 Wang 认为，Meta 在这条消费级未来赛道上拥有独特优势：Instagram、WhatsApp 和 Facebook 已经掌握人们的兴趣与关系网络，公司日活用户达到 35亿。由此，战略押注的核心是个人生活，而不是工作、编程或企业级 AI。
真正的催化剂出现在 2026 年初，当时 Opus 4.5 和 OpenClaw 让 Wang 与 Nat Friedman 看到了个人智能体可能成为的样子。 Friedman 的智能体通过监控安防录像，确认他是否喝了足够的水；Wang 则让自己的智能体调用邮件、照片和连续的心理探测，完成了一次仿佛“3年心理治疗”的过程。这些“介于恐惧与狂喜之间”的体验，让 Wang 认为超级智能体在很多意义上就是“终极消费级产品”，团队也在 1到2周内做出了原型。
Meta 随后花了 7个月，把一个偶尔能创造魔法的原型打磨成可靠的消费级产品；Senra 认为，这大致相当于传统科技行业的10年。 团队列出了100多项必需行为，为每一项建立评测，即便还有20行未达标也拒绝发布。首个全面通过的检查点来自专门训练的 Musepark 1.3，体现了 Wang 的判断：产品体验对模型质量、沟通方式以及何时向用户寻求指导极其敏感。
Muse 由不到200人打造，因为 Meta 把它当成一件“完整的艺术作品”，而不是委员会拼装出的功能合集。 Wang 说，大公司产品往往会变成各路产品经理议程冲突后的“Cronenberg 融合体”或“奶昔”；Muse 则坚持单一而完整的产品观点，其中 Friedman 的品位尤其关键。Zuckerberg 的贡献是耐心：他容忍了7个月的质疑，也接受了华尔街认为 Meta 可能在 AI 上一事无成的叙事。
Wang 只愿透露 Muse 的用户数达到“数百万”，同时指向自己转发过的图表，暗示它可能是有史以来增长最快的消费级 AI 应用。 分发带来了认知，但增长依赖早期的惊喜时刻，以及不断升级的“信任坠落”：用户先交给智能体一个小任务，看到它完成，再逐步把更大的问题托付出去。最具传播力的案例是一位 CEO 让智能体在不打扰自己的情况下，找回遗落在 Flexjet 上的身份证；截图显示，智能体安排了快递，并一路协调取回物品送到他的办公室。
Muse 的破圈营销把具体用户收益、原生适配截图的设计，以及 Wang 越来越大胆的梗图攻势结合在了一起。 他的 feed 是亲自撰写的，并非由“100个 Muse”生成；他希望 Muse 和 Meta 都能冲破 AI 产品发布的“滔天洪流”，重新变得有文化影响力。他重新意识到的一条规律是，“互联网奖励风险”和意外；而 Jolly 吉祥物让每张被分享的截图都能被立刻识别为 Muse 的互动。
对投资者而言，Muse 也是 Zuckerberg 在 Llama 4 之后重置 AI 战略的一部分：人才密度、消费级分发和耐心执行，集中在一个小而扁平的实验室里。 Meta 对 Scale 的投资持股49%，让 Scale 得以继续运营，同时 Wang 与数名同事加入 Meta；Wang 在 Scale 工作9年后决定离开，是因为这个陷入困境的 AI 项目提供了真正“重新奠基”的机会。他如今把部分 AI 研究视为挖钻石：大量实验失败，但少数结果的价值可能高出100万倍；与此同时，200多名直接下属、pod 级技术负责人和反官僚组织结构，为研究人员提供了“尽情发挥的空间”。
1. Muse 始于一个关于人类应如何使用超级智能的命题
Wang 将 Muse 的起点追溯到 2025 年 6 月那份“个人超级智能”备忘录。该备忘录由他与 Zuckerberg、Nat Friedman 及 Meta Superintelligence Labs 的其他成员共同制定。目标不只是做一个助手或聊天机器人，而是打造强大的 AI，“真正让所有人的生活变得更好”，并从广义上“提升人类体验”。
其底层划分是：哪些事是人们想做的，哪些事是人们不得不做的。Wang 预计，先进 AI 会接管后者中的很大一部分，让人类有更多时间投入关系、兴趣、爱好和抱负——而这些恰好就是 Meta 现有产品一直覆盖的领域。
这套哲学也对应着商业优势。Meta 的应用每天有 35亿人使用，公司长期专注于连接亲友，并拥有规模巨大的消费级分发能力；因此，Wang 认为个人智能体和消费级 AI 天然属于 Meta 应该争夺的战场。
2. OpenClaw 提供了那一刻既恐惧又狂喜的概念验证
2026 年初，Opus 4.5 把智能体推到了聚光灯下，而 OpenClaw 则粗略展示了个人能动性可能是什么样。Friedman 成为 MSL 内部最早的支持者之一，他的体验在 Wang 看来“介于恐惧与狂喜之间”。
Friedman 对智能体的信任足够深，甚至让它在自己说想多喝水后监控安防录像；智能体会检查他是否照做，并向他表示祝贺。Friedman 的判断之所以令 Wang 重视，是因为他曾提前押中 Stripe、开源开发者工具和 GitHub Copilot：“我觉得这里面确实有东西。”
Wang 的入门体验来自 Friedman 撰写的一条心理探测提示词。智能体可以访问他的邮件、照片及其他个人信息，连续展开数轮探问，过程仿佛把“3年心理治疗”压缩在一个瞬间完成。Wang 说，自己在这个过程中“非常脆弱”。
两人都在 2月写下备忘录，此前他们利用每一个空闲时刻沉浸在智能体中。Wang 将超级智能体描述为在很多意义上的“终极消费级产品”；Friedman 则用了两个令人印象深刻的标题：“爪子即法律”和“信任是必须的”，帮助定义了产品的核心原则。
3. 两周原型变成了7个月的可靠性攻坚
Meta 大约用 1到2周组装出首个原型，并向董事会展示。Jollybot、早期的 Muse Charm 以及许多最终交互思路已经出现，尽管完整的概念集合和标签页当时还没有成形。
Zuckerberg 也经历着同样的狂喜与恐惧。他在家里、与孩子相处时以及 MMA 训练期间使用智能体，让它观看录像并提出改进建议；员工会议也开始频繁讨论各种最奇怪或最有用的智能体体验。
真正困难的不是展示魔法，而是让它稳定可靠。Wang 说，OpenClaw 的现象之所以达到峰值后消退，是因为用户会在频繁故障中偶尔遇到惊艳时刻；Meta 花了 7个月，把一个“有时有效、有时很神奇，但大多数时候都不工作”的原型，推进到适合数十亿用户的产品。
4. Meta 选择个人智能体，而竞争对手挤向编程
Wang 回忆，年初时 Opus 和 Claude 在编程上“遥遥领先”，市场一度看起来像是 Anthropic 即将接管。行业共识也越来越把编程智能体视为唯一有意义的 AI 产品形态、能力方向和竞争战场。
Meta 没有顺着这股引力走，是因为自身的智能体体验已经让团队在其他方向形成了确信。它没有追逐同一个开发者用例，而是把模型路线图对准个人超级智能，以及个人智能体所需要的一系列具体行为。
Senra 将两者的差异概括得很直接：“不是工作，不是编程，是个人。” Wang 表示认同；Meta 的消费级基因和分发能力，让它在个人智能体上拥有一个合理的胜算，而在企业软件或开发者工具领域，这个优势会弱得多。
5. Musepark 1.3 按逐项行为定义的发布标准通过考验
模型团队维护着覆盖100多项具体能力和行为的电子表格，为每一项建立评测，反复检查缺陷，再逐项训练——这是一个“非常细致”的过程，目标是逐渐磨平模型的毛刺。
Wang 将其与 Meta 过去的产品纪律联系起来。公司曾经优化转化、推荐和新用户引导的每一步；到了 Muse，这种本能转向底层模型本身：任务完成、沟通方式、可靠性，以及何时需要回到用户身边寻求指导。
每一行表格都有阻止发布的门槛，以及更高一级的“优秀”定义。即使80行已经变绿、20行仍为红色，也不能接受；首个清空整张表的模型检查点，是专门训练的 Musepark 1.3。
数字门槛之外，还有体验判断。团队反复开展用户测试、模型版本 A/B 测试和首次上手测试，因为消费者必须理解 Muse“不是聊天机器人”——它能够采取行动，但这种陌生能力需要被谨慎解释。
6. 小团队保护 Muse 免受大公司产品熵增
Muse 的模型和产品层共有不到200人参与。Wang 承认这个数字听起来并不小，但在 Meta 这样的公司内部，这仍然体现出组织上的克制，因为高管预计它会成为当年最重要的发布之一。
统领全局的理念是：Muse 是“一件完整的艺术作品”。Wang 担心大组织常见的失败模式——变成一个“弗兰肯斯坦怪物”、一个“Cronenberg 融合体”或一杯“奶昔”，每个产品经理都把各自的议程塞进同一个最终却不连贯的界面。
Friedman 的品位提供了大部分统一的产品观点。Senra 将这种作者性类比为人们对 Apple 的描述——“拥有1万条命的 Steve Jobs”：产品传递的是某一个人的感受力，而不是把背后的组织架构暴露出来。
Zuckerberg 扮演的是耐心和克制的角色。“AI时间里的7个月”感觉像过去了10年，华尔街也在质疑 Meta 是否能在 AI 上做成任何事情；但他仍持续公开提及个人智能体，并相信团队可以等到质量达标再发布。
7. 可靠性决定魔法会复利，还是让用户流失
个人智能体处在“刀刃边缘”：可靠的魔法可能造就史上最伟大的消费级产品之一，而非确定性的失败则会让用户认定整个品类都是垃圾。Wang 说，OpenClaw 的精彩时刻不足以维持现象级热度，因为用户最终会碰到它的局限。
Muse 的产品体验对底层模型高度敏感。它不仅要完成任务，还要以适当方式沟通，识别不确定性，在必要时请求用户输入，并且稳定做到这一切，让用户下一次愿意把更大的任务交给它。
这也解释了 Wang 所说的发明与商业化之间那条并不光鲜的差异。他将 Peter Steinberger 和 OpenClaw 的贡献者视为愿景型人才；Meta 的贡献则是“真正把所有细节一遍遍磨出来”，持续打磨模型和产品，直到两者组合出的体验能够稳固运行。
8. Muse 通过让 Meta 再次变得出人意料实现破圈
Wang 只愿透露 Muse 的用户数达到“数百万”。当 Senra 指出这可能意味着数亿用户时，Wang 仍拒绝给出具体数字，只是指向自己转发过的图表；那些图表如果属实，意味着 Muse 可能是有史以来增长最快的消费级 AI 应用。
Meta 的分发能力可以确保市场知道 Muse 的存在，但 Wang 认为，仅有认知无法穿透模型、产品、初创公司和功能组成的“滔天洪流”。团队明确讨论过，如何让 Muse 变得有文化影响力，以及这场营销能否让 Meta“重新变酷”。
Wang 说，那些被认为“发疯”的 feed 内容全部由他亲自撰写。梗图像等待解决的 bug 一样闯入会议和晚餐；一些随手发出的内容意外获得很高传播后，他重新想起“互联网奖励风险”和意外，于是进一步加码这种表达方式。
Friedman 开玩笑说，团队原本想了解 Muse 的用户，结果却在“深入了解 Alex 的大脑”。Wang 认为这场营销确实重要；Senra 自己收到的反馈是，讨论这些梗图的人，也下载并开始使用了这款应用。
9. 当用户收益能装进一张截图，产品营销就奏效了
反复出现的问题是：人们为什么要在意？Wang 的答案是，大多数人并不关心 Meta、OpenAI、Anthropic 或研究本身；他们关心的是产品能为自己做什么。
因此，产品推广的是具体用例，而不是机构品牌。Wang 最初转发了几十个有趣的互动；小吉祥物 Jolly 则成了一个灵活的文化载体，可以拎公文包、进入梗图，也能传递更轻松的产品个性。
Wang 说，Meta 确实投了几则广告，但令他意外的是，Anthropic 和 OpenAI 的营销中几乎看不到围绕用户收益展开的内容。他认为，现代消费级产品必须“在截图里成立”。Jolly 的存在让被分享的对话一眼就能被识别为 Muse，在群聊和社交 feed 中，截图因此与少量广告一起成为可辨识的分发渠道。
最有力的例子来自 Senra 的一位朋友：他把身份证落在 Flexjet 上，并告诉智能体，自己不想参与找回过程。智能体处理了快递、办公室安保和取回流程；3到4张截图，比一场抽象的品牌营销更有效地说明了产品能做什么。
10. 留存来自不断升级的信任坠落
Wang 将用户采用描述为一个连续过程：先给智能体一个小任务，看它成功完成，再交给它稍大一点的任务。每完成一步，用户对可能性的判断就扩大一些，最终形成与 AI 智能体之间的“信任坠落”。
Senra 注意到，一些大学生有时会直接完成一次完整的信任坠落：他们录下语音备忘，把脑中的想法全部倾倒给 AI，再相信它能把这些内容整理成计划和行动。这展示了在可靠性成为默认前提后，消费级行为可能走向哪里。
因此，实际目标不是泛泛追求参与度，而是尽早制造惊喜，再让用户持续顺利升级任务。如果智能体“每一次都能做到”，好奇就会变成信任，信任则会同时扩大任务规模和产品黏性。
11. Wang 将个人智能体定义为恢复人类能动性的第二个心智
Wang 主要把产品用于重复性工作流，并把它当作“第二大脑”，因为他的工作仍要求他出席会议。他更大的问题意识是：如果“每个人都有一个超越自身的第二个心智”，会发生什么——一种能够支持个人愿望、梦想和未完成意图的 AI。
Senra 认为，这款产品可能给人们带来“童年时梦想拥有的成年生活”。他描述了孩子们想成为宇航员、拯救地球或改变世界的想象，而工作和义务最终会侵蚀这些抱负。Wang 回应说，成年人都是“僵尸”，因为“生活已经把他们身上的东西打掉了”，他们失去了能动性和抱负。
Wang 所承诺的是一部“能动性自动扶梯”：完成一个抱负，提出更大的梦想，再重复这一过程。他认为 Elon 和 Mark 等标志性创业者已经在实践这个循环；个人 AI 则可能把它的某种版本提供给所有人，而不再只属于极度狂热的创始人。
Senra 给出了最宽泛的推演：历史一直由少数把欲望变成现实的狂热分子塑造，但人类从未见过“每一个人的能动性都被完全打开”。他想象数十亿人获得充足的智能，创造出一个奇异、多元且充满艺术性的世界——他的类比是《Rick and Morty》中的“跨维度电视频道”成为现实。Wang 则引用剧中的一句话回应：“宇宙会吞掉聪明人。”
12. Muse 源于 Zuckerberg 对 Meta AI 实验室的快速重建
Wang 早年与 Zuckerberg 的关系，主要是 2021年或 2022年经人介绍后偶尔交流；他甚至认为，两人的第一次见面可能花了1年才约成。在 Llama 4 令 Meta 内部失望、AI 的战略重要性进一步上升后，Zuckerberg 打电话询问公司应该怎么做。
从那通电话到 Scale 交易宣布，整个过程只用了5到6周。Scale 的企业和政府业务与 Meta 标志性的产品收购并不相似，因此 Wang 起初怀疑其中的产业逻辑；当讨论转向人才后，真正的目标才逐渐清晰。
最终结构是：Meta 持有 Scale 49%的股权，Scale 继续独立运营，而 Wang 与数名同事加入 Meta。Wang 称之为“三赢”：股东获得收益，Scale 保留未来，Meta 则得到了一批能够帮助重建 AI 业务的人才。
离开仍然在情感上很难接受。Wang 在 Scale 工作了9年，大约是自己人生的三分之一，也确实把它视为毕生事业。改变决定的，是重新奠基一个看起来像“坏资产”的项目、建立新的原则，并从头搭建大部分实验室的机会。
13. 新实验室的组织方式是挖钻石，而不是搭官僚大厦
Wang 创立实验室时的组织命题是：小型、扁平、高技术含量，并拥有极高人才密度。Meta 能提供一种类似创业公司的塑造机构机会，同时避开创业公司在算力、基础设施和分发上的瓶颈，所有工作都围绕个人超级智能这一清晰的北极星展开。
招聘并不只是给出一份显然更高的现金方案。Wang 说，随着 OpenAI 和 Anthropic 股价上涨，留在原公司对很多人来说本身就可能非常赚钱，因此 Meta 的报价在总薪酬上对不少人而言只是净中性。真正的吸引力，是参与塑造一间新实验室，并借助 Meta 的资源做有意义的工作。
Wang 自己的角色也从作者转向教练。在 Scale，他相信创始人应该亲自做每一项工作，这让公司带有他的观点，但也使他成为瓶颈；到了 Meta，他转而招募卓越人才、识别不寻常的天赋、设定方向，并创造条件，让他们“把自己的才华发挥到最大”。
其运营类比是挖钻石，而不是建摩天大楼。Senra 将前沿研究与数据业务的运营优化区分开来；Wang 说，这正是自己在 Scale 和 Meta 所做事情之间的差别。在一些 AI 领域，大多数想法都会失败，但某一个结果“价值可能真的比其他所有结果高出100万倍”。
Wang 用自己19岁生日的经历说明这种幂律：当时有10到12名哈佛和 MIT 学生在场，包括 Hyperliquid 的 Jeff Yan、Cognition 的 Scott Wu、Decagon 的 Jesse Zhang 以及 Anthropic 的 Vicky Ye。他开玩笑说，HRT 干脆投资自己的实习生就好。
实际上，200多人向 Wang 汇报，技术决策由 pod 级技术负责人解决。他承认，自己并不是管理这么多人最理想的传统经理；让这么多人直接汇报，本身是一种反官僚的表态：已经证明自己的研究人员“不需要经理，他们需要一个优秀的环境”，也需要“尽情发挥的空间”。
完整逐字稿
Tell me the history of Muse.
1. Rebuilding Meta's AI lab from scratch
The first thing to start with is probably what our vision was when we started Meta Superintelligence Labs. When we started it, Mark published this memo that Nat Friedman, myself, and a bunch of others worked on closely with him called “Personal Superintelligence.” The whole idea was: How do we make AI—how do we build powerful AI—that actually makes all of our lives better? How do we build something that actually lifts the human experience, broadly speaking?
In that memo, we spoke to a bunch of those ideas, and we alluded to a lot of what Muse ultimately became, but we were gesturing at it very early. This was in June 2025, before agents had happened in any meaningful way. In the summer of 2025, AI was in a totally different spot. But that was what we wanted to build toward. That was the mission of MSL, and that was really where we wanted to take everything.
Then we got to work building models because the end-to-end process of developing any of these frontier models takes many months. You have to build your entire stack, produce the data sets, pre-train the model, do post-training on the model, and build a whole stack for that. It’s a very lengthy, end-to-end process.
2. Discovering the potential of personal AI agents
This whole production line continued into the start of this year. Opus 4.5 obviously caused many people, including us, to see what the future of agents could be—or was—and really brought the potential of agents to the foreground. OpenClaw was happening at the start of the year.
Nat Friedman, whom I work closely with, was, I think, the first person within MSL to work with OpenClaw and try OpenClaw. He had an experience I would describe as somewhere between terrifying and euphoric. I think he really went all in and trusted it entirely. He’s told some of these stories—he had a Stripe Sessions interview where he talked about some of this—but he told his OpenClaw that he wanted to drink more water. His OpenClaw would watch him on security footage to make sure that he was drinking water and tell him, “Good job.”
He had many crazy stories like this, and it also makes you realize, “There’s something quite magical about what’s happening here.” I remember one of the things that Nat told me, and this is one of the things I remember the most. He said, “I think there really is something here.”
3. Leaving Scale AI for Meta
Early-adopter Nat Friedman is a pretty good indicator because he was one of the first customers of Stripe back in the day. He was also really early to a lot of open-source developer tools, helped create GitHub Copilot, and did many other things. That was a clear signal: There’s clearly something pretty magical about personal agents, generally speaking.
4. Three years of therapy in one AI session
When you actually trust the personal agent and really bring it into your life, it can do so many things. After hearing all of these crazy stories, I started using it very deeply. For many weeks, I think both of us spent every spare moment talking to our OpenClaws and really immersing ourselves in what personal agents could be.
What was your initial reaction: euphoria or terror?
Nat had built this prompt to do basically a full psychoanalysis of yourself. There were multiple waves of deep psychological probing. He had written this prompt for you to use with an OpenClaw, and I think that was one of the very first things I did with my own OpenClaw. It was maybe the equivalent of 3 years of therapy compressed into one moment.
I’m curious: When you get your first shot at OpenClaw, why is this one of the first things you do?
I downloaded it and installed it, and then Nat said, “Oh, you should do this.” I was next to him, and he said, “Okay, I’ll just do the thing you’re telling me to do.” Then I had this crazy experience.
Say more about the crazy experience, though.
It was very intense. It was like 3 years of therapy because it was successive rounds of deep psychological probing. It had other information about me, including access to my email and a bunch of my photos. It could look up other information and pull it all together. I was very vulnerable through that process.
This was in February. I think both of us had come to the joint conclusion that there was clearly something really special happening here. We both wrote memos at that time and sent them up through the Meta board. For whatever reason, it was board season.
I wrote a memo about how I felt that, in many ways, this was the final consumer product—that the super agent was, as far as consumer products went, a clear endpoint in many ways. Nat actually wrote 2 memos, which had very memorable taglines: “The Claw Is the Law” and “Trust Is a Must.” Those ultimately defined a lot of the core of how we built these products.
5. Building Muse in 7 months
We came to the conclusion that there was something very important here. This was in February. Then we built our teams, built a prototype within a week or 2, and demoed it to the board. It was a very fast cycle, and many of the decisions that made their way into the final Muse product were present in the first prototype.
The Jollybot was there. An early version of the Muse Charm was in that first meeting. Not all of the ideas and not all of the tabs were there, but a lot of them were present.
From there, obviously, that was February, and we launched the product 7 months later. The really hard process was taking this prototype—which sometimes worked and was magical, but most of the time didn’t work, constantly broke—and turning it into something highly reliable. We wanted it to be a beautifully crafted consumer product that we felt confident offering as a contribution to the consumer world. That was a very lengthy and quite painful process.
Before we get there, I want to go back to the idea that you created the prototype in 1 to 2 weeks and showed it to the board. What was Mark’s role in all this?
He was experiencing the euphoria and terror of the product at the same time as all of us. I think he’s talked about this in various interviews, but he used it for a bunch of things around the home, used it with his kids, and used it in his MMA training. He had it watch videos of himself doing MMA training, and it would give him tips on how to improve.
There were also staff meetings—and this became a thing for much of the Meta team—where we would have a staff meeting and ask, “What were the crazy experiences you had with your agent?”
Again, this was back in February. It’s interesting to think back to this whole cycle because the OpenClaw phenomenon peaked and then disappeared. It didn’t persist. I think a lot of people had this experience: You’d have these magic moments, but for the most part, it didn’t work that well. There were all these problems, and then it just died down.
Part of building the product was figuring out how to make this an actual, reliable consumer experience that you feel comfortable putting in front of, over time, billions of people—and feel confident that most of those people will have an amazing experience.
Why were you able to do that and the developers of OpenClaw were not?
First of all, we respect Peter Steinberger and everyone who contributed to OpenClaw immensely. I think Peter is definitely a visionary. Whatever happened in his beautiful brain that created OpenClaw is magical.
The unglamorous part of what made Muse so magical was really grinding out and sanding down all of the details to make the combined model-and-product experience amazing. We did many iterations on the model itself. We aimed the overall model roadmap toward building personal superintelligence, broadly speaking, but certainly personal agents as—
Say more about that and how it differentiates you from your other competitors.
This is useful to go back to the moment because if you go back to the start of the year, in February, Opus and Claude were way ahead in coding—way ahead of anyone else—and they were exploding. It was the beginning of the Anthropic takeover of the whole industry, and it was a very scary moment, generally speaking. It was like, “Wow, they’re so far ahead on coding.”
Obviously, they had Claude Code, and they had this clear lead.
I think the conventional wisdom in the industry was that coding agents were the only form factor that mattered, or maybe the only technology or capability that mattered for advanced AI. There was this incredible force in the industry to crowd that use case and compete with Anthropic on coding agents because of our experiences and our conviction, which I think formed around this idea of personal agents.
Taking a step back, I think the reason why we did personal superintelligence in the first place was that it was also what we believed Meta was uniquely positioned to do. We can talk more about that, but I think that was a big part of it.
Well, let's talk about that now.
Okay, we'll do the weave. I think one of the first things I told Mark, even before the whole Scale and Meta deal happened and everything, was that I thought Meta was actually a really special company if we had AGI or ASI, or whatever you want to call it.
Part of the promise of AI is that all of us have a mix of things that we want to do and things that we have to do. The promise of AGI is that we're going to spend almost none of our time doing things that we have to do, because the AIs will just start doing a lot of that for us, and we can spend all of our time on things that we want to do.
For me, Meta's products represent the things that we all want to do. My use of Instagram, WhatsApp, or Facebook—these are products that represent what I want to do. I follow my interests and hobbies. I follow people that I like, my friends, and the people I care about. The product span of Meta encapsulates the realm of what we want to spend our time doing.
I think on a long arc, Meta is an incredible winner through all this. That's one of the first things that I believed about Meta and superintelligence. Ultimately, a lot of those ideas made their way into this concept of personal superintelligence.
You can look at it a few ways. One is the heritage of the company. The heritage of Meta is connecting people with their friends and family and helping them discover their interests. These are things that are very human.
Then you can take the business lens. Meta has distribution to 3.5 billion people using the apps every day, and the natural area where Meta was going to be successful is consumer AI.
It's not work. It's not coding. It's personal.
Yeah, exactly. Personal agents, personal AI, and consumer AI, broadly speaking, were always the zones where we felt that Meta was going to be able to do something very special.
That brings us to early 2026. Now you have the prototype, you have the buy-in, everybody knows Mark's on board, and now you're doing the 7 months of difficult work refining this product.
Yes. There was kind of a specter around personal agents. They came and went, and we always felt that it was because the experience wasn't perfect or polished yet, and all that kind of stuff. But we knew that was something we had to think about as we were developing this product.
We focused our model development roadmap on personal agents. The top focus, or one of the top focuses, was building personal agents. We had these long spreadsheets of hundreds of different—more than 100—specific behaviors and capabilities that we knew the model needed to have to build the ideal personal agent.
It was very meticulous. Over time, we built evals for all of those. We figured out how to train the model to be really good at every single one of those things. We had reviews where we looked at that spreadsheet, figured out the things we were bad at, and made sure that we were improving on them.
We did this very meticulous process to slowly keep improving and sanding out all the edges of the model so that it could power the product.
Is that really the culture of Meta, too? This constant sanding of the products?
Yeah. A lot of the culture of Meta is that once you know what you're measuring, and you have very talented people who can optimize for those things, you can sand out all the details and make those things amazing.
In the pre-AI consumer product era, a lot of this was about how to optimize your conversion, how to optimize referral rates, and how to optimize every part of the onboarding flow. How do you make all that stuff perfect?
In the AI era, a lot of that, especially from Muse, has been in the form of sanding out every detail of the model to make it perfect for the product.
6. Mark Zuckerberg's obsession with improving products
Did you see this profile on Zuck in Colossus written by Jeremy Stern?
I saw that. He's writing the best profiles on entrepreneurs and investors in the world right now. I think he's literally the best in the world at what he does. In that profile, Zuck was talking about the advantage that he has relative to Elon, Sam Altman, and others. He was saying, "Listen, my skill set is very unsexy, but I'm really good at building teams and then improving a product slowly over a long period of time."
Relative to Elon, Sam Altman, and others, he's not going to lose control of his company, and he doesn't need to raise any more money, so he can just do this for an excessively long period of time. That section of the profile is popping into my mind when you're talking about this 7-month stretch of trying to refine and build Muse.
I think that's exactly right. It took immense patience. 7 months might not sound like that long, but AI is fierce.
7 months in AI time is maybe like a decade in non-AI time. I think it was that way for all of us, but I really give Mark a lot of credit for this. It took a lot of patience, restraint, and trust in the process.
You can go back and look at what Mark said in earnings calls for this entire period, from February through September. He alluded to the fact that we were building personal agents, and we weren't really trying to hide it. We were building personal agents and talking about it.
But I think there was a lot of doubt in the company for that entire period. One of the dominant Wall Street narratives was, "Oh my gosh, Meta is just burning all this money. Are they even going to pull anything off in AI?"
It's more than that. It's that Zuck cannot win AI. That's been said over and over again for a long period of time.
Going back to Jeremy Stern's profile, he starts talking off the record to his competitors, including one former researcher who used to work with Zuck and is now at another lab. The researcher says, "I don't want to have to compete with them because Mark's the [__]." I think the line is, "Mark is the [__] Terminator."
He just will not stop.
And then he talks about one of his superpowers: He's never happy with the state of the product, so he always wants to constantly improve. But he's not an asshole. He is definitely a dictator, but he's not rude or the kind of person who will destroy the chemistry of a team. He actually talks about, in the piece, how excessively important team cohesion is to him.
I think it's all right. One of the magical things about Muse was that we took quite a bit of time to sweat all the details before ultimately coming out with a product, because I think that personal agents was one of those product areas that rested on a knife's edge. If it can reliably deliver these magical experiences, it has the potential to be one of the greatest consumer products ever.
But you're dealing with models, something nondeterministic, and agents. They can break; they can be unreliable. If people have these unreliable experiences, they think it's trash. It was one of those things where getting the product to a point of quality and value such that it resonated with so many of the people who downloaded and used it was really important.
7. How Meta knew Muse was ready to launch
I want to talk more about this process. I'm curious: How many resources did you have? How big was the team? How many people were working on it? Do you guys have any data on, okay, I'm using Muse—9 great experiences, 2 inconsistent or shitty experiences—and then I tap out? Do you have any information on this?
You can see this from a lot of the other agent products at the start of the year. Almost every one of those turned out badly, and it's because they were unreliable.
We did a lot of user testing of Muse. We did many rounds of new people trying to download it and use it from scratch, just seeing what that experience was like.
A huge part of the premise of Muse, and I think what's really landed, is that it is a different thing from what most people have experienced with consumer AI. It is not a chatbot. It is an agent, and it can do agent things for you.
I think to a developer this is old news because they've had to download agents and have been using them for a while. But for most people, even that graduation process needs to be explained, and people need to be onboarded in a very thoughtful way.
For sure, we've seen that the product experience is very sensitive to model quality. We've done tests with different versions of the model and continuously do those A/B tests. The product experience is very sensitive to how reliably the models can undergo these tasks, how reliably they communicate with the user about them, and how reliably they check back with the user if there are things that require guidance.
But how did you know it was ready to be released?
Part of it was pretty numerical. We had set—you go back to that spreadsheet of 100-plus things—for each of those rows what threshold was launch-blocking and above what threshold was good.
We had many checkpoints that were green on 80 of the rows but red on 20 of the rows. We knew that wasn't good enough. We had our first model checkpoint that was green across the board, which was a specially trained version of Musepark 1.3.
So there was the very numerical part of it, and then there was just the experiential part. We tried the model, and we could tell that it was a lot better than anything we had iterated with before.
So, going back to that 7-month period you were talking about, how many people were working on the product between the model work and the product work?
Under 200 people total, which I guess sounds like a lot, but in big-company land—
It's actually a lot smaller.
It's not very many. That was kind of cool, looking back on this and thinking about it. We knew this was going to be the most important product. The whole executive team—everyone knew that this was probably going to be one of the most important products that Meta would ship this year.
But we had the restraint, organizationally, to make sure that it was a really small, focused team working on it and that it didn't balloon into huge swaths of the company working on it, because we knew it was a single work of art. It required a strong point of view and strong focus to pull that through into one beautiful experience.
8. Why great products need a single point of view
Say more about this. Why did you just call it a single piece of art and say that you need a single perspective or point of view?
I think one criticism of products that come from larger organizations—and I think this is even now starting to plague a lot of the AI labs—is that they become kind of this Frankenstein, almost like a Cronenbergian amalgamation of a bunch of different people's points of view, visions, and beliefs.
It's sort of PM hell, in some sense, where every PM has a thing that they're trying to get into the product and jam into the product, and then all that blends together into this smoothie. People feel it as users and consumers. You can feel when a product feels like lots of different people worked on it and there were clearly different people who were goaled differently, so they just jammed it all together.
Muse had to be the exact opposite. There was one clear point of view. There was one cohesive experience that we were trying to deliver to the world.
I love that you said that, and it just happens to be pure coincidence. I don't know when this is going to come out, but we're recording this on the 15th anniversary of Steve Jobs' death.
When you think of somebody like a pro at consumer products who had a single point of view, you obviously think of Steve first. I've read every single book on Steve and the history of Apple, and I've done 15 or 20 episodes on him for my other podcast, Founders.
One of my favorite lines to describe Steve's impact says that Apple is just Steve Jobs with 10,000 lives. The products are to his taste, his perspective, and his point of view. What he wanted to see in the world is exactly what the end user got.
In Muse's case, Nat Friedman deserves a lot of credit. His taste and sensibilities ultimately dictated and shaped what that product became.
Do you talk about how many people are using Muse? Are you guys talking about this publicly or not?
We've said millions, and I've retweeted other people who have shown graphs where we would imply that Muse is the fastest-growing consumer AI app of all time.
So you won't say a number, though?
We've said millions.
Yeah. Well, millions could be hundreds of millions. That's also millions. So you're just not adding the beginning to that.
9. Making Meta cool again with memes
I want to go into—you've been kind of going, and I mean this in a loving way, unhinged on X. You talk about your—what is this? Is this a marketing strategy? What is going on? I feel like there's a distinct difference in your public tweeting since Muse launched. What are you doing?
There are a few pieces of what's happening here. The past few weeks have been some of the most fun I've had building—certainly of my career.
This has got to be more fun than Scale AI when you were building Scale. It has to be. We'll talk about that later, but it has to be.
I think there are a few things happening. One is that it felt important that we figure out a way to make Manus break through.
Going back to this, it's a pretty torrential stream of updates. For most people, they're constantly being bombarded with new AI products, startups, releases, models, and features. Here's another model, here's another product, here's another feature. It's just this torrential stream.
10. Why Meta bet on personal AI
We actually had large meetings where we talked about this as a team. We needed to break through. Obviously, at Meta, we have lots of distribution, so we can make sure everyone knows about it, but that's not enough to make it break through in an interesting, cultural, or fascinating way for the world to pay attention.
I separately happened to have met and become friends with memers—people who literally spend all day thinking about memes on the internet and understanding them. I have people who are obsessed with Manus, and they're also obsessed with your X feed. They're convinced that you have 100 Muses coming up with memes to be used.
It turns out you just have a collection of schizophrenic friends who are terminally online.
The X feed, just for what it's worth, is literally every one of those tweets I write. There were quite a few periods when it was really hard for me to pay attention in meetings because I would just think of a meme and be like, "Oh God, I have to get this out."
They were moments of inspiration. I was trying to have dinner with one of my coworkers, and during dinner I was like, "Wait, I figured it out." Then I posted a different meme. It was almost—I almost felt like I was an engineer solving bugs again, because you could tell when something was going to click.
One thing we wanted was to break through. Another thing that I thought was important was, how do we make Meta cool again? For whatever reason, despite having these incredible pieces of cultural software, like Instagram and Facebook, historically Meta was not cool for a while.
I wanted to make Meta cool and interesting and kind of funny. The unhingedness happened naturally over time. When Manus first came out, I was mostly focused on promoting all the interesting use cases that I saw. I was retweeting tens of interesting use cases that I saw of people using Manus.
And then, at some point, because I was tweeting so much, there were a few tweets where I didn’t think that much about it and just said, “Fuck it, post.” Some of those tweets did really well. Have you seen this? “Time spent thinking about it and then banger.” Have you seen that? Great.
I just didn’t spend that much time on some of these. I tweeted them, didn’t even check my phone for a few hours, and then I was like, “Oh, wow. That one went really well.” I remembered something about the internet: It rewards risk, surprise, and things that people don’t expect.
Then I just went off. A lot of my coworkers were somewhat surprised at something that Nat said during the week when I was posting all the memes. He said this in a bunch of meetings: “I thought we were going to learn a lot about Muse users and how people were using Muse, but actually, we’re just learning a lot about Alex’s mind.” It was a lot of fun. I felt like I was drawing on all of my internet knowledge from years and years and years to pull into those moments.
I think it genuinely made an impact on growth, which is kind of hilarious.
I definitely think it did. I can’t stop hearing about it. All the people I’m hearing about it from also downloaded the app and are now using it.
Go back to what you were saying, though. We have distribution, but that doesn’t mean that we can make something that people are going to come back to and that’s actually sticky. How did you do this with Muse?
I think the question was: How do you make this breakthrough as a product? There were a few parts to that, but one is, how do you get someone to a wow moment with the product as early as possible?
This product naturally lends itself to that because it’s AI. It can do a lot. If someone has never used an AI agent, there’s a lot it can do that will surprise you. I think that was a big part of it.
Another part was: How do you harness all of those stories of people having these wow moments and use them to help make the product sing and make the product fly?
How are you using the stories, though? You guys aren’t running ads for Muse, are you?
We ran a few ads. Honestly, at first, it was just the fact that I would retweet so many of them. Anytime I saw someone do something interesting with Muse, I would quote-tweet it.
I couldn’t understand why both Anthropic’s and OpenAI’s ads were so bad. Anthropic was doing this huge outdoor campaign, and it was all about them. Same thing with ChatGPT. They’d have the icon, which I don’t even think people recognize as a logo, and it would just say “ChatGPT” and show 2 people sitting at a desk. I’m like, “What the fuck does this mean?”
You have all these use cases. All you have to do for your ad is show the benefit of your product. That’s it.
You have millions of them. Run those ads. It’s so elementary that it’s interesting that people make this mistake over and over again.
Yeah, 100%. I think there are a few insights. First, most people in the world don’t give a fuck about OpenAI, the research, these other companies, or even us and our research. They care about what it can do for them. Everybody’s self-interested.
We had this gift of the mascot, which was the little jolly guy. I remember the tweet where I realized, “Oh, wow. This is mileage.” I was driving into work, and on the way I generated a version of the guy holding a briefcase. I tweeted, “MFW I go to work on Monday.” I remember thinking, “This is one of my best tweets in a while.”
Then I realized you can put him into a lot of situations. One thing I realized as we were going through that process was: How do you create surprising moments that are very different from any other marketing you’ve seen for a product? I don’t really remember the last time someone leaned into racy memes to market a product. I don’t know if it’s been done before, but I realized, “Oh, wow. This is something we can do with this little guy.”
11. Building trust with AI agents
That’s interesting. It drives intrigue for the average person.
Okay, but that might raise awareness and might even get them to download the app. How do you get them to stay engaged once the app is on their phone?
There’s a cycle that people who really love the app—and we even had this back in February—go through. You try to give it a little thing, and it does it. You’re like, “Oh, wow. It did that.” Then you try to give it a little bit more, and it does it, and you’re like, “Oh, wow. It can do that. It could do that first thing, and it could do the second thing.”
Then you give it a slightly bigger problem. It’s almost this trust fall that you have with an AI agent. At first, you don’t really trust it, but you’re intrigued, so you give it a little nibble. Then you give it something a little bigger, and something a little bigger, and something a little bigger.
Certainly, when I talk to college students, some of the ways people use AI are a full trust fall. They’re recording voice notes and brain-dumping to the AI, sending them, and then trusting the AI to organize all the thoughts, give them clear plans, and tell them clear things to do.
For successful Muse users, it really is about getting them onto this trust-fall process of using it for slightly bigger and bigger things, and then having it work every one of those times.
I wonder how much of this is actually driven by word of mouth. Obviously, talking to you, I was going to download the app no matter what. But what really piqued my interest was that a friend of mine was sending me screenshots of how he was using Muse.
When you download the app, it doesn’t really tell you too much about what you can do. It’s kind of open-ended, like an open text box. But a friend of mine who’s running a $100 billion company left his ID on a Flexjet, and he told me, “I don’t want to be involved in this process at all. You have to figure out how to get the ID. Figure out where it is, figure out where I am, figure out all the steps along the way, and arrange the courier.”
Manus went there, got through security at his office, somehow got upstairs to his office, and retrieved it. It was like, “Holy fuck, this is the best marketing.” It's just 3 or 4 screenshots of Muse going back and forth, with him saying, “I don't want to be involved. You have to do everything.” And it's figuring it out along the way.
You know, this is actually one of the things about modern product marketing: it's very screenshot-driven. I think for a modern consumer product, it has to work in a screenshot. The screenshot has to fly, so to speak, either through group chats, word of mouth, or online.
12. The future of human ambition
I think that's been a big part of it. This wasn't intentional or engineered, but I think the fact that we had the little guy in the screenshot was a big deal. Having Jolly, or whoever your Muse is, in the screenshot immediately communicates, “This is a Manus screenshot.” It's just the best marketing you could possibly have.
So, are you on your Muse all day long? What is your own personal usage?
I mostly set up a lot of workflows using it and use it in my meetings a lot of the time. In practice, in my job, quote-unquote, I'm supposed to pay attention, be present, and interact with people. I set up a bunch of workflows and treated it as a second brain.
You said that in the post you wrote, and it was really short. I think the last time I saw it, it was at 5.5 million views. I love how you describe Muse: “What if everyone had a second mind beyond their own?”
Philosophically speaking, I think this is one of the most interesting questions about what it means to be human when you have powerful AI, and what that relationship looks like. Different people have different takes on what exactly that looks like. The fear case, obviously, is that AI is above us and we're just doing what it tells us to do. There are other worlds where that relationship is different. Before you go on, what is your own personal view?
I really believe in Muse. I believe in the personal agent as a long-term form factor. It is human nature to have wants, desires, and dreams, and I think AI and personal agents like Muse will be the mechanism and the bridge that enables us to accomplish those things and continue accomplishing them.
Humans will dream bigger and bigger and bigger and bigger. Kids will dream of having their own star systems or whatever, and Manus will help them do that. There's this thing that Bezos wrote in one of his shareholder letters—I think—which is that the beautiful thing about consumers is that they're always beautifully unhappy, or something like that, or beautifully unsatisfied with the options they have.
That is one modality of our relationship with AI. I think there are other modalities. AI will start doing more and more of the scientific discovery process. AIs will start inserting themselves more and more into various parts of the economy. But I do believe in this form factor long-term, of everyone having an AI that supports them and what they want.
So you see AI as removing all the kludge of life—the stuff that we don't want to do. In the piece you said something like, “The world is full of gatekeepers and obstacles, and Muse can get around this for you without spending any more of your time, so you can focus on the stuff you want to do.”
I think Manus can give everyone the adulthood they dreamed of in childhood. When people are kids, they have big ideas and big dreams. When they imagine their life and play it out, they imagine a life where they can be an astronaut, save the planet, change the world, or whatever it might be. They think big.
Then, for a variety of reasons, by the time people finish school and enter the workforce, and by the time they've been in a job for a while, all that hope, ambition, and ability to dream has been sucked out of them.
No, it's been beaten out of them. The life has been beaten out of them. As adults, I don't think we appreciate the degree to which we're all zombies. We lost our agency, and we lost our ambition. I think ambition is a good word for it.
One of the promises of AI, broadly speaking, and of products like Muse, is to keep that going from when you're a kid. You're a kid, you have big dreams, and you use AI to help make those things happen. Then you dream bigger, make those things happen, dream bigger, and make those things happen. You experience this escalator of agency—an increase in agency throughout your life—versus having lots of agency and then having it crushed.
You mentioned earlier this experience that you and Nat Friedman went through, where you had this disturbing psychological audit coming in waves from this AI. What if we had that, but it was the opposite of what you're saying? What if it actually gave you more self-confidence and more understanding that the world is malleable, and that if you push on it hard enough—if you go after it with enough energy and drive—you can actually change the world around you?
Yeah. I think this is something that the most impressive entrepreneurs exhibit. They dream big, accomplish that, then dream bigger. Maybe they work on that for a decade, and if they accomplish that, they dream bigger. Elon is obviously a great example of this. I think Mark is a great example of this.
A lot of iconic entrepreneurs have this as their lived experience, and I think there's a version of that that should be true for every person. What's tragic is that, for most people, you have bigger dreams and big ideas, then you enter the corporate workforce and become a zombie. Maybe at some point you want to go tackle your dreams, but it's really hard because you have all these commitments. Maybe you have a family, maybe you have whatever it is. You almost become trapped in that sort of zombiehood.
There's one more thing on Muse before I want to get to Scale AI and the partnership you did with Meta and how you made that decision. In this essay, or this short post, that you wrote, I love what you said: “The world until now has been shaped by the small number of fanatics who have somehow found a way to make their wants real, but we've never seen humanity with every single person's agency fully switched on.”
I think the promise of this world, where every single person has an increase in agency throughout their lifetime and has the ability to accomplish their wants and dreams, looks crazy in a very good way. I think it will be very interesting, artistic, and cool. You could go into different pockets of the world and it would be very diverse. It would be kind of insane to think about what that looks like, where literally billions of people have, all of a sudden, because of abundant intelligence, the resources to make incredible things happen.
I think that's part of the promise. In my head, it's almost like the Rick and Morty Interdimensional Cable, but somehow manifested into reality for humans. It could just be really awesome.
I keep hearing this line from that show where it's like, “The universe eats smart people.”
13. The call from Mark Zuckerberg that changed everything
Okay, so I want to go back. What I'm personally curious about is that you founded Scale AI, you're running Scale AI, and then one day Mark Zuckerberg reaches out and says, “Hey.” I assume he says, “Hey, I want to talk. Can you tell me about this?”
Yeah, I think the exact message was something like, “Hey, do you have time for a call?”
Did you have a relationship with Mark previously? Did you spend any time with him? What was the background there?
I had a friend, Alex Schultz, whom I've known for many years, since Scale was maybe 1 year old. I met him in San Francisco. He's currently the chief data officer at Meta, but he was a longtime Meta executive and lieutenant, and he introduced me to Mark—I want to say in about 2021 or 2022—to talk about AI.
I think at that time Mark was getting a lot more interested in AI, and Scale was doing a lot of stuff in AI. My memory is that it took 1 year to schedule that first meeting. From the introduction to the meeting being scheduled, I think it was a full year.
Why?
Mark obviously has an insane calendar and a bajillion different things to deal with. He's also very good at prioritization. He'll spend a lot of time on the things that are really important—in some ways, too much time on the things that are really important.
The only way he's able to do that and be a good dad, be a good partner, and do all this other stuff is to ruthlessly prioritize. I had spoken to him a few times after that and had gotten his advice as a founder. Getting advice from Mark Zuckerberg is obviously a really big deal.
I remember at that time, he would write these really long responses. I would ask him a question on WhatsApp, and he would write a really long response. I was really confused. I thought, “How does he have time to write these really long responses?” But now I realize he's just really fast and rigorous at writing a lot of WhatsApp messages.
He'll write a very rigorous, long response very quickly. It's a very impressive skill of his.
It gives you a sneak peek into what's going on in his mind.
Yeah. There's a reason. Meta was an important customer of Scale. Prior to this instance, maybe we spoke once every 6 months or so. Maybe that was the cadence of our interactions.
Then he called, and I think it was actually not obvious at first what the endpoint was going to be, because the first conversation was just—this was after Llama 4—and it was kind of just asking, “What do you think we should be doing?”
Why is it an important point that this conversation is happening after Llama 4?
Llama 4 was not on the trajectory that Meta wanted as a company. I think, internally speaking, it was definitely a disappointment, and it was also at a time when AI was becoming increasingly important. It was clear that AI was going to be really critical to the future of Meta.
The first few conversations were on the phone, and he was asking for advice: “What do you think we should be doing? What do you think we should be focused on?”
This is what I've heard about him privately: He has this insane—it's not a board of directors, like a Meta board, although there are some people on the board that he also does this with—but he's got this group of world-class entrepreneurs around him. I've spoken to some of them, and you don't understand—he asks for advice constantly. He's hitting us up: “Here's what's going on in my life or at work. What would you do?”
He'll do that over and over and over again. I think this is quite an impressive trait because it takes humility, obviously, to continue doing that even after all the incredible things that he's been able to accomplish.
The whole timeline from when he first reached out to when we announced the deal was 5 or 6 weeks. It was pretty quick.
Another thing that world-class entrepreneurs have in common: We had Jonathan Ross, the founder of Groq, on this podcast. The first time we ever spoke publicly about the $20 billion deal he did with NVIDIA was on the show. He's like, “From the time of the first call from Jensen to the money being in my bank account was 3 weeks.”
Yeah, that's amazing. I had lots of ideas about what I'd be doing if I were him and how I'd be thinking about Meta's AI strategy.
One of the first things I said was what I said earlier, which is, “Hey, I actually think that because Meta is, heritage-wise, focused on the things that people want to do, it's one of the most well-placed companies for this incredible shift to AI.”
I had lots and lots of ideas that I sent his way, and that was this conversation happening in parallel with, “Oh, yeah, maybe we should potentially consider if there's a way to work together.”
So when he says that, what do you think? This was not predictable to you.
Not predictable. No. The whole thing was a really crazy sequence at the time. I was just like, “Oh, that's nice, but there's no way that—”
Well, you know what that means when they say that, right? They want to buy you. It's always the same thing, but they say, “Oh, we have to find a way to work together.”
Yeah. I don't think I'd had that kind of coaching at the time. One thing that I explicitly felt at the time was, “Yeah, it probably doesn't make that much sense, even.” I think I personally was like, “Yeah, maybe he's teasing it or throwing it out there, but does it even make that much sense?”
Why wouldn't you think it makes sense?
Obviously, a deal ended up happening, so what do I know? But—
No, no, but back then—not now that we know what happened—I'm very curious about your thinking as you were experiencing this. The fact that you said, “Oh, yeah, this doesn't make sense.” Why wouldn't it make sense?
Taking a step back, the iconic Meta acquisitions have been Instagram and WhatsApp, and these were very clear, product-logic-driven decisions. Scale is a deep enterprise- and government-sales business, an entirely different kind of business from Meta.
The logic that I thought would have made sense at the time was if Meta wanted to get into those things. Then maybe that's the industrial logic that would have made sense, but this wasn't that, because I think Mark was clearly predominantly interested in getting the Llama program and his overall AI program on—
He wanted talent.
Yeah, he wanted talent. It was shrouded and unclear the whole time, in a lot of ways.
When did it become clear?
After the money hit the account—when we started talking numbers, I was like, “Oh, wow. Okay. All right.” That's when things became more clear.
At first, you're like, “Oh, this will make sense.” Then you guys keep talking. How fast did he convince you this was a good path for you?
Part of it was just, “Wow, this is a really fascinating deal construct.” We ultimately landed on this deal where Meta invested and owns 49% of Scale, Scale continues, and I and a few people join Meta.
What I was kind of incredulous about at the time was that it's such a weird kind of deal, but it's one that I think genuinely was this very interesting win-win-win, so to speak. There's a win for the shareholders of Scale because all the shareholders of Scale got a great deal and benefited a lot. Scale continues, and I truly believe the best days of Scale are ahead of it.
My question to you—or what I'm personally interested in is this: He can explain this to you, and you come around to that perspective in one phone call? Is this just, “Wait, Mark, I've got to think. This is so fucking crazy. I've got to think about this for a few days”? Explain this process as much as you can.
It was many weeks of thinking, “This is insane. Is this even real? Does this make any sense? If it does make sense, how do I feel about it?”
The predominant emotion was more like, “This is kind of insane.” Then, as I thought more about it, I realized, “Oh, I actually have to make a decision about whether or not I do this.” That was it, in and of itself.
There were a lot of conversations with the people at Scale and with our investors. There were a lot of interesting conversations on that side. It's obviously really hard to let go of or give up your baby—
Because you worked on Scale for how long? How many years?
9 years, from founding until the—
And you're still young. You're not even 30 yet. This is a third of your life.
There was something that Paul Graham said for a long time, which is, “If you think of your company as your life's work, you will operate differently.” I genuinely thought of Scale as my life's work for the whole time that I was working on it.
That was a tough emotional process. What ultimately got me was a combination of, “Wow, this is a win-win-win. It's a good deal for all parties,” and also that I saw the potential of what could happen at Meta.
This was a point at which I think Meta on AI certainly looked like damaged goods in many ways. It was maybe not the most appealing place to work on AI, but I think that triggered my entrepreneur side, where I was like—
You can essentially rebuild it. It's like a refounding of the lab.
Yeah. I think this is one of the things that ended up being fascinating. Because it was clear that there was so much work to do, it really cleared the way for me to build a lot of stuff up from scratch, set up the right principles, and build the right culture.
It created the conditions for us to build something amazing.
Is that how you and Mark discuss it? Is it like, “Hey, we have to hit a reset here. Obviously, if we keep on this path, we're essentially going to throw out what we have and redo it with a completely different level of talent”?
14. Diamond mining vs. building skyscrapers
Are these the conversations that are occurring between you and him? It didn't seem like he wanted a minor adjustment. You understand what I'm saying? He's just like, “Oh, this is not working at all. It's not going to work. Then let's rip it down to the foundation and rebuild.”
Correct. It was an evolving conversation. One of the things that I believe really strongly is that Elon has demonstrated this at various times in his career: You can have a small, incredibly cracked, highly technical team that's very flat, and if you do that, you can accomplish a lot very, very quickly.
That was something that I had a lot of conviction in, and I believed in AI. That is the right approach, especially if you have to do what we had to do, which is move very, very quickly.
Early on, that was one of the things I talked to Mark a lot about, and I think he was excited about it as well. That ended up becoming one of the tent poles of our overall strategy: a small, very flat, highly technical team with very high talent density. How quickly can you move if you have those ingredients?
From an outside perspective, it just seems like Mark empties the clip. That's the way I think about this, right? Have you ever read this book called The Mind of Napoleon?
No.
Okay. So, The Mind of Napoleon was published in 1957. I found it because I saw an interview—I think Tyler Cowen interviewed Sam Altman—and I think this was in 2018. He asked him, “What’s the most important book that you read this year?” and he said, “The Mind of Napoleon.”
It’s very hard to find, so I bought the book. It’s 300 pages of just Napoleon’s own words, organized by topic. In it, he talks over and over again about how hesitation is fatal. But he says that before you engage in a course of action, there’s a lot of deliberation: he’ll study it from every angle and make sure he’s making the right decision.
I feel like there’s an echo in the way Mark operates. He does a lot of deliberation and makes sure he’s on the right path, but once he makes that decision, he goes all in. He just empties the clip. That’s the way I felt about what was occurring at this point, where he comes and gets you and then you guys start rebuilding this entire organization.
I think he’s really internalized that failure doesn’t matter. What matters is when you win and how big you win. I think this overall concept enables him to operate truly without fear in a lot of circumstances, because he’s very comfortable taking lots of risk.
He’s got that great line where it’s like, “In a world that’s changing all the time, the biggest risk is not taking any.”
Yeah, exactly. I think he’s internalized this quite deeply. There are certainly versions of the world where Meta looks a lot more like Google, in some sense—it’s more slow-moving and bureaucratic and just looks very different as a company. But because he’s internalized some of these lessons so deeply, he keeps the thing dynamic, moving, and alive.
15. From founder to coach
Talk me through the difference between how you were as the founder running Scale and how you’re approaching your work now, refounding this lab and working with Mark. What are the different ways that you approach your work in those 2 environments at Scale?
One of the things I really internalized—and I don’t even remember where this advice came from—was this idea that if you want to be a good founder, you have to be able to do every job. I had a very strong point of view about how every little thing at the company should happen and should work. That resulted in a company that was very much a product of my effort and my point of view. It expressed itself in all sorts of ways throughout the company, but it also made me the bottleneck in very big ways.
There are pros and cons to that for a company. I definitely think that if you were to start a company and it’s your first time starting one, you should take that approach, because it’s much more likely to be successful than if you don’t.
Scale was your first company?
Scale is my first company. But at Meta, it was a very different assignment, because we had to accomplish so much in so little time. I didn’t have the luxury of being able to “do every job” within the lab. I certainly am not an AI researcher. I’m also not a product designer. I’m not all these things that are totally pivotal to the organization being successful.
At Meta, I very much adopted this philosophy around creating an environment where we could hire incredible people. Job number 1 was to hire brilliant and incredible people who were at the top of their field across every discipline.
Why do you think you’re able to do that at Meta? Is it just because you have more resources? Why can you do that?
I think one of the things that was charismatic and attractive about the founding moment of Meta Superintelligence Labs was that there was an opportunity to build a lab from scratch with a really, really small team. We had a strong point of view about where we were going: personal superintelligence. I think that resonated with people.
Everyone who works on AI wants to help people. They want it to ultimately mean something to their mom, their grandma, or their grandpa. They want it to be something meaningful to every single person in the world.
Because we were doing this with Meta’s resources, there was an opportunity to not be bottlenecked by compute, infrastructure, or distribution. We weren’t going to be bottlenecked by the things that startups often are. But it was an opportunity to really build something that you could put your stamp on, have a lot of influence over, and shape. I think that was quite exciting for a lot of people.
And so that’s why he chooses to spend what looks like an insane amount of money getting a handful of top talent, because he knows that those top people will recruit other top talent in turn. But you need that as a starting point. Correct?
Yeah. I think one thing that’s maybe somewhat underappreciated about that specific part is that the stock prices of OpenAI and Anthropic ran up a lot. If you looked at how much these people were making by staying at OpenAI or Anthropic, it was also a lot of money. So, for a lot of the talent, it ended up being kind of net neutral in many ways, comp-wise.
Obviously, I think it’s the same thing: oftentimes, a lot of us don’t think about how much the early employees or early researchers at these labs that have had immense stock-value appreciation are making.
We started this fully on the belief of having and hiring the most brilliant people. That’s 1 pillar. Personal superintelligence, that’s 2. How do we create the environment where people do their best work at Meta and in MSL?
I’m doing much more of what you might call traditional management, so to speak. These are the things I’m thinking a lot about: how do I create the environment where all these brilliant people, who are exceptional and very, very special, can all operate and achieve the greatest expression of their talent?
I’m thinking more like a coach in some ways. At Scale, it was very different. I was the person whose point of view had to be expressed in everything that we did.
Say more about that. You feel like a coach?
I think my job is to help set the north star and the directionality of where we want to go—which is, again, personal superintelligence, winning in consumer AI, and building AI that means something to every person—and then spotting special talent on the team and giving people the environment and resources to express that talent maximally.
I think one thing that big companies often mess up is that there are exceptional people at most big companies, but they’re not empowered to do their best work. That’s why they often leave to go to smaller companies, startups, or wherever else it might be.
You mentioned Jeff Bezos’s shareholder letters earlier. That’s something he would repeat as well, in all the books I read about him and in his shareholder letters: if you have great people who can’t build, they’re going to leave.
Yeah, exactly. AI is such a special thing. This term “research” now feels like it almost doesn’t mean anything, but by definition, humanity doesn’t know the limits of what these models are capable of or what you can do with them. They really are these unknown artifacts that we don’t understand super well.
It really is science and research. We’re exploring the limits of what you can do with this technology, what can be built, and it’s a process that requires lots of trial and error. It requires brilliant people to have incredible insight. You can’t think about this process as anything other than a very intellectually intense process. People need space, time, and resources to do their best work.
I think about this in many dimensions. Nat Friedman is absolutely brilliant, and as I mentioned, so much of MSL—if there were a single person whose single point of view came through on MSL, it’s Nat. It’s important for him to be in a position where he can fully express all of that.
Then we have some brilliant researchers who have very, very exciting ideas. How do we create an environment where they’re able to explore those ideas? Most of them won’t work out. Some of them will work out incredibly well. The cases where they do work out justify all the investment.
There are 2 kinds of problems, so to speak, in that analogy. It’s like diamond mining or building skyscrapers, and it depends on what the distribution of the payoff is for the things that you’re working on. There are some areas where the distribution is a super power law: most of the things you do will be kind of useless, but some of the things you do can be literally a million times more valuable than everything else.
VC is a great diamond-mining kind of industry, to give you a sense. It’s all about being able to identify the special ideas and see those through.
16. The future billionaires at Alexandr's 19th birthday party
And that’s easy to do. Just look at your 19th birthday party.
Yeah. I mean, that’s unbelievable.
For people listening who don’t know what the hell I just referenced, can you talk, real quick, about who was at your 19th birthday party and why that would have been a good bet if they had just invested—everybody around the table?
This is a really crazy thing to think about.
My 19th birthday party. I was a freshman at MIT. It was in January, and I was doing a winter internship at a trading firm called HRT. There were 10 or 12 of us, something like that, almost all of whom were students at either Harvard or MIT. Most of us already knew each other from various math, science, or computer science competitions.
Some of the other people there were Jeff Yan, who went on to start Hyperliquid, which is very successful; Scott Wu, who went on to start Cognition, which is very successful; Jesse Zhang, who went on to start Decagon, also quite successfully; and Vicky Ye, who is, I think, a brilliant researcher who works at Anthropic. Everyone was incredibly talented. I actually talked to the founder of HRT recently, and HRT itself, by the way, is now crushing it. Unbelievable.
What does HRT do?
HRT is a proprietary trading firm. They trade their own money. Q2 of um 6 had 11 billion of revenue and like I think 8 billion of profit. That’s the record number. Fewer than 1,000 people work at HRT. It’s unbelievable.
I recently met up with the founder of HRT, and they were talking about, “Oh yeah, we’re starting to do some private investments.” I was like, “Just invest in the interns. All the interns get seed checks.” That’s great.
I grew up doing math, science, and computer science competitions and Olympiads, and got to know all sorts of people from all around America doing this kind of stuff. It’s quite surreal to see the people I did competitions with become so successful.
It’s going to be fascinating to see what all of you do over the next few decades.
Before I interrupted you, did you have more to say about diamond mining and skyscraper stuff? Basically, is it a power-law payoff, where some things just pay for literally a million times the investment? Or is it more like building skyscrapers, which I would describe data annotation or data businesses as being more like?
Amazon Prime deliveries are another example, where the payoff is pretty linear and you have to develop a process where you’re really, really amazing at doing it. If you’re building skyscrapers, then it’s a very operational business, and it’s all about squeezing every ounce of efficiency out of that process. You look at it and figure out, “I can make it 2% more efficient this way, 1% more efficient that way, and half a percent more efficient that way,” and you just try to squeeze every ounce of optimization out of it.
Then there’s diamond mining, where you’re kind of like, “Hey, everyone, just go forth, try all your crazy ideas, and we’ll nurture the ideas that seem promising.”
That’s what you feel you’re doing at Meta right now?
That’s what I feel I’m doing at Meta. And building skyscraper stuff, generally speaking, is what I felt I was doing at Scale. That’s this big paradigm difference.
17. Why Alexandr has more than 200 direct reports
Going back to the coaching analogy and the difference in how you’re essentially managing now at Meta compared to the way you managed at Scale, why do you have so many direct reports? You have, what, 100 direct reports? How many direct reports?
Yeah, I think more than 200.
How and why?
Going back to it, one of the founding theses of MSL and what we want to do in rebuilding this lab was: How do you have really high talent density? How do you make it the best place for a lot of these brilliant people to do their best work? How do you build an organization that’s technically focused, where the technical people are doing their best work and have the opportunity to really do their life’s work?
One of the design principles around this was making it very flat. Effectively, all the researchers we hired into this group called TBD report directly to me. It’s more a statement of the fact that this is anti-bureaucracy. TBD is literally anti-bureaucracy. The point is, we’re hiring brilliant people, and you’ve all demonstrated clearly that you can do brilliant work. You’ve done it before in your careers, so I don’t need to micromanage you to do brilliant work. You’re going to be able to do that on your own.
We need technical leadership, so some of you are going to be responsible for setting clear technical direction for your groups. We have this pod structure, where we have various pods and various technical leads of the pods who tie-break on the technical decisions. There’s a technical leadership structure in place, but from a quote-unquote people-management perspective, or from a bureaucracy perspective, we’re just anti-bureaucracy.
The other part of this is that we do make a lot of group decisions. We want to debate because it’s a very talent-dense group of brilliant people. We want to have those conversations where we’re all discussing what we think the right path is. Everyone in the lab has done brilliant things. We want to hear all these opinions and arrive at things that we all believe are the best path forward.
It’s certainly an unconventional way to run the team, and I would be lying if I said I was the best manager to more than 200 people. But part of the point is that they don’t need managers; they need a great environment. They’re all brilliant and extremely capable. They just need the room to cook.
Yeah. So maybe not a coach. You’re kind of the steward of a great environment so they can do great work.
Yeah. Exactly.
Alex, this was awesome, man. Thanks for taking the time.
Yeah.