How Meta Built Muse: Trust First, Jolly Included
Lenny's conversation with Mona Sarantakos on launching Meta's Muse: a Raspberry Pi weekend, a trust-first build, and the stuffed avatar nobody could kill.
Product update8 min read
The consumer AI product Lenny calls the number-one app in the app store — the one Zuck has hinted is the fastest-growing AI consumer product of all time — started with a borrowed Raspberry Pi and a stuffed animal nobody was allowed to kill. That sentence sounds like a joke. In Mona Sarantakos's telling, on Lenny's Podcast, it is the operating history of Meta's Muse: a weekend experiment that became a company bet, built by a small team that chose trust over speed and an avatar over a logo.
Sarantakos is VP of product at Meta Superintelligence Labs and head of product for Muse. She had been at Meta about ten years, two of them in AI, and was in the middle of zero-basing the old Meta AI assistant when everything changed. What follows is her account of how.
The weekend on the Raspberry Pi
The trigger was a ping. Toward the end of January or the beginning of February, Nat Friedman messaged a few people that he was trying OpenClaw, and that it was changing his view of personal superintelligence. Sarantakos had written in an H1 brief that this sort of thing would come to a head at the end of 2026, maybe 2027. The message collapsed that timeline to now.
That Friday she told her husband she needed a Mac Mini. The Apple store had none, so he dug out a Raspberry Pi somewhere. The kids were away skiing, so she spent the whole weekend talking to her own OpenClaw — email access, calendar access, home automation. Her husband's verdict, in her retelling: that is crazy, that is so not safe. Her answer: I'm going maximalist.
Monday morning did the convincing. A newsletter her agent had assembled overnight pulled together the school portals and newsletters she never reads — "spirit week", the special days, what to expect today. A tiny prompt, she says, and for the first time she felt on top of her personal life on a Monday morning. She pinged Friedman back: this is it, I'm in. My read: the product was sold to its own builder before it existed, on one concrete morning, not on a vision deck.
"Trust is a must"
The team's internal saying, in her words: trust is a must. Launch too early with issues, and people will never trust the agent — and if they don't trust it, they will never use it. So the consumer build had to be very different from the weekend hack: way safer, way easier to set up, scalable.
That principle shaped the machinery. Priority safety evals gate everything that ships; secondary and feature-level evals sit underneath. PMs, engineers, and researchers all write evals, agents help generate them, but people review all the data — "when we don't, we always have issues," she says. And one rule is absolute: the system prompt and the behavioral spec must be handwritten. No part of them may be written by AI.
Zuckerberg's contribution, in her telling, was direction, not deadlines. His framing: Muse exists to make people's lives better and help them achieve their goals — alignment with the user's priorities, not task completion for its own sake. That unlocked what the team called dreaming and studying: a 24-7 agent that thinks about you and your priorities in the background, plus an explicit goals tab in the UI. And because the CEO in question is, in her words, a consumer CEO thinking about democratization, scale and affordability were on the roadmap from the start. Alex Wang, she says, was the one who finally pushed them out the door: "guys, I think it's ready. Just go."
Jolly the cockroach
The mascot nearly died a dozen times. Lead designer Alex Cornell — the same designer behind the floating emoji on Facebook Live — had put a one-shot creature into the first prototype. It was called JollyBot because he called his daughter JollyBug. It was meant to be temporary; everyone assumed corporate marketing would eventually demand a logo.
Instead Jolly became, in Sarantakos's words, "our cockroach — it just never dies." Every few months someone declared there was no way to launch with a stuffed animal at the top. A week later the team was keeping it. Her argument: when you talk to something that personal, the last thing you want is a corporate logo — "that breaks the spell completely." Employees making their own avatars sealed it; the system grew around the creature until removing it was unthinkable. Zuckerberg himself, she says, loved the avatar. The decision landed about a week before launch. My read: the cutest thing in the product survived because it did structural work — it told the user this thing is yours, not Meta's.
The small team on the roof
Muse was built by a small team tucked on the roof of a Menlo Park building, doing stand-ups in a carpeted common room called the library, on floral couches that look, she says, like they were bought in 1995. The culture she describes is builder-first: one or two engineers disappear for a few days with a kernel of an idea and emerge with a prototype. The internal mantra for that is "let them cook."
The stories she tells are all speed plus improvisation. Cornell built the first working prototype in days; it ran on his phone as a local build. Days before a board meeting, the team decided at the last minute to show it instead of the planned pre-read — then it broke an hour before the meeting and got fixed five minutes before it started. Later, after a canceled flight on the way to the World Cup, two of their Muses negotiated with the airline's agent and got the refunds. And the use case that surprised everyone was shopping: the team had predicted productivity, health, finance, and home automation, but shopping kept popping up in dogfooding until they rebuilt their evals around it. Engineering head Christine Wadd shops so much through Muse, Sarantakos jokes, that she wants to take her credit card away — a line her own husband echoed when seven packages landed on the doorstep.
Word of mouth beats ads
Activation, she says flatly, is the number-one problem: "Why? I already have a chatbot. I don't get it." The most effective lever, in her words — "that's 100% the most effective lever" — is other people's stories. She watched a power-user mom pass her phone around a group of parents, two of whom downloaded Muse on the spot. No amount of advertising, she says, did what one relatable demo did. It rhymes with her own ChatGPT memory: friends at dinner going "can you believe it did that?"
The second lever is the first prompt. People freeze — "what's the weather today? Hi. What do I say to this thing?" The answer is the Ideas tab: pre-packaged starter packs (a cron notification, a dashboard, "let's do it") that show rather than explain. And onboarding now splits by cohort, because early adopters wanted to push the safety boundaries while later users just ask why this differs from every other AI product. Even the chat structure was fought over: one long-running main chat versus threads split employees into warring camps ("team main chat, team side chat"), until side chats shipped while the agent stayed aware of everything. The team, not the executives, made that call.
Ambient, not app-bound
Sarantakos is unusually candid about the fear that comes with scale. A nurse, a single mom of two, runs meal planning, doctor searches, appointments, and emails through her agent — "I hope it's not fucking up," Sarantakos says. Non-determinism breaks a PM's instinct for control: no fixed onboarding flow, just guardrails, evals, and an agent that reasons it out. No two Muses, she notes, are really the same.
The vision she sketches is ambient: out of the app, into glasses and charm-like devices, affordable enough to be an equalizer. Less babysitting — she compares today's agent-tending to feeding a Tamagotchi — more autonomous action as context and trust compound. And more than time saved: help getting healthier, more financially secure, even starting the side hustle. The closest thing to a roadmap request she makes is feedback itself: tell your Muse what's working, it sends the team a report. A neighbor, she says, files weekly feedback addressed to "Mona, your product manager" by name.
So here is the open question the episode leaves behind. If the winning consumer agent is the one people trust with their inbox, their kids' schedules, and their money — and trust is slower to earn than speed is to ship — how many labs will actually choose the slow path?
More podcast deep dives from Podwise.


