One Hour Before the Board Meeting, the Muse Prototype Broke
Mona Sarantakos speaks publicly for the first time about Muse: the Raspberry Pi weekend, the seven-month delay, the fixed team of ten, the placeholder Jolly, and why the system prompt can't be handed over to AI.
One hour before Muse was first demoed to the board, the prototype broke.
These are details Muse product lead Mona Sarantakos shared publicly for the first time on Lenny Rachitsky's podcast. Muse is Meta's AI consumer product; after launch it shot to No. 1 on the App Store and was called by Lenny the most successful AI consumer product of the year. Mona is VP of Product at Meta Superintelligence Labs and has been at Meta for ten years.
The story starts with an internal message. At the end of January, Nat Friedman messaged a few people at Meta saying he'd had a new realization from using OpenClaw and the team had to try it. Mona couldn't find a Mac Mini, so over the weekend she installed it on a Raspberry Pi and gave it access to email, calendar, and home automation. Her husband said that was way too unsafe.
On Monday morning, it summarized what was happening that week at her child's school, pulled from school emails she never kept up with. She replied to Nat: This is it. I'm in.
A commenter in the replies pointed out the key: it worked that weekend because she handed over her own credentials; the next seven months were about turning permissions into scopes other people could revoke.
The prototype for the board demo was built by design lead Alex Cornell in a few days. The team decided at the last minute to show it. One hour before the meeting, the prototype crashed. Alex and an engineer rushed to fix it, and it only came back five minutes before the start. Mona said, in another universe we didn't pull it off, and maybe Muse would never have been born.
Jolly was originally just a placeholder. Alex created the character for the first prototype, called Jolly Bot, inspired by his nickname for his daughter. Every few months after that, someone would say you can't use a plush toy as a logo. Mona's exact words were that Jolly is like our cockroach—it just won't die. One week before launch, the team and Mark Zuckerberg made the call: Jolly stays. Mona's reasoning was direct: when you're talking to something that feels very personal, the last thing you want to be talking to is a corporate logo. A commenter summed it up: placeholders always win the logo debate.
Muse took seven months from start of work to launch. It began earlier than all AI assistants. After the product definition was clear and the end-to-end version was done in a few months, all the remaining time went into reliability, safety, and privacy. The team's internal motto is 'Trust is a must.'
The initial core team was only 10 people. The team deliberately stayed small to keep communication and decision-making tight. They worked in a library-like common area in a building in Menlo Park and ate all three meals together every day. Mona said it was like bottling lightning, and she couldn't imagine this product being built by another team.
Shopping became Muse's most unexpected use case. Mona had predicted personal productivity, health, and personal finance. Others bet on home automation, because it was hot among OpenClaw users. Then internal testers started shopping. Engineering manager Christine Sed became the top dogfooder, sending shopping bug reports every morning. Mona said all our evals were elsewhere, so the team had to start building shopping evals.
They used 'agent swarms' to find user insights at scale. Muse conversations are long and multi-turn, and existing tools couldn't analyze them. The team put custom groups of agents into anonymized conversations and feedback to report what worked well and where people got stuck—down to 10 people being blocked while trying to shop on a certain website. Mona said, this makes me wonder how products were ever built before.
A commenter asked how the things reported by the swarm were scored, and whether humans still needed to sample and see what it missed. That gets exactly at a hard rule in Muse's approach: agents are used everywhere, but humans always review the data.
The system prompt has to be written by humans. Mona said you cannot use AI to write the system prompt—any part of it. The team uses agents heavily elsewhere, including helping write evals, but humans always look at the data, because when they don't, things always go wrong. The closer you get to the master switch for behavior, the less you can outsource it.
When hiring PMs, Mona looks first for product taste. She agrees that the PM role has changed more in the past two years than in the previous twenty. Now she can come up with an idea, build it herself, and have real discussions on a 24-hour cycle. What really distinguishes a PM is representing the customer, holding a high bar, and being able to tell an engineer who has fallen in love with their own idea: this is cool, I'm glad you really like it, but I don't think anyone else will. A commenter said this is the hardest one.
Muse's next step is ambient presence. In the near term, Mona wants to cut the babysitting. She said, if it gets to the point where I have to supervise an agent, I'd rather just do it myself. She wants Muse to move beyond the app and into glasses and other devices. But she doesn't pretend to know the endpoint: a year ago, I would never have explained Muse to you the way I'm explaining it today.
The full conversation is from Lenny's Podcast; the video and link are below.
发布时间: 2026-11-17 08:34