Muse 产品负责人 Mona Sarantakos 分享 7 条幕后心得
My biggest takeaways from @mona_sarantakos, Head of Product for @Muse, the most successful AI consum...
这篇文章把 Muse 从 Nat Friedman 一条消息到上线的全过程拆给你看:10 人团队、7 个月、被保留的吉祥物 Jolly。做 AI 产品的人能直接抄作业,尤其是他们用 agent swarm 分析用户对话、临时补建购物 evals 的做法,很实用。
Lenny Rachitsky 总结了与 Meta 旗下 AI 消费级产品 Muse 产品负责人 Mona Sarantakos 对谈的 7 条要点。Muse 从立项到上线用了 7 个月,核心团队仅 10 人,刻意保持小规模以便沟通决策,团队在 Menlo Park 一处类似图书馆的公共空间办公,一日三餐都在一起。
产品起源是 1 月底 Nat Friedman 在内部发的一条消息,称 OpenClaw 改变了他对可能性的看法。Mona 用一个 Raspberry Pi(当时 Mac Mini 缺货)搭好原型并接入她的邮箱、日历和家居自动化,周一早上它汇总了她一直没跟进的学校邮件,她说"This is it"。第一个原型在董事会演示前一小时崩溃,设计师 Alex Cornell 和工程师在会前 5 分钟修复。吉祥物 Jolly 原本是占位符,以 Alex 女儿的昵称命名,上线前一周团队和 Mark Zuckerberg 决定保留。
产品方面,Mona 原本预判生产力、健康和理财是主要用例,但内部测试者开始高频购物,工程经理 Christine Sed 成为头号 dogfooder,团队为此专门补建购物 evals。由于 Muse 对话长且多轮,没有现成工具能分析,团队用自定义 agent swarm 跑匿名对话和反馈来规模化挖掘用户洞察。
My biggest takeaways from @mona_sarantakos, Head of Product for @Muse, the most successful AI consum...
My biggest takeaways from @mona_sarantakos , Head of Product for @Muse , the most successful AI consumer product of the year: 1. The first Muse prototype broke an hour before the board meeting pitch. Lead designer @AlexCornell built the working prototype in a few days, and the team decided at the last minute to show it at a board meeting. An hour before, it broke. Alex and an engineer raced to fix it, and it came back five minutes before the meeting started. “In some alternate universe we didn’t figure this out and maybe Muse was never born.” 2. Jolly, the now-iconic Muse avatar, was just a placeholder. Alex created the character for the first prototype and called it Jolly Bot, after his nickname for his daughter. Every few months, someone said they could never launch with a stuffed animal as the logo. “Jolly’s like our cockroach. It just never dies.” A week before launch, the team and Mark Zuckerberg made the call: “Jolly stays.” Mona’s reasoning: “When you’re talking to something that feels so personal, the last thing you want to talk to is a corporate logo.” 3. Muse began with a ping from @natfriedman . On a Thursday in late January, Nat messaged a few people internally: OpenClaw was changing his whole view of what’s possible, and they needed to try it. Mona spent the weekend setting it up on a Raspberry Pi (Mac Minis were sold out) and gave it access to her email, calendar, and home automation. On Monday morning it sent her a summary of what was coming up that week, pulled from the school emails she never keeps up with. She pinged Nat back: “This is it.” 4. Muse took seven months to launch, even though work started before every other AI assistant. After figuring out what the product was going to be, and building a fully functioning end-to-end version in a few months, the team spent the rest of the time on reliability, safety, and security. Their internal motto was “Trust is a must.” 5. The original Muse team was just 10 people. The core team stayed small on purpose, to keep communication and decisions tight. They worked out of a library-like common area in a Menlo Park building and ate all three meals together each day. “It was like lightning in a bottle. I couldn’t have imagined this product being built with any other team.” 6. The team didn't expect shopping to be one of Muse’s top use cases. Mona predicted that personal productivity, health, and personal finance would be the primary use cases. Others bet on home automation, which was big with OpenClaw users. Then internal testers started shopping. Engineering manager @christinesed became the top dogfooder, sending shopping bug reports every morning. “All of our evals have been somewhere else,” so the team started building shopping evals. 7. The team uses “agent swarms” to surface user insights at scale. Muse conversations are long and multi-turn, and no existing tool could analyze them. So the team sends swarms of custom-built agents through anonymized conversations and feedback. The swarms report back what’s working and what’s breaking, down to the 10 people who tried to shop on one site and got blocked. Mona said, “It makes me wonder how we even built products before.” 8. Muse’s system prompt has to be written by a person. “You cannot use an AI to write a system prompt, any part of it.” The team uses agents everywhere else, including to help write evals. But a human always reviews the data, because “when we don’t, we always have issues.” 9. Product taste is now the first thing Mona looks for in a PM. She agrees that the PM role has changed more in the past two years than in the two decades before. She can now come up with an idea, build it herself, and have a real conversation about it “in a 24-hour cycle.” What sets a PM apart is representing the customer, holding a high bar, and telling an engineer who loves their own idea: “That’s really cool. I’m glad you really like it. I don’t think anyone else will.” 10. The next frontier for Muse is ambient presence. Near term, Mona wants to cut the babysitting (“At the point where I’m babysitting an agent, I might as well do it myself”) and move Muse beyond the app into glasses and other devices. But she doesn’t claim to know where this goes: “A year ago, I would’ve never explained Muse to you the way that I did.” Watch the conversation: youtube.com/watch?v=WY0cdD… Lenny Rachitsky @lennysan In late January, @natfriedman pinged a few people at @Meta . He’d been using OpenClaw, and told them they needed to try it. Mona Sarantakos couldn’t find a Mac Mini, so she spent the weekend setting it up on a Raspberry Pi and gave it access to her email and calendar. Her husband told her, “That is so not safe.” But by Monday morning, it had pulled together everything her kids had going on at school that week. She pinged Nat back: “This is it. I’m in.” The team spent the next seven months building something safe and simple enough for everyone. They called it @Muse . It became the fastest-growing AI consumer product ever. Mona is head of product for Muse and VP of product at Meta Superintelligence Labs. In her first podcast interview, she shares the inside story of how they built and launched Muse. We discuss: 🔸 Why they waited seven months to launch, despite getting started before all of their competitors 🔸 The two levers that actually move activation for an AI agent 🔸 How @natfriedman , @alexandr_wang , and @finkd each left their imprint on the product 🔸 Why the Muse team relies on "agent swarms" more than dashboards 🔸 Why taste is now the most important PM skill Listen n youtube.com/watch?v=WY0cdD… 67p3pT 🔗 View Quoted Tweet📊 11 ⚡
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Muse, Meta, Nat Friedman, OpenClaw.
Source attribution: curated from Lenny Rachitsky (@lennysan).