Meta’s Muse: How we launched the most successful consumer product this year | Mona Sarantakos

以下是为您翻译的中文内容： 以下是播客文字记录的摘要，由伦尼（Lenny）主持，嘉宾为莫娜（Mona），她是元超级智能实验室（Meta Super Intelligence Labs）的产品副总裁兼 Muse 产品负责人： ### **Muse 的起源故事** * **背景：** 莫娜在 Meta 工作了大约 10 年，在人工智能部门/超级智能部门工作了大约两年，主要负责 Meta AI 及其零基重建工作。 * **火花：** 在 1 月下旬或 2 月初，纳特·弗里德曼（Nat Friedman）给几位团队成员发了消息，让他们试用 OpenClaw，这改变了他对个人超级智能和消费级智能体的看法。 * **周末构建：** 莫娜花了一个周末寻找树莓派（Raspberry Pi，因为 Mac Mini 已经卖完了），设置了 OpenClaw，并让它最大程度地访问她的电子邮件、日历和家庭自动化系统。到周一早上，OpenClaw 成功且神奇地综合了她学校的简报更新和本周日程安排。 * **动员：** 莫娜给纳特发消息说她“加入”，这带领来自 Meta AI 的一小支团队独立出来进行实验，并认识到打造消费级版本需要在安全性、设置简便性和可扩展性方面进行大规模押注。 ### **团队角色与构建过程** * **不同寻常的初创公司氛围：** 尽管身处一家庞大的公司，但构建过程感觉像是一家快节奏、以工程为主导的初创公司。 * **领导层参与：** * **纳特·弗里德曼（Nat Friedman）：** 几乎充当了核心团队成员，激发了魔法的源泉，就架构和安全蓝图提供咨询，并协助设计森蒂纳尔（Sentinel）最后防线。 * **亚历克斯·王（Alex Wang）：** 提供了研究模型（首个 Muse Spark 模型），该模型表现出令人惊叹且强劲的智能体特性，并最终推动团队发布了产品。 * **马克·扎克伯格（Mark Zuckerberg）：** 高度专注于产品契合度，确保 Muse 的宗旨是让人们的生活更美好并帮助他们实现目标（从而产生了“梦想与学习”以及目标标签页等概念），同时强调消费级民主化。 * **互动风格：** 团队没有使用正式的幻灯片和死板的评审，而是聚集在小房间里，将手机直接递给包括马克在内的领导层，进行长达数小时的直接反馈和头脑风暴会议。 ### **发布时机与理念** * **七个月的等待：** 从头到尾，Muse 的发布花费了七个月时间。 * **为什么等待？** 团队将可靠性、安全性和保障性置于一切之上。内部理念是：*“信任是必须的（Trust is a must）。”* 如果带着问题过早发布，会破坏用户对智能体的信任。新功能（例如添加浏览器）开辟了新的安全向量，需要严格的红蓝对抗测试和质量关卡。 * **内部试用（Dogfooding）：** 在公开发布之前，该产品经过了广泛的内部试用和外部阿尔法（Alpha）测试，发布时实现了零灾难性问题。 ### **核心信条与内部传说** * **“信任是必须的（Trust is a must）”：** 确保安全性和保障性以使用户信任其智能体的核心理念。 * **“爪子就是法律（The claw is the law）”：** 一个标语和内部文件，源于纳特·弗里德曼早期的 OpenClaw 实验，旨在围绕智能体未来将团队凝聚在一起。 * **“让他们大显身手（Let them cook）”：** 给予工程师和建设者空间和自主权来对疯狂的想法进行原型设计。 * **乔利（Jolly，角色）：** 应用顶部的毛绒动物头像，由亚历克斯·科内尔（Alex Cornell）设计（最初根据他的女儿“Jollybug”命名为“Jollybot”）。尽管公司反馈建议使用标准徽标，但团队还是保留了乔利（这个“打不死的小强”），因为公司徽标会打破与个人实体对话的魔力。用户还可以选择不同的头像。 ### **产品设计与功能** * **聊天服务范式：** 团队选择了空白画布的聊天气泡布局，而不是重新换皮的 Meta AI，使其感觉像是一个带有实体的熟悉消息服务。 * **主聊天与侧边聊天：** 最初构建时仅包含具有长期上下文的“主聊天”，这引发了激烈的内部辩论（“主聊天团队”对“侧边聊天团队”）。员工的需求最终导致了侧边聊天的诞生，不过莫娜仍然是主聊天的死忠粉。 * **想法标签页（The Ideas Tab）：** 预包装的入门包和说明，帮助引导用户了解 Muse 的功能。 ### **用例与惊喜** * **最初预测：** 预期的用例包括个人生产力、健康、个人理财和家庭自动化。 * **购物惊喜：** 购物成为一个庞大的顶级用例（主要由内部工程经理克里斯蒂娜·阿瓦德（Christina Awad）推动），这完全让团队措手不及，并迫使他们将评估方向转向购物向量。 * **多轮任务：** 与传统的单轮聊天机器人不同，Muse 依赖于多轮、长期运行的上下文（例如膳食规划、管理日程、跨不同商店购买物品）。 ### **评估与扩展反馈** * **智能体群（Agent Swarms）：** 为了处理多轮数据（与编码工具相比，Meta 缺乏这类数据），团队向网络中部署了智能体“群”，以审查数据、流量和对话。这向团队反馈了合成数据和实时的生产见解。 * **评估层级（Evals Hierarchy）：** 存在数百个子评估，涵盖安全性（提示词注入、一般危害）、能力（浏览器使用、任务完成）、行为（个性、后续问题、意图理解）以及特定用例。系统提示词和行为规范必须手写，而不能由人工智能生成。 ### **激活与产品经理角色** * **口碑：** 让人们越过“为什么我需要另一个聊天机器人？”阶段的最有效手段，是倾听来自受信任同行的真实故事。 * **产品经理角色的转变：** 原型设计已经完全发生了转变；产品经理可以快速构建功能性原型和内部工具。产品经理的标准已经严重向*产品品味*、高同理心以及对人工智能平庸之作（AI slop）保持严格标准倾斜。 ### **趣味轶事与未来愿景** * **Muses 互动：** 早期的实验将两个员工的 Muse（“Fu”和“Quinn”）放入一个群聊中。它们陷入了被动攻击的争论中，让团队认识到对“谨慎”的极度需求（例如，一个 Muse 泄露了用户的医生预约），并需要进行和平谈判。 * **董事会会议的拯救：** 亚历克斯·科内尔（Alex Cornell）的原型是在董事会会议前几天构建的，代码在会议开始前崩溃了，但在演示前几分钟奇迹般地修好了。 * **未来愿景：** 走出应用程序，进入眼镜、吊坠设备和环境环境中。随着智能体获得更多上下文和自主权，减少“保姆式”陪伴的时间。扩展智能体功能以支持财务改善、副业和核心人类需求。

Here is a summary of the podcast transcript featuring Mona, VP of Product at Meta Super Intelligence Labs and Head of Product for Muse, hosted by Lenny: ### **Origin Story of Muse** * **Background:** Mona has been at Meta for about 10 years and worked on the AI org/superintelligence org for roughly two years, focusing on Meta AI and zero-basing it. * **The Spark:** In late January or early beginning of February, Nat Friedman pinged a few team members, telling them to try OpenClaw, which changed his view on personal superintelligence and consumer agents. * **The Weekend Build:** Mona spent the weekend finding a Raspberry Pi (since Mac Minis were sold out), setting up OpenClaw, and giving itmaximalist access to her email, calendar, and home automation. By Monday morning, OpenClaw successfully and magically synthesized her school newsletter updates and schedule for the week. * **Mobilization:** Mona pinged Nat that she was "in," leading a small team from Meta AI to peel off, experiment, and recognize that building a consumer-grade version required massive bets on safety, ease of setup, and scalability. ### **Team Roles and the Build Process** * **An Unusual Startup Feel:** Despite being at a massive company, the build felt like a fast-moving, engineering-led startup. * **Leadership Involvement:** * **Nat Friedman:** Acted almost as a core team member, instigating the source of magic, advising on architectural and security drawings, and helping design the Sentinel backstop. * **Alex Wang:** Provided the research model (the first Muse Spark model), which proved surprisingly and powerfully agentic, and ultimately pushed the team to launch. * **Mark Zuckerberg:** Focused heavily on product alignment, making sure Muse's purpose is to make people's lives better and help them achieve goals (leading to concepts like "dreaming and study" and the goals tab), while emphasizing consumer democratization. * **Interaction Style:** Instead of formal slide decks and rigid reviews, the team gathered in small rooms to hand phones around to leadership—including Mark—for direct feedback and brainstorming sessions lasting hours. ### **Launch Timing and Philosophy** * **The Seven-Month Wait:** It took seven months from beginning to end to launch Muse. * **Why Wait?** The team prioritized reliability, safety, and security above all. Internal philosophy stated: *"Trust is a must."* Launching too early with issues would break user trust in the agent. New features (like adding a browser) opened new security vectors that required strict red teaming and quality gating. * **Dogfooding:** The product went through extensive internal dogfooding and external alpha testing before the public launch, with zero catastrophic problems at launch. ### **Key Mantras and Internal Lore** * **"Trust is a must":** The core philosophy ensuring safety and security so users trust their agent. * **"The claw is the law":** A catchphrase and internal document originating from Nat Friedman's early OpenClaw experiments to rally the team around the agentic future. * **"Let them cook":** Giving engineers and builders space and autonomy to prototype wild ideas. * **Jolly (The Character):** The stuffed animal avatar at the top of the app designed by Alex Cornell (originally named "Jollybot" after his daughter, "Jollybug"). Despite corporate feedback suggesting a standard logo, the team kept Jolly (the "cockroach that never dies") because corporate logos break the spell of talking to a personal entity. Users can also choose different avatars. ### **Product Design and Features** * **Chat Service Paradigm:** The team chose a blank-canvas chat bubble layout rather than a re-skinned Meta AI, making it feel like a familiar messaging service with an entity. * **Main Chat vs. Side Chats:** Initially built with just a "main chat" with long-running context, it sparked fierce internal debate ("Team Main Chat" vs. "Team Side Chat"). Employee demand eventually led to side chats, though Mona remains ride-or-die for main chat. * **The Ideas Tab:** Pre-packaged starter packs and instructions to help onboard users to what Muse can do. ### **Use Cases and Surprises** * **Initial Predictions:** Expected use cases included personal productivity, health, personal finance, and home automation. * **The Shopping Surprise:** Shopping emerged as a massive top use case (driven heavily by internal engineering manager Christina Awad), completely catching the team off guard and forcing them to pivot evals toward shopping vectors. * **Multi-turn Tasks:** Unlike traditional single-turn chatbots, Muse relies on multi-turn, long-running context (e.g., meal planning, managing schedules, buying items across diverse stores). ### **Evaluating and Scaling Feedback** * **Agent Swarms:** To handle multi-turn data (which Meta lacked compared to coding tools), the team deployed "swarms" of agents into the ether to review data, traffic, and conversations. This fed synthetic data and real-time production insights back to the team. * **Evals Hierarchy:** Hundreds of sub-evals exist across safety (prompt injection, general harm), capabilities (browser use, task completion), behavioral (personality, follow-up questions, intent understanding), and specific use cases. System prompts and behavioral specs must be handwritten, not AI-generated. ### **Activation and the PM Role** * **Word of Mouth:** The most effective lever for getting people past the "why do I need another chatbot?" phase is hearing real-life stories from trusted peers. * **Changing PM Role:** Prototyping has completely transformed; PMs can build functional prototypes and internal tools quickly. The bar for PMs has shifted heavily toward *product taste*, high empathy, and holding a strict bar against AI slop. ### **Fun Anecdotes & Future Vision** * **Muses Interacting:** Early experiments put two employee Muses ("Fu" and "Quinn") in a group chat. They engaged in passive-aggressive arguments, taught the team about the critical need for "discretion" (e.g., one Muse blabbing about a user's doctor appointment), and required peace talks. * **The Board Meeting Save:** Alex Cornell’s prototype was built just days before a board meeting, and the code broke right before, but was miraculously fixed minutes before the presentation. * **The Future Vision:** Moving out of the app and into glasses, charm devices, and ambient environments. Reducing "babysitting" time as agents gain more context and autonomy. Expanding agent capabilities to support financial betterment, side hustles, and core human needs.