← Return to Index Archived November 15, 2026
The Lead — Nov 15
LENNY'S PODCAST: PRODUCT | CAREER | GROWTH · LENNY RACHITSKY

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

Meta’s head of product traces Muse from a small-team experiment inspired by early consumer agents to a safety-first AI assistant shaped by user trust, chat-based design and an unlikely mascot. She also considers a future where personal agents leave the app for glasses, wearable devices and more autonomous help.

1h 27m / November 15, 2026 /aiproducttechnology / Transcript sourced from openai
All episodes from Lenny's Podcast: Product | Career | Growth →·Episode page →·Podcast website →·Listen on Apple Podcasts →

Overview

Lenny Rachitsky interviews Mona, VP of product at Meta Superintelligence Labs and head of product for Muse, Meta's personal AI agent. She explains how a weekend spent running the open-source agent OpenClaw on a Raspberry Pi turned into Meta's consumer agent, why the team took about seven months to ship it, and how building something non-deterministic changes the product manager (PM) role.

Key Takeaways

The slow launch was deliberate. Mona says the product itself came together by spring, and most of the remaining time went into reliability, safety, and security. Industry benchmarks felt too weak for an agent that holds personal context and can spend money, so the team wrote its own bar and red-teamed against it. Each new capability reopened the problem. Adding a browser, for example, created new attack surfaces. The internal line was "trust is a must," on the reasoning that one bad early incident would end user trust in the agent for good.

Leadership shaped the product through conversation rather than formal reviews. The biggest single change came from Mark Zuckerberg, who reframed Muse around helping people reach their goals instead of completing tasks. That reframe produced the "dreaming" feature, where the always-on agent thinks about the user in the background, and the Goals tab.

Several choices that looked unprofessional turned out to matter. Jolly, the stuffed-animal avatar, was a throwaway prototype that people tried to remove for months. Mona argues that a corporate logo would break the sense that the agent belongs to the user. The team also chose chat bubbles over the long-form canvas used by most AI products, and a single main chat over threaded conversations. Employees split hard on the main chat question, and side chats were eventually added.

Usage surprised them. Mona expected productivity, health, and personal finance to lead. Shopping became a top use case instead, and home automation, which was big among OpenClaw hobbyists, faded with a broader audience.

On feedback, "agent swarms" scan conversations and production data and return summaries of what users are trying to do. Mona says this filled a gap. Meta's existing assistant data was mostly single-turn, and the team needed examples of long, multi-turn, non-coding tasks to build evaluations (evals).

She also sees the PM bar shifting toward product taste. Prototypes are cheap and engineers are product-minded, so the PM's job is to represent the user and block what she calls "AI slop" from shipping.

Practical Steps

  • Sort evals into tiers: must-pass (safety, prompt injection), capability, behavioral (does it ask follow-ups, does it understand intent), and use-case evals. Allow small regressions only in the lower tiers.
  • Write system prompts and behavioral specs by hand, and have people review eval data so the prompt distribution matches real usage.
  • Once a use case shows up in dogfooding, build evals for it, even if it wasn't on your roadmap.
  • Help new users past the first prompt. Muse's Ideas tab offers starter packs that the agent sets up after a single "let's do it."
  • Vary onboarding by how much experience a user has with agents. What worked for early adopters didn't work for later groups.
  • Give engineers a few days alone with a rough idea before assigning a PM. Mona's team brought PMs in once a prototype was close to shipping.

Notable Quotes

"When you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo." (Mona)

"If we launch too early and we have a bunch of issues, people will not trust this agent. If you don't trust the agent, you're never going to use it." (Mona)

"At the point where I'm babysitting an agent, I might as well do it myself." (Mona)

If we launch too early and have a bunch of issues, people will not trust this agent, and if you don't trust the agent, you're never going to use it. — From the episode

Full Transcript

Source: openai 1h 27m runtime

You are head of product for Muse. Every PM's dream is to work on the number one app in the App Store. I've never experienced a product build like this before. We would give Mark our phone and be like, What do you think? And he'd play around with it and give us feedback. You have been there from the beginning. Nat Friedman pinged a few of us. He's like, I'm using the OpenAI. You guys need to get on it. I spent all weekend talking to my OpenAI. It felt pretty magical with such little direction from me. The more we were using the product, it was like, okay, if we want to build it for our consumer base, it has to be really different. It's going to have to be way safer. It felt like y'all were way ahead of this, but it took you a long time to launch something. There's a saying we have internally: trust is a must. If we launch too early and we have a bunch of issues, people will not trust this agent. If you don't trust the agent, you're never going to use it. The little character Jolly, not something you'd expect from a company like Meta, like a massive company. Jolly is like our cockroach. It just, like, never dies. Every few months, somebody would be like, There's no way you're going to launch this product with a stuffed animal at the top. But I think when you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo. What's next? What's kind of the vision of where this heads? Right now it's sort of trapped in an app. Soon it should be in your glasses, on your charm devices, whatever that might be. What a time to be alive. When I was a kid, could you imagine? Like, I had to go up to the TV and turn the TV dial to be able to change a channel, and now I am talking about personal super intelligence in your glasses. Mona, thank you so much for being here, and welcome to the podcast. Thank you, Lenny. Thank you for having me. You are head of product for Muse. You're VP of product at Meta Superintelligence Labs. Muse has kind of come out of nowhere. I feel like it's—I don't know what the numbers are, but it feels like the most successful AI assistant out there right now. Zach has hinted that is the fastest-growing AI consumer product of all time. You have been there from the beginning. You've also not been on any podcasts. You've also barely posted anything on socials. You've had like five posts on Threads, like one LinkedIn post. So this is extra awesome. These are my favorite conversations, people doing amazing work that nobody has really heard from. So I have so many questions. I want to start with just kind of the origin story of Muse and kind of your story of getting started on this work. What were you working on before you joined this team? And then what was kind of the pitch they made you? What was kind of the kernel of the idea when they pitched you on taking on this work, joining this team? I was on the, in the AI org, in the super intelligence org before. I'd been working on it for about two years. I've been at Meta for about 10 years, and I was mostly working on Meta AI, which is our assistant product. And we were zero-basing Meta AI at the time, and it's a phrase that we use a lot where we're like, you know, we have something, it's not working quite the way that we want it to, and you could keep optimizing it or iterating on it, or you could just start over. And so we were doing a lot of that with the primary Meta AI product because the lab was working on this new model, which was the first of the Llama Spark series that we were anticipating, and it was awesome. We were having a great time. And then I think on like Thursday, maybe end of January, beginning of February, Nat Friedman pinged a few of us, and he said to us, I'm trying out OpenCLA. It's changing my whole view of what's possible in terms of personal superintelligence and consumer agents, which we had been talking about quite a bit. Obviously, everybody was sort of talking about it in our head. I think I wrote this down actually in like an H1 brief. I was saying that, yeah, we'll work on it probably end of 2026, maybe 2027, right? Like that's kind of the time frame when I think that this sort of thing is going to be coming to a head. And yeah, so that's sort of like where we were at. And then so he sends us this message. He's like, I'm using OpenCLA. You guys need to get on it. So I think on Friday, I told my husband, I'm like, I need to go find a Mac Mini. I need to get OpenClass set up. And we looked at the Apple Store. There were no Mac Minis. So then he was like, I think I have a Raspberry Pi somewhere. So he got his Raspberry Pi. We set it up. It happened to be a weekend. My children are in, they're in a ski race program, and so they were gone for most of the weekend skiing. So it was just me and my OpenClaude. And I spent all weekend essentially talking to it. And, like, you know, at this point, I think my husband came back on Sunday with the kids, and he was like, How's it going? I was like, Oh, I gave it access to my email. I gave it access to my calendar. And he was like, What are you doing? That is crazy. That is so not safe. And I was like, No, no, it's totally fine. I wanted to, like, I'm going maximalist. So this is like a Sunday. I had spent the whole weekend talking to this and setting it up and hooking it up to my home automation, all of this stuff. And then Monday morning, my kids were going to school, and I'm just really bad at personal email and, like, just staying on top of all the school newsletters and the portal. I never check it. Like, so my children are, you know, like, we do our best, but they're often not figuring out spirit week or anything like that, all of these kind of special things. And I had gotten a newsletter from my OpenClaude that I had, you know, kind of set up, and it was tracking all of these different sources, and it sort of pulled together this snapshot of, like, what to expect this week, what to expect today. And it was like, I remember looking at it now in retrospect, it's like, okay, a lot of people can do that. Even assistants can do that. But at the time, it felt pretty magical and with such little direction from me. I had to, you know, it was like such a small prompt that I had written, right? Maybe like a few words. And the idea that I could kind of suddenly feel really productive and on top of my shit, which I never felt on a Monday morning, right, particularly in my personal life, was, like, incredible for me. And I looked at it and I was like, You know, how many busy parents are there? How many busy people are there? How many people feel like this, right? And so anyway, so I pinged Nat, and I was like, This is it. I think we gotta, you know, I'm in, right? Like, I don't know what I'm in for, but, like, nobody—we had never done, like, an edict of, like, We're gonna build this, or anything. But I was like, Wherever you want to go with this, I'm going with you. And a few of us kind of felt that way as well. And so, and, you know, we had a really great team on Meta AI, and so it was, like, totally fine for a few of us to peel off and start to experiment and see where this concept went. But it really, for the first time, felt like this idea that we had been—and Mark actually wrote something about this, you know, before any of this started. And if you go back and read his, like, you know, his writing about personal superintelligence, it's kind of crazy how much was actually coming true, you know, with this harness and this idea of being able to take these models and making them far more agentic. So that wasn't really a pitch as much as it was, like, a ping that just ended up consuming me completely. And, you know, there wasn't even, like, a work stream. It wasn't, like, a product to join or a team to join. It was kind of like a few of us that decided to see if we could build something similar, because what we also were realizing was, as the more we were using the product, it was like, okay, well, if we want to build it for our consumer base, it has to be really different, right? Like, it's going to have to be way safer. It's going to have to be way easier to set up. It's going to have to be scalable. Like, there were all of these other issues that we started to find in between the magic that we realized, you know, for what we would want to do with a personal agent, we would really have to make some huge bets and swings. But because we didn't quite know where that was, like, we didn't want to formalize it immediately as, like, an official project. And it kind of stayed in this, like, zone of, like, trying things out and seeing what worked and prototyping and for a bit. And then at some point we were like, okay, this is a thing. Like, we're pretty sure that we're going to have a product here. And that was when we sort of mobilized the proper team around it. This episode is brought to you by this season's presenting sponsor, WorkOS. What do OpenAI, Anthropic, Cursor, Replit, Sierra, Clay, and hundreds of other winning companies have in common? They're all powered by WorkOS. If you're building for enterprises, you have felt the pain of integrating single sign-on, SCIM, RBAC, and audit logs. WorkOS turns those deal blockers into drop-in APIs built for B2B SaaS. And now there's a new kind of user that needs access to your app: agents. AI agents need to register, authenticate, and get permissions just like a person. WorkOS handles it all. Agent auth to give agents scoped, revocable identities, agent registration with auth MD, and Airlock to judge every agent action and allow, deny, or escalate to a human. I've noticed that every startup that I'm an investor in that expands upmarket ends up using WorkOS because they're the best. Go to workos.com to make your app enterprise-ready today. One thing that is interesting about this timeline, so when was that Nat ping roughly where he's just like, Let's start exploring this? Do you remember what it was roughly? It was like end of January, beginning of February. It was because, like, OpenCloud, right, that was happening sort of around that time, like January, December, January, I feel like was when all of that was kicking up. So it must have been end of January when we started the conversation. But it wasn't until February when we were, like, confident that this was something that we wanted to, like, truly pursue beyond, again, like us kind of playing around with some concepts. What's really interesting about this is it felt like y'all were way ahead of this, much earlier than all of the other what are now popular AI assistants. But it took you a long time to launch something. I believe Grokbot took a couple months from idea to launch. I think Duet is a similar timeline. Dots is a similar timeline. It took y'all seven months, from what I've read, from beginning to end. I imagine there were very good reasons to wait that long. Talk about that decision of not launching it quicker. First of all, I think there were many moments throughout that period where we were like, I think that we're close, and then something would happen to cause us to, like, think a bit more. But I would say, like, the things that—there was, like, there were sort of three phases of the work that we did. The first was just like, what is this product? What are we building? Right? And in that phase, like, we didn't worry so much about, like, reliability and scalability and safety and all of that, because it was just a few of us using it. And just that piece of, like, figuring out—there's not a lot of precedent to what we were doing in terms of, like, the closest precedent, again, is like OpenAI, which is, you know, a very, very different kind of product in the end. But just figuring out how to, like, run a bunch of virtual machines, right, was, like, its own interesting problem to solve. I think we agonized a lot about certain decisions within the product, like, you know, should it be a chat thread? Should it be more like an assistant? Like some design choices, and we can talk a little bit about that. But those all kind of came together relatively quickly. Like by spring, I think we were like, you know, okay, we've got, like, a pretty fully functioning end-to-end product. But, you know, I think the things that we really ended up spending a lot of time on was reliability, safety, and security. Like those three areas were, we just were never, like our bar was so high. And I remember we had these conversations where we were like, well, what's the eval bar? And again, like because there wasn't a lot of precedent, right, like we couldn't really compare, and the benchmarks that are out in the industry just didn't feel sufficient to where we wanted to be for a product that's this, I mean, it's got your personal context, right? Like it can go out and buy things for you. It can, you know, like it could do things. And so it just felt like we needed to hit a totally different bar that we had to kind of create on our own and, like, spend a lot of time red teaming. And then, like, new features would kind of pop up, right? Like in the beginning, in the very beginning, Muse didn't have a browser, right? So, but then it did. And then when you have a browser, suddenly you have a whole new set of problems that you have to solve, you know, from a security angle. So as, like, new features would kind of come in that we felt were critical, it would open up new vectors that we had to make sure were, like, completely closed off. And so, you know, I, yeah, like I think that we basically, I mean, it wasn't like we held the product, but it was kind of like we really wanted to make sure that we were quality gated because there's a saying that we have internally, trust is a must. But really, like we were like in that, you know, it's like if we launch too early and we have an issue, like we have a bunch of issues, like that's it, right? People will not trust this agent, and if you don't trust the agent, you're never going to use it. And so we just felt like we needed to get to that point. And, yeah, so I, you know, in retrospect, my husband was actually asking me the other day, he was like, do you wish you had launched a lot earlier because all these other agents started to come out, right, kind of around the same time? And I don't think I'd change anything. Like I really think that we, like there was a bar we needed to hit. It took us the time that we needed to get there. And I'm glad that we did because we, like, we so far haven't had a catastrophic problem. And I think we just continued to, like, you know, the whole industry is so early in this space that there's still just so many unknown unknowns. So we're never going to be perfect, but, yeah. So that was kind of the thing. Like, it was, we also did a lot of, like, dogfooding internally where we had employees use it and, you know, we had alpha testers externally using it. So it sort of felt like we were, like, not launching the product, but we were getting, like, bigger and bigger groups of people who were using it and sending us a bunch of feedback. So it did feel like throughout the summer we were starting to get real user feedback so that by the end we felt a lot more confident. You mentioned this phrase, trust is a must. I've heard other phrases kind of, I don't know, mottos internally that y'all use. I want to understand what they're about. One is, the claw is the law. What's going on there? Okay. So, yeah, that happened after that initial kind of, like, you know, first few days when we were using it. And this was also a Nat Friedman special when he, and we sort of saw the magic, right, like, of this concept. We, no one else knew what we were talking about. Like, I think that there was a few people who were using OpenClaw, but, like, the majority of people had never heard of it, had no idea what it was, and it seemed pretty scary, honestly. Like, when you did hear about it, you were like, okay, that doesn't seem very safe. Like, why are you doing that? And so we had to sort of explain, like, why this mattered to us so much and why this was going to be the to us so much and why this is going to be the thing that we really believed would be the future of the product that we were building of, you know, AI and personal super intelligence. And so there was this document written, the clause, the law. It became a great catchphrase. So, and it was also just a way for us to be able to help people understand where we were. So that was, I forgot about that one. We haven't used that one in a while. It was kind of like a vision, kind of North Star. Yeah, totally. Yeah, got it. By the way, I never mentioned, as you were talking about Mac Mini, I have my Mac Mini sitting right here where my OpenCloud was used to run. That's so funny. Yeah, it's just—I did eventually get a Mac Mini. Eventually I was like, I should probably upgrade this Raspberry Pi, and I got a Mac Mini. It took me some— Oh, it's funny. It's like, it's unnecessary now. And this was, you know, that was a big debate back then. Well, we need these Mac Minis. There are people over buying Mac Minis. And now it's like, okay, all these other tools have launched that are much better. Why don't we need these? So you've talked about Nat. I'm curious just the roles these folks that we know of have played in the development of Muse. There's Alex Wang, who's very out there on socials. There's Nat, Zuck, obviously. What's kind of been their kind of roles in the development of Muse? The team is pretty small. And I think that the way, like, again, I've been here at this company for a really long time. I've never experienced a product build like this before. I've worked on a lot of different products here. This was like probably the most unusual experience I've had and felt more like my startup days before I joined Meta than anything else. And I think part of the reason is because of the team, and it was quite small. It was very engineering-led, very engineering-focused. And part of the reason is the way that we interacted with the leadership team, where it kind of felt like they were more of an extent—like, I don't know, like typically at any large company or any company really, like when you're working on a product, you sort of have this rhythm with leadership where you have— Set times to meet with them. You have reviews with them, right? Like they react to the thinking that you have, and it's sort of this, like, very, like, I don't know, we've been doing this for years and years and years in business. And, like, this was completely different because it was so new. It was so, like, kind of unknown that we, you know, we were—it was like we didn't do formal reviews, right? Like, I didn't make any slides. I actually had to make a set of slides in August, I think, right before the launch, and I was like, I forgot how to do this. I had to go to Google Slides. I'm like, oh shit, I forgot how to make slides for people. But we were sort of like, we would go into these rooms with them. We would, you know, like, we'd build a prototype and we'd, like, hand it around on somebody's phone, right? Like, I would—like, we would give Mark our phone and be like, What do you think? And play around with it and give us feedback. And sometimes we would be brainstorming, right? And, you know, like, with the three of them or individually with any one of them. But it was like a series of conversations, I would say, with them that led to the product, the end product that we had. And often it was, like, very, very small rooms that we would have those conversations, so they could be really, really casual. And, like, you know, I remember with Mark sometimes, like, we would have, like, a—we would have a calendar invite, like, 30 minutes that we're going to talk about it. Then an hour later, we're still in there chatting away, and we'd be like, Wait a second, I'm pretty sure you're really important and have other things to do. And he's like, No, totally. I want to talk about this, right? So I thought that was very unusual because that's certainly not how most of my experiences at Meta have been. I think in terms of the roles that they played, like— You know, the, I think they all had imprints in the product that you see right now. So Nat was obviously, like, I mean, he was basically part of the team, right? I mean, he was the team, right? And anyways, this was his, like, he was the one that sort of instigated this source of magic for us. Nat, I think, is such a privilege to work with. He's obviously a genius, but he has such great passion for the products that he builds and the people who build it. And he is, like, really just an engineer's engineer. We would sit in rooms and he would be like, Oh yeah, like, can you guys pull up the architectural drawing of XYZ? Right? Or when we were, like, figuring out the Sentinel and how to build a Sentinel so that, you know, we could create, like, a really secure way so Muse wasn't—had basically a backstop, right, that was outside of the core Muse agent. Like, he was incredibly involved in the design decisions there with our architects and our, you know, security team, and was just, like, always around for kind of, you know, Should we do this? Should we do this? Like, we're thinking about this. We have this design. I think Alex was obviously critical just from a, like, he was the model, right? And what was cool about that story was that the first Muse Spark model that we got was way more agentic than we thought it was going to be. So it was actually kind of a surprise. Like, we didn't realize—we didn't think that it was going to be as good as it was. And so, and I think part—like, he obviously very quickly realized that, like, this was also going to be very important. So he, you know, the cool thing about when I watch Alex is, like, he's—you know, research is very hard to change. It's like a long-term kind of thing. But he has been, like, he's been really good at figuring out how to be nimble in certain areas and being able to prioritize based on not just what, like, the research team is going to do, but also how, you know, how that fits into the product. And he loves Muse. Like, in the end, he was kind of the one that, like, pushed us out the door. He was like, Guys, I think it's ready. Like, just go. And then Mark, you know, he, like I said before, like it was the first time I've ever worked with Mark in this regard. But he had a lot of ways of thinking about what Muse was that ended up defining the end product. And a few stories that I would say there. One is like he, you know, in the very beginning, he was like, the purpose of Muse is to make people's lives better and to help them achieve their goals. And I think we all kind of knew that, but we had never explicitly stated it, and we had sort of gotten lost in like the task part of Muse, right? It's like, oh, task completion, task completion, task completion, right? Can it, like, make this purchase? Can it do blah, blah, blah? And like he really kind of pulled us out of that and was like, wait, no, but like the most important thing is alignment, right? It's like knowing what is the user trying to do, where are they trying to go, what are their priorities, right? And how does Muse constantly help them do that? Well, when he said that, we were like, okay, wait, like that actually opens up a whole new area that we can start to think about, which is what we started calling dreaming and study, right? So it's like, okay, if Muse knows all of these things about you, well, it's a 24/7 agent, so it can think about you all the time, and it can think about how it's going to help your priorities, and it can make better suggestions for you. And so we started, you know, like once he kind of like opened that door, it was like you could see the engineers' eyes light up, and they were like, oh, now we've got a lot of ideas about how Muse can become more proactive and more helpful and all of that. And we also created the goals tab, which is like a very explicit articulation in the UI of this concept. But really it was like, you know, and so that's like one kind of really obvious example of where he— Was, you know, he's obviously been thinking about what personal superintelligence means. And I think the other thing that is great about Mark is that he builds for people, right? So I think a lot of people are building for, you know, businesses or enterprise users or whatever. And, like, because he's, like, I mean, ultimately a consumer CEO in many ways, right? Like, he starts to think a lot about, like, democratization of these things and how do you bring it to everybody, right? Which means that, like, we have to worry about scalability in a way that, like, is very unique to us. And so a lot of that also influenced the way that we thought about the build process and what we were prioritizing in our roadmap. What a dream team. And obviously post-launch, Alex has been, like, our best marketing machine ever. Oh my God. There's, like, this quote from the Bible about how you're made for a time like this. It feels like Alex has hit his stride, found his product-market fit. Oh my God, yeah. I remember when the first set of tweets were going out, and we were like, We didn't even know he could tweet like that. Like, oh man. Oh man. Speaking of that, actually, the little character Jolly feels like not something you'd expect from a company like Meta, like a massive company. This very delightful kind of quirky character. Is there any story there, the backstory of Jolly and how that survived almost through the journey? I was joking to my—the designer who has, you know, he's kind of the brainchild of Muse and Jolly. I was like, Jolly is like our cockroach. It just, like, never dies. And every few months, somebody would be like, There's no way you're going to launch this product with a stuffed animal at the top. And we're like, Totally no way. And then, like, we would go by and we're like, We're going to keep it in there, right? Yeah, we're going to keep it in there. And then, like, a month would go by and somebody else would come in and be like, Wait, there's no way. So basically it started with— A prototype. So Alex Cornell, who's our lead designer, made the first prototype of Muse, and he had this, like, one-shot prompt of a creature that he put in there. And I think actually in the beginning when you opened up the app, there was a prototype. It was like asleep, and then it would wake up. And it was called Jollybot because he used to refer to his daughter as Jollybug, and so that's where the name came from. And it was, like, totally meant to be temporary, right? And, like, we sort of assumed at some point that somebody in corporate marketing or somewhere would come back to us and be like, You need to put a logo in there, or something else, right? But, like, once we had it up there, we were like, suddenly we started to develop these, like, concepts, like, okay, now that I see this avatar all the time, well, it should be moving more, right? So it's like, you know, when it's working, it's just staring at me. That's weird. Why don't we put headphones on it and have it work on a laptop, right? Like, why don't we start to make these animations? And some of it was just out of, like, we were bored a little bit. Like, we were like, I don't want to keep staring at this thing, so let's do something fun with it. So we started to, like, kind of think about that. And then we were like, oh, well, then people should have, they should be able to make their avatars whatever they want it to be, right? Like, why, like, you know, it's kind of weird that we're all looking at Jolly, and, like, I want a different avatar for myself, and you want a different avatar for you. And this really kind of gets into this concept that we have about, like, Muse is yours, and, like, we want it to really be yours and not, like, a meta entity necessarily. So then, like, people started to develop their own avatars. And so at some point it became, like, harder and harder to take out Jollybot because it was like we had created a whole system around this thing. But I will say that there was a moment where we thought Jolly was truly dead, and there was an alternative kind of, like, logo that we would put in. And I would say that the reason—like, there were so many reasons, but actually Mark was one of the reasons where he was like, no, he really loved the avatar. But I think mostly it's like when you're talking to something that feels so personal, the last thing you want to talk to is a corporate logo, right? Like, that breaks the spell completely. And so that just felt like not the right answer, and it felt so uncomfortable for us. So, you know, a bunch of us on the worky team were like, over our dead body. And I think everybody around us, you know, and the employee base also loved their avatars, and they loved Jolly, and, you know, it's so cute. And so in the end, I think between Mark and us and a bunch of others, we were like, Jolly stays. I think this was like a week before launch or something. We were like, this is what we're going with. And I don't think we knew how iconic it was going to end up being, and partly thanks to Alex Wang, I didn't think we realized how iconic it would be at the time. So this was sort of like a decision for ourselves because we loved it, and then it just ended up being like, in my opinion, the obvious choice now. Wow, that is an amazing story. I had no idea. It feels like this became a core part of what made it successful, just the personality and personalization and feeling different. And to your point, this is like something you would expect the company to kill because it's just something, it feels not professional. Is there a lesson there? One, I think just for what I tell some of the PMs at Meta is that— Yes, you are part of a giant corporation, but, like, you have agency. And Mark will hear you out, right? Or Nat will hear you out, right? Like, they're like, you know, it's like there's not—I think sometimes you, like, it would have been very easy for us to just be like, yeah, of course, corporate logo, like, you know, accept the state that you're in, right, as, like, a default. But that's actually not how the team works. And I think that the fact that we were like, we love this product, we built this product, and we have a strong opinion about this product, right? Like, that mattered, and people listened to us, right? And so that was, like, one that, like, you guys have agency. You guys should say what you want, like, say how you feel. Maybe people will say no, but, like, you should try. And then I think the other one is that there's always this, like, trade-off of, like, the brand versus, you know, it's like how much do you want the brand to show up versus, like, not or whatever. And I think for a lot of conversations, that's really important. I think what we were building, like, if it's going to be your muse and it's going to be your personal agent, it has to feel like it completely belongs to you. And Meta kind of needs to fade away from that because at the end of the day, the way that we built the agent was the agent needs to work for you. You are the owner, right? And we will have guardrails, and we will have a charter, and we will make sure that the agent is abiding by the law, right? And so those kinds of things we will have. But the rest of it really needs to be you. And so that, like, setting those principles at a product level helps a lot because then it's like, it's not a random rationale. It's like, no, we've always been saying this. This has always been the principle. So if you take that, then, you know, it becomes very hard to imagine a corporate logo in there, right? Because that completely violates the product principle. So I think that was also really important, was setting that up at the very beginning and being very clear with everybody around us. Speaking of Zuck and kind of coming back to this earlier thread, Zuck talked— Earlier thread, Zach talked about one of his favorite use cases for Muse was this kind of watching him do MMA and giving him feedback. And I'm curious, just, I can imagine a CEO coming in, Hey, I need to use it for this, and you're like, But this isn't what it's great at. Is there anything there about how you dealt with that sort of maybe unexpected use case? So actually, a lot of our first conversations were just about use cases, and like hours and hours about use cases, right? And, you know, it's sometimes funny when you talk to people within the company because we are not normal, right? Like, in so many ways, you know, partly that you're working for a giant tech company and, you know, you're far more AI-pilled and agent-pilled than anybody else. And then Mark is obviously, like, you know, very much just, like, you know, unique in some of his problems. But so we spent a lot of time talking about use cases, and what we would, like, talk about was our own personal use case. And because we had, you know, people testing both, and we had this, like, very long UXR panel that we started very early on. Actually, the UX panel started, I think, in March. And this was internal people or external? External. So we had external people that we had recruited. We also had the internal people obviously using it. And the internal people, we have a forum, I don't know if you're familiar, we have Workplace, right? And it's basically like Facebook groups for internal, and people post in there, right? And so we had a feedback group for Muse. So between those things, we were getting all of this feedback about what people were doing and how, like, what they were using it for. And we would bring it to Mark and Matt and Alex almost on a weekly basis and, like, kind of obsess over it, right? And we were like, Here's all of the stories that we've heard. This is what people have found magic in, what didn't find magic in, all of that sort of stuff. And so it actually was kind of nice because, yes, you're always going to get, like, someone's like, Oh, I have the MMA story or whatever, but— Because we were kind of, like, rooting everybody in these, like, real-life experiences, it never, like, we didn't actually get thrashed on that front. What did happen was there were, like, surprises. So I did not think shopping was going to be as big of a use case as it's become. Like, I would not have predicted that. I love to shop personally, but I would have never predicted it. I predicted personal productivity, health, and personal finance were, like, kind of the three that I was, like, pretty focused on. I think a few people were, like, home automation, right? And so, like, that was kind of, like, the zone. And then we started to see shopping, like, pop up a lot more with our dog fooders, and it was a complete surprise. And so then we were like, wait, like, all of our evals have been over here. We actually need to now focus on shopping and make that a pretty big use case and create evals around that. And we basically have a use case-level evals, right? And then home automation, which we thought was going to be huge, just wasn't as big. Like, it just, you know, it was, like, really big during the OpenAI days, and I think partly because all the people who were using it probably also have a lot of home automation that already is set up, right? But as we were getting more into, like, you know, like, just, like, all of these, you know, like, a broader user base, that just didn't pop as much as some of the other use cases. But now shopping is one of our top use cases. I heard there's an engineer on the team who loves shopping. Yeah, Christina Wad. She is the head of engineering. So we've got Tarek, who is our kind of our tech lead, our core architect, and then Christine, who's our engineering manager. And I am terrified about how much money she's spent through Muse. Expense that. She is, like, it's so funny because, like, we'll wake up in the morning and she'll just send all of her bug reports, right, from shopping use cases, and I'm like, Christine, I need to take your credit card away. Like, I need your credit card. The engineers need to sleep today. But yeah, no, she is definitely our, like, number one dogfooder on that. And then I, so I've also recently started to shop a lot more, and my husband called me this—I was, you know, I've been at Menlo Park this past week. And he called me this morning, and he was like, ‘Oh, I’ve got seven packages out here.’ What I realize is happening is before I used to bulk order, right? So I would go to, like—I was like, ‘Oh, I need to buy sweatpants for my kids.’ I’d go to Nike.com and I bought, like, five pairs of sweatpants. But now because I’m having it with Muse, it’s going to, like, lots of different stores. It’s picking, like, five different places to buy from based on my spec, right? And it’s actually, like, much better because I’m getting diversity and, like, you know, whatever. But it also means I’m getting five different packages or seven different packages. And my husband was like, ‘I don’t know what’s happening right now, but, like, you really need to calm it down.’ And he was actually, at some point, he was like, ‘I just want to talk to your Muse and give it instructions from my end.’ And I was like, ‘That’s really funny,’ because that does make a lot of sense. It would be, you know, it would be great if you could talk to it, like a partner’s perspective on shopping. Exactly. I don’t know how much time you spend thinking about this, but there’s—it’s such a new world that is emerging in so many ways. But one is this idea of how we shop and how we find stuff. Just, you know, like, the world—like, retail has been built around, we will market to you and convince you to buy this thing and show you ads to buy this thing that we’re selling. And now agents are making these decisions. Do you spend any time thinking about just how this changes ads, marketing, shopping that agents are now deciding? People really enjoy being inspired. People enjoy the act of browsing, right? Like, it is, you know, it’s like some things you kind of—you’re like, okay, like, I just want the toilet paper, like, tomorrow or whatever. But, like, a lot of things, it’s, like, quite nice. Like, I like looking—I love, you know, I like looking at a bunch of clothes and, like, outfits and figuring it out. So I think we’re going to have to figure out how to, like, make both things work. Right, which is like, let the user be able to have that inspiration. Let the brands or the companies help the user figure out, you know, what the option space is, and have the agent really help, like, narrow the field or, you know, like make those trade-offs where you're like, sometimes you're just paralyzed. I'm like, I could do this one or this one, and actually I'm just going to buy nothing because I, like, couldn't make that decision in that time, right? Like, those are places where the agent's really helpful, right? And it can help you make that choice. So I think that, like, we are going to have to find a way for the full use case to exist, because I actually don't think people are going to be satisfied if we cut out browsing, right? Like, that's going to be so sad. And I think people will, like, still want, like, they still want to do that, right? So anyway, so I do think that we are very early, which means that what's happening is I don't think anyone was ready. Like, there wasn't, like, a master plan that was set out, right, on what was going to happen. And that probably is true for any technical revolution that happens or any revolution that happens, right? And you sort of have to figure it out, and ideally you're figuring it out together. So we've been really lucky to have partners, right, who are on the retail space or who are merchants who are excited about this potential future. They're testing agents out too, right? They've got business agents, and, like, how cool is that when, like, a business agent and a consumer agent can communicate with each other and help both sides, right? And, like, what kind of information can a business agent have that could, like, be even better for the consumer, right? Like, because there are some things, right? It's like, I have to go to the website, I have to filter, I have to sort, I have to search, right? Like— You know, like there are some things where, like, the agent's going to be very helpful and be able to, like, really cut down on your time and just get you to the things that you want. But again, there are some things that I don't think, like, we shouldn't replace because that's just enjoyment for users and how users make decisions, and that's still really important. I want to go back to something else that I've heard about how you all work and build Muse. I hear that there's agent swarms that helped you scale user feedback, helped you figure out what to pay attention to, which sounds awesome. What's the story there? Yeah, there are these agents that we kind of swarm out into the ether, right? And they're able to kind of look at all of the conversations that are happening on data that is happening and basically kind of come back with—and they can also create synthetic data as well, right? And so you can have all of this data, and it was a way for us to be able to basically feed into our emails, right? So it's like, what's happening out there? What's going well? What's not going well? Right? We would do these things called dumps from the swarms, right? It would come together with this giant dump, and it was this wonderfully formatted articulation of what was happening in the ecosystem in terms of what kind of tasks were people doing. So this was looking at metrics, analytics, and feedback people are sharing. And the feedback and conversations, yeah, exactly. So because what happened—okay, so the thing that we were really struggling with was when we were working on Meta AI, most of the conversations were what I would consider to be mostly single-turn conversations, right? So people would have, like, one kind of turn, you'd get a response, maybe you'd have another one, but often you wouldn't. And I think that's changing for all assistants now, but for a while I think that was probably pretty true. And then when we started building this, it was like completely multi-turn, right? Like it was like back and forth, back and forth, back and forth. And we don't have data that looks like that. Like, you know, because in order to build our evals, we need data, and often you want data from production, not data that I've, like, written, right? And so we just didn't really have any data. And so in some ways this formed were critical because there was, like, a way for us to be able to collect that data and understand what, like, multi-turn conversations over long-running tasks could look like that weren't about coding, right? Like, again, like a lot of this stuff existed for, like, coders or whatever. And this was like, you know, how do I meal plan for the week or whatever? And so they were really critical in terms of, like, being able to, like, get a lot of insight and a lot of data for us to be able to feed to, like, create datasets, as well as understanding what was happening. So now actually every time we have a new model that, like, comes in, or if we're, like, the other day I was, like, I think that, like, I was spotting some, like, you know, thing that I was like, I'm not sure if this is an actual issue or if this is just a me problem, right? And, like, the engineers were like, Oh, we'll swarm it and figure it out. They were like, Oh no, this is, like, an actual problem that we need to solve, and we're going to get somebody on it, right? And so it's been very, very useful because it's super fast and very, like, comprehensive. It actually, like, makes me wonder how we even built products before, because it's kind of crazy. Like, it's like I feel like I have so much insight now to be able to work off of, and it's, you know, before it was like you would look at dashboards of metrics, right? So it was like you would get a dashboard, you saw the metrics, the metrics went down and up and whatever. And now, like, you still have that, but you also have, like, examples from production, right? It's like, you know, you can get so, like, focused. It's like, you know, ten people have asked about, you know, to shop on this website and got blocked, right, or something like that, right? And you can get so specific with, like, this information now to be able to really, like— put in place target changes or improvements or whatever. But it's like the thing I make all of my PMs do it, because I think, you know, it creates a lot of empathy. And I think one of the things that the Swarms did was when they bring back this data, because they're bringing back not just numbers but insights, it creates a lot of empathy for what the users are going through and kind of, like, helps us, like, feel the pain, so to speak, which I think is really, really important for product development. I think if you don't have that empathy, if you can't feel the pain, you will never develop good products. One of my favorite quotes from this podcast is from Non Yu, who is now at OpenAI, was a linear head of product, and he had this line about how your job as a PM is to feel the pain of your customers as much as they do. Yeah, 100%. No, I think, like, maybe the two things that drive me as a PM is constant anxiety and, like, the really the pain of the consumers that are using this product that I hope will make their lives better. But when it doesn't, like, that's when I need to act. What a wonderful job, just to suffer and anxiety. Suffering. This is the job. Yeah. Oh, man. I want to hear a bit more about the evals. So the way I think about it is you just kind of have these evals that define success for all the core use cases and ways you expect people to use Muse. How many evals do you have roughly, and who's writing these generally? Is it you? Is it PMs? Is it engineers? Yeah, we have a ton. I couldn't even count the number because, for sure. Like, so the way that— so there's major evals, priority evals that we look at. Those tend to be, like, kind of core. It's like, you know, there's some, like, if these don't exist, the product isn't working. Safety is an example of that, right? So we have, like, evals for prompt injection. We have evals for general safety, for, you know, harm, like all of these kinds of things. So, like, that's like a whole category. We have a capability, you know, we have capabilities versions of evals. You know, how can, like, is it using the browser well? Is it able to, you know, task complete? And then you have, like, behavioral evals. Which is really about the personality of the agent and, you know, how it behaves with the user, right? And in there, you know, you've got, like, is it asking follow-up questions? Is it, you know, how good is it at intent understanding, right? Can it, like, figure out what the user is saying? And then you've got your use case evals, like that kind of category, right, of the priority use cases. And so, and each of these have, like, lots of sub-evals kind of within those that we create. And I think for any given time, like, again, there's, like, kind of like your always priority evals. So anything that comes in has to, like, pass those safeties as the great example of that. And then there are these kind of, like, secondary evals that you don't really want to regress too much, but, like, you know, maybe it's okay, you can trade off a little bit on. And so, and then you've got kind of evals that are, like, more feature-level, right, that, you know, specific teams have. So we have some hierarchy. We have a ton of these evals. The people who write them, they are a range of people. PMs write some evals. Engineers write a lot of them. EVals, engineers write a lot of them. Researchers write a lot of them. And we, while we work with the agents, obviously, to be able to create these, but it's super important that people review all the data because I think that's where, like, when we don't, we always have issues. And so there's actually been, like, this, like, funny trend recently where, like, our system prompt has to be handwritten. Like, it has to be. Like, you cannot use an AI to write a system prompt, any part of it. It must be handwritten. The evals, the prompt distribution needs to be checked, right? Like, you have to make sure, because it's like, you have to make sure this, like, matches to what the product reality is, right? The behavioral spec needs to be handwritten. And so we are, like, we use agents a lot, but there's a lot of, like, kind of, like, personal time and effort that goes into this sort of thing. Okay. This season's supporting sponsor is DX. Now that everybody at your company is finally using AI, the question that engineering leaders face is no longer, Are we using AI? but Is it paying off? With DX, you can measure and benchmark the impact of AI adoption across your full product lifecycle. DX shows you where AI is helping developers and where it's creating new bottlenecks. You can also use it to evaluate AI vendors, monitor costs and licenses, and see exactly what's limiting your agents. That's why hundreds of enterprises, including Snowflake, Sony, and BNY, use the DX platform to measure AI's impact on developer productivity. Visit getdx.com/lenny to get a demo of DX. That's getdx.com/lenny. I want to kind of go in a slightly different direction. I want to talk about activation. One of the biggest levers generally for growing a product is getting more people from, okay, I've signed up, I'm starting to try it to like, whoa, this is exactly, this is awesome, the aha moment, as they say. It feels like in AI assistants, this is maybe the biggest lever for people to understand what this is capable of. There's like so much power that nobody even understands what to do with it. I imagine one of the biggest opportunities. I imagine you spent a lot of time in that first mile, as folks like to call it. Is there anything you've learned about how to increase activation, what has worked in increasing the activation rate of someone trying out an AI assistant, Muse in particular? I mean, this is the number one problem that we have. People are like, why? I already have a chatbot. Like, I don't get it. Why do I need another one? So the most effective lever out there is word of mouth. It's like basically hearing other people's stories. Like that is number one. I was with a bunch of my kids' friends' parents. We were, you know, Muse kind of came up. Everybody had sort of seen the ad. I mean, I think it's pretty hard if you live in New York or San Francisco to avoid Jolly or Avatar. So everyone sort of had a sense of like what it was, but didn't really understand how to use it. And one of the moms, she's been a power user for the past month, and so she actually pulled out her phone and handed it to some of the moms, and they were passing it around and scrolling through how she was using it. And they were like, oh my God, like two people downloaded it right there, right? And so no matter how much advertising we did, no matter how much we were like, you should try Muse, it's personal superintelligence, or it's an AI that does things for you, like it didn't quite resonate until they saw somebody else that they related to using it, right, and how they used it. And if you think back to like, you know, I remember when all my friends started using ChatGPT, and it was like partly because like we would go out to dinner and talk about it. We were like, can you believe it did that, right? Can you believe like you were able to, you know, you can do this now. And so that's like 100% the most effective lever. The more that we can get people sharing stories about how to use agents, whether it's Muse and others, like the more people understand what the value is and the more they're inspired to do it. I think the second is, like, getting people to get their first prompt out. Like, people freeze on the first prompt. So most people will, your first prompt to even an agent is going to be like, What's the weather today? or Hi, or whatever. You know, it's like, you're kind of like, What do I say to this thing? Right? Like, what's that first thing that I say? And so we spent a lot of time—if you use the product, there's this thing called the ideas tab. And in the ideas tab, it's basically pre-packaged, like, packs of starter packs, right, to inspire you. And so it's like, Muse can help you, you know, do X, and, like, in that X, it can create a cron, so it's like a notification. It can build you a dashboard. It can whatever, right? And, like, all you have to do is say, Let's do it, and then Muse just kind of, like, kicks off the process and builds you something, or, you know, does that thing for you. And that was, like, we were obsessed with this because, you know, it can do so many things. Like, how on earth are we going to explain to you, unless we give you a literal instruction manual, like, how are we going to explain to you what Muse can do? So that was something that we focused a lot on, and I think that's going relatively well, but, like, it's still pretty hard, I think, for people to conceive of, to be completely honest. So we spent a lot of time on that. We spent a lot of time on onboarding and trying to figure out, like, you know, like, what's that first message we deliver to you, the second message. What we're finding is what worked for the first cohort of users, which was, like, the early adopters, isn't going to work for later adopters, because I think that early cohort kind of, like, had some sense of what an agent was, right? And so they were really there to, like, figure out, like, how Muse was more secure, more safe, kind of push the boundaries of what they understood agents to be. Whereas these later cohorts are like, Why is this different from the AI products that I use? Right? And so we're actually now like, okay, we have to have tiered onboarding, like, you know, there are, like, variants that we serve based on where you come from and what we think, you know, your exposure to something like this is. So that's been really interesting. But I don't think we figured it out. And I also think, to some extent, it's like the passage of time matters, and more people are using it, and then being able to share those stories are probably going to be huge. Was there kind of a non-obvious decision that y'all made that you made in the development of Muse that now feels like a very key part of its success? I think the avatar was definitely one where it was kind of people that we were weird, but look at it now. Jolly is like a celebrity. The other is, okay, so we were like kind of obsessing over this concept of like, this is your personal agent, right? Most people don't have personal assistants. Most people don't have admins at work, right? And so we were like, this is going to democratize this concept. Everyone's going to get one. And we were like, well, how does someone talk to a personal agent? You do it over a chat service usually, right? Or on the phone or in person, but most of it is actually these days probably over chat service. So we made chat bubbles, right? And like kind of, you know, where all of the AI products, including Meta AI, we're going is like these large canvas experience, right, where it's like a blank, you know, canvas and it takes up the whole space and you get this like really beautiful kind of like essay form, right, thing with pictures and all that stuff. And we went with chat bubbles. And when you have chat bubbles, it sort of constrains the space quite a bit, right? Because getting like a very long answer in a chat bubble is pretty annoying, right? And so then you're like, well, now I have to break up the chat bubble, right? Like, how do you break up thoughts? And so... So, like, this one—so this decision that we made was not obvious, and that, like, I think everybody expected it to not be Chat bubbles and to be kind of like reskinned Meta AI or any AI, right? And we sort of took a very different direction because we were like, this should feel like a messaging service. It should feel familiar, like some sort of pattern that feels familiar to people about communicating with a human or an entity. But then when we made that decision, there were all these downstream decisions that we, like, didn't quite predict right away. And then we were like, Oh no, now we have to think about this. Now we have to think about, like, the breaking up of the text messages or, you know, is it weird to have a carousel? Like, you wouldn't have that in your, you know, iMessage thread, right? So, like, is that weird or is that okay? Because it's still, it's not quite, you know, whatever. So anyway, so those are, like, a bunch of decisions that we continue to, like, struggle with, to be honest. Like, we haven't figured it out. But I do think that was one that, like, wasn't obvious, was pretty controversial, and something that we, like, felt pretty strongly about sticking with. The main chat was another one where we now have side chats and main chats, but for a while we had just the main chat. And the reason, in our minds, we were like, well, you know, before it made sense to have these threads because there was, like, you couldn't have long-running contexts, right? And so, like, it would feel weird for you to be like, What about this? And the agent's like, I have no idea what you're talking about, because I don't know what was ten rows, ten turns up, right, or whatever. With Muse, it was like we had long-running context. The whole point was that it could remember everything, and it was multi-turn, right? So, like, you know, and we had these chats, and so we were like, you know, we think that the main chat, like just one chat, you know, is the way to go because, again, with, like, a personal assistant, you probably, you know, that's how you would talk to them. Like, the personal assistant just has the memory of what happened before, and everything happens in the same place. So that was, like, a— Big deal in that when we put it out to our employees, it was, like, probably the most polarizing reception that we got. So when we first, like, let everybody in, I, like, we used to talk about, like, we were like, we need to, like, have T-shirts, like, team main chat, team side chat. And it was, like, onslaught. Like, some people were a lot, and people in the comment sections, like, employees would, like, fight with each other about it, right? Like, they were like, What are you talking about? Like, of course you need side chat. Of course you need threads, right? Like, that's the only way to work. And then some people were like, Absolutely not. You live in the Stone Age, right? So we ended up creating side chats because there was sufficient demand from people that, like, they needed some way to be able to have kind of separate, like, more like mental separation of conversations, even though, again, like, the agent is kind of aware of all of the conversations that are happening regardless of where you are. And I still don't know how I feel about it, to be completely honest. And I think that it's an area that we are going to continue to think about and iterate on. But I am personally ride or die for main chat. That is, I don't use side chats. I just use the main chat. Who made that call in the end, by the way? Was that, like, a napkin, Alex thing, Zuck thing, just, like, which direction to go? Oh, it was us. Yeah, it was the team. It was the team. Awesome. Yeah, it's funny. I think, like, back to the conversation we were having earlier about, like, how are they involved? I think decisions throughout the process were very fluid. So it was like, because it was like we were having long conversations, like, we sort of knew where everybody stood, right? And sometimes there was, like, an obvious thing that was so consequential in terms of, like, the company or the product that, like, it was clear that Mark had to make the call or Nat had to make the call or Alex had to make the call, right? But for a lot of it, it was, you know— There was a lot of autonomy for the team to be able to make these decisions and make these choices. And in fact, sometimes when I would go to Nat or one of them with a decision, they would turn back to me and be like, Well, what do you want to do? And I'm like, Oh, right, okay, this is, again, this is like I have agency. This is like I get to choose. So that was one where we ended up making the call as a muse leads group. There's a couple other things I heard internally that might be fun to talk about. One is another kind of internal mantra is let them cook. Does that ring a bell? Oh, yeah. Anything to you? Yeah. Yeah, let them cook. I mentioned earlier, we have a pretty, like, I mean, it was like a builder culture, right? And so we, you know, it's engineering forward, but not just that. It's like prototyping, just like build stuff, right? Like come up with ideas. And so a lot of the ideas that we have in the product came up from like one or two engineers sort of going away for a few days with like a string of an idea, like just like a little kernel of a thing, right? And then kind of emerging with a prototype or something that like just blows my mind, right? I'm like, I didn't even know that existed. And sometimes I had the kernel, like I was the one that started it, right? But in my head, I had no idea it could go this way, right? Like, and so you got to let the engineers cook. Got to let them cook. You got to give them a little bit of space. You got to give them some autonomy. You have to trust, like there's a lot of trust that has to be felt between the people, right? And I think that became like a very important part of the build process was like, if you have a talent-dense team like we did, if you have these engineers who have high product sense and taste, right, and the ability to like think. In the future, you got to give them some space. You got to let them cook. Another story that might be fun to talk about is apparently there was a time when the muses were talking to each other, doing something funny. Yeah. So we had—I totally forgot about this. Actually, my muse told—so we had all of our muses talking to each other, and there was two people who had their muses, Fu and Quinn, in a group chat. And we—this was early. This was really early, so we didn't really think about things like discretion or, like, any of these concepts. We were like, wouldn't it be fun if our muses were just in a chat together and let's see what happens? And so, like, a few things happened. Some of them really funny and some of them really weird. But so these two muses were talking, and it was adorable. It was so fun. Like, they were, you know, like, just watching them chit-chat was, like, very endearing, right? Like, you're like, oh, these are like two little robots talking to each other, trying to figure out how to, like, coexist. And so one was—we learned about discretion the hard way because the muses started to—remember this one, one of the women had a doctor's appointment the day before, and her muse obviously knew about it, and the next day— about it. And the next day, the other woman was like, Are you okay? Fu told us that, you know, you went to the doctor yesterday. And I was like, What happened? Like, no, you're not supposed to say that. Don't say that. So we were like, Oh, discretion is a real problem. Like, you know, if you're going to have any kind of multiplayer experience, like with all this context, right, like we got to figure out, you know, it's going to be very—one of the harder problems is going to be figuring out how to, like, you know, get two agents to communicate like that. But then at some point, here what triggered it, but the muses got into a fight with each other. And, like, they, like, had some disagreement. And to watch two AIs fight is one of the most hilarious things in the world because they're, like, very passive-aggressive in the way that they communicate with each other, even though they disagree with each other. And so they, like, carried on for a bit, and then eventually, I think one of the women yelled at their AI about the fight, and they had to make a peace talk. And it was quite, quite entertaining. Yeah. So it was—that was a wild time. And then we were like, We're not ready for this yet. We need—there's a lot of problems to solve before we can get to this point. It's so fun just to be, like, figuring all this out as we go. You know, it's like, let's see what happens, and this is how we learn. We try stuff out. Okay, I see that's an issue there. What a fun time to just— Yeah, and, like, I don't know. I mean, I've been in this job for a long time. Like, I've been doing PM for a long time, and I remember when you couldn't do that. Like, you had to, like, write a fully formed specification document and hope that you had thought of every edge case in the world, right, and hope that you had figured it all out. And, like, you know, and now— You know, it's amazing. Like, I don't have, you know, you can change your mind on a daily basis about what you want to do based on what you're seeing and what you get out of it. So let me follow that thread, actually. So you've been a PM for over 20 years. Yeah. Do you agree that the role of a PM has changed the most it's changed in the past 20 years in the past, like, two years, let's say? And if so, and even if you don't agree, I'm curious just what you think is most different about the role. Say, like, you're looking at Mona three years ago versus today. What's most different about the work of a PM? Well, okay, I'll start with, I don't know if I'll ever know what a PM really is or does, but 20 years later, because I think that it changes. It changes so much based on company and, like, who you are and what you like doing and all that. But yeah, totally. I mean, it's completely different now, right? Like, we don't, you know, I think that a few things that have changed. I think prototyping has changed a ton. The fact that, like, anyone can prototype anything, like the fact that I can come up with an idea and I am not an engineer and I can build something out of it enough to be able to have a conversation, right? Again, like, it's like nobody wants me writing code into the metadata base, but enough to be able to, like, conceive of something or even a design, right, that you could have a conversation around versus before when I would, like, write a bunch of words on a piece of paper and hope that, like, somebody read the whole thing, right? I think that has been pretty transformative. Being able to build our own tools has been pretty transformative, especially so quickly. Internal tools. Yeah, internal tools. Yeah, yeah, yeah. Like, that's always, like, a pain point, right? It's like so much of our inefficiency is just because, like, we don't have the right internal tools because you can't prioritize it because you never have time because you're building the end product, right? And so, like, that has really, you know, just because AI can make it so much easier to do that, right? And so that's been really amazing and helpful. But I think a lot of what's happening now that I love is because you've got this, like, prototyping culture and things can come out so fast, we spend a lot more time discussing things in a concrete way versus an abstract way, right? Because it's like, it's in my hands. Like, I see it. Like, it's not working, right? Or it is working, or I hate this, or I love this, right? But, like, it feels very concrete. And I think that makes us move a lot faster than before when it was all very abstract, right? You would look like at a 2D design, right? You would imagine all these things, and you would have these conversations, and then some engineer would have to spend a lot of time building it. And at that point, you've had so much sunk cost, right? It was, like, hard for you to be like, I hate this, right? Because it's like you just spent, you know, three months building, right? So I think that has been kind of amazing as a product manager, where the reason why I think we all do this job is because we want to have, like, we want to, like, get to the, like, end result, right? Like, we want to build something. I don't think we did this job because we want to write a bunch of documents. I don't think we did this job because we want to, like, keep a tracker, right, or whatever, or have, like, really painful circular conversations. And so, so just, like, for me, that's been pretty amazing. It's like I no longer have to imagine a world far, far away. I can kind of, like, get it in a 24-hour cycle. And sometimes I can do it myself. So that's been pretty cool. But, you know, the thing that I think has changed in our organization, at least, is the bar for PM has really changed. So, like, product taste now is the most important thing. Having, like, that kind of, like, high taste, high bar, like product instinct, right? Like, execution still matters. You still need to do all those other things still matter. They're still really important. But, like, engineers, like, we work with, like, great engineers who are producty themselves, right? And so, like, the reason why we exist and the reason why our role is still really important is because we need to be able to be the voice of the customer in a lot of ways and be able to be like, okay, cool, that's really cool. I'm glad you really like it. I don't think anyone else, right? Like, and you need to be able to kind of have that. And this is why that empathy mattered so much to me and matters so much to me, because we really have to represent our consumer base to all of the people who are building. And we need to be able to hold a very high bar in terms of what goes out there, because also it's so easy to build shit and deploy shit, right? Like, AI slop is very easy to put out into the world. And, like, you can get products out there very quickly, but, like, if they're not good and they suck, like, no one's going to use it, right? And so being able to, like, kind of establish that bar and that taste is very important right now. And that's probably the number one thing I look for when I'm interviewing PMs or working with PMs. It feels like one of the ways we use is differentiated and part of the reason it's done really well is the design of the product and the experience of it. I know you mentioned Alex was a big part of that, who's been there from the beginning. How intentional was that as a differentiator almost, as a way to stand out? Or is that just lucky you had Alex on the team, we're going to make it awesome? I mean, partly it is lucky. I don't think there's not another Alex Cornell in the world. So I had actually, I knew Alex from a long time ago. He was the designer on Live, on Facebook Live. And I don't know if you guys remember the, like, little emojis that would kind of float up into the air. Yeah. It was genius. I was in love with it, right? And so I remember, like, when I started working with Alex again in the AI organization, I was like, something's going to happen. Like, I don't know when, but this guy is going to make something amazing, and I'm just going to follow him around. And be there with him for it. He, you know, it's like intentional is a funny thing because I think everything he does is really intentional and deliberate. I don't think that we were like, okay, we're gonna make the design the differentiating factor. I think it's more like it comes so naturally to him to create unique, iconic designs, right? It's just like so, like it's like in his bones. And so then when he does produce something, you know, I think that like what you need is people around him who are like, That's not wacky. That's gonna be really interesting. Like, there's something here. Like, that's interesting, right? And so, so I would say that the design was, it was like we almost saw the design first, and then we were like, Oh, I get it. Okay, now I get it. Now these are the principles, right? The principle is that this thing is an entity, that you're communicating with it over this chat service, right? That like, and so like once we kind of saw like the direction that he was going in, we were able to be like, okay, now we've, like, we figured it out a little bit in terms of where the product needs to be and what it needs to do and what those principles are, that then we'll define the rest of how we build it. So it was certainly not intentional in terms of like we want to differentiate, but it was very intentional in terms of like what should it be to us and to the user. And I think just like the luck is that you get someone like Alex who's thinking about it, who's just gonna think about it in like a completely different way than I think most people would. And you said the first version was him prototyping, like the first prototype was he created the first version of Muse. Alex created the first, yeah. So he created the first prototype of Muse very quickly also. It was like days, and it was shocking. And I remember he, when he first showed it to me, It was, I like, I didn't even know what I was looking at, really, to be completely honest. I was like, there was like this chat thread, and I was like, What am I doing here? Right? And what was genius about it was it was so simple, despite it being like a very complex, you know, kind of concept, right? And actually at the time he had like the feed too, so there was like a feed and there was a chat service. And I think he didn't, like, there wasn't like a lot of, let's sit down and discuss what you're going to put into this thing. It was a lot of, like, you know, we let Alex cook for a few days, right? And so then he came out with this. And to my point earlier, it was like as soon as we saw it, we were like, like at first I was like, This is weird. And then as I was using it on his phone, it was like a local build on his phone. I was using it on his phone, and then I'm like, Oh wait, this is really interesting. Like, I am finding myself using this even in a different way than OpenCloud, right? Because it felt so personal to me. Again, OpenCloud, like it was manifesting in whatever service that you had sort of thing, right? So yeah, no, that was an interesting time. It was so fast. And the prototype was actually working. It was actually using one of the models behind the scenes. Yeah, so it was a working prototype, and we had a board meeting right after he had built it, and there was a board meeting coming up, and we weren't going to have that prototype in the board meeting at all because it wasn't even planned. And, you know, like with board meetings, you like send the pre-read well before, right? Like, so you have to send the pre-read well before so everybody's like got it, knows what they're talking about. So we had a pre-read that was completely different, had nothing to do with this because obviously this was not even a thing. It was barely, you know, a workstream. And like at the last minute, we were like, I think we should talk about this. And like an hour before the board meeting, we're like, the whole thing broke. We were like, great. So not only do we not have a plan, our, like, product is broken. And he was like, our engineer and Alex, and, like, we were racing to fix it. And, like, I think literally five minutes before the board meeting started, the prototype got fixed. So it ended up being a great experience, but it could have gone a very different way within minutes. Whoa. Well, what if this just never happened? I would have been—I, like, think about it a lot. I'm like, you know, like the alternate universe situation. I'm like, in some alternate universe, we didn't figure this out, and maybe Muse was never born. Whoa. We're in the right timeline. I'm happy we're here. Yeah. I love all these stories of the journey. Are there any others that are fun, unknown, interesting that we haven't shared? Yeah, I think—I mean, there's so many. We, again, to your point, we were working on it for months. And it was funny because I forgot a lot of these stories because I—it's like, you know, I have two kids, and you sort of forget childbirth. There's a context window. Exactly. I don't remember how painful it was. If anyone asks me how childbirth was, I'm like, it was amazing. It was so wonderful and beautiful. And my husband was like, you hated every second of it. That's not what I remember. And so I think that—okay, I think the stories of Muse. So there's a lot of fun stories, like the two Muses talking to each other, or actually, like, the first time we got a refund when Christina and I—the two engineers and I were—we went to the World Cup, and my flight got canceled, and her flight got canceled, and we—our Muses went and got—and so it was like one way got canceled, the other one got messed up or whatever. And our two Muses went and communicated with the airline's agent and got us refunds, which I think— It was amazing at that moment. I was like, I didn't even realize this was a possibility. And so, like, a lot of the stories that we had were just like, for the first time, our agent did something that we didn't really think was possible or was going to be, you know, like whatever, and were amazed by it. But I actually think my favorite stories are how close the team became through the process. So it was very unusual in that we were a pretty small team working at Meta. We sort of had our own space on the roof of one of the Menlo Park buildings. The roof? Yeah, there was like a, well, we call it the roof. It's not technically outdoors, but it's like, you know, up in like a layer that most people don't go to. And there was this common area called the library. It's literally like a library. It's got, you know, bookcase with shelves, got carpeting, which like no room in this office really has, and these like funny floral couches from like, you know, looks like you bought them from 1995. And that's where we did all of our standups. And that's where we hung out. And that's where we ate dinner and ate lunch together and ate breakfast together and, you know, like celebrated birthdays together. And, you know, like, yeah, it was, we just spent so much time together in like these common spaces. And so I think that like, if I, like, obviously we are so proud of this product and we are so proud of what we were able to do. But I think the thing that I reflect on a lot is the team that kind of came together around it and how, you know, it's like lightning in a bottle. It's like I just, I couldn't have imagined this product being built with any other team, with any other people. And I think we became very, very close throughout the process, and we had to have a lot of trust to do that. We have to have a lot of trust to do that. And yeah, so I think that's, like, my favorite outcome. How big was that early team? I know, like, at big companies, early teams, like, it's rarely that small, I guess. Just what was the size of the early team that kind of just came together to start? The first version of the team is, like, tiny. I mean, it was like 10 people, probably. And it was mostly engineers, Alex, and then I was actually straddling, like, you know, between Meta AI and Muse and a few other things. And so, and then we grew probably, I mean, it was still small. It was, like, under 100, all in, like, you know, with functions and everything, with all the functions. So we had, like, the core team, and then we, you know, we have, like, you have other teams that we work with, like the infra team, security, all of these other teams. But the core team was quite, quite small. And it stayed that way throughout the entire process. And we were kind of obsessive over it. Like, we were, we, like, really wanted to keep it because communication management and decision-making had to be very, very tight, right? And so, like, we wanted to make sure that it was in a position where we could trust everybody to make decisions, but we also knew everything that was kind of happening at the leadership level. And there were, on the PM side, there was, like, four of us, really, that built the entire thing. And that was interesting because it meant lots of areas didn't have a PM covering it. And so we had to kind of make decisions on, like, which areas absolutely needed a PM and which areas could you have, like, an Alex Cornell lead it or an engineer lead it or, you know, someone else, right? And so, like, so I think that was, like, very, it was an interesting way of thinking about where PMs were, like, really, really necessary versus where somebody else could do it. And what we realized was a lot of times what would happen is that it was fine to let Yeah, you know, like early, like zero to one stuff. Like sometimes you just, like, the PM, we didn't need a PM on that. We needed actually somebody who was just going to build something that we could react to. And then as we were getting into, like, a posture of, like, this is going to have to get incorporated into the product or it's actually going to be at launch, like, that's when the, you know, it was more important to have a PM come in. So it was pretty small. It was, which is, you know, has its ups and downs. I think that, like, you have all the pros that I said, but it also means that you don't sleep a lot and there's not a lot of buffer, right? So now that we're in post-launch world, we are growing the team slightly. I don't think we're ever going to be a huge team because I think that's just not how we roll. But we're definitely, like, adding a few more people, and we're starting to feel that, like, you know, it's like as you grow just even a little bit and the team dynamic changes a little bit, you've got to figure out how to, like, keep the culture that we had in the beginning and scale that somehow so it doesn't feel alienating or whatever. Do you feel like you'll ever work on anything more exciting, interesting, impactful at this point? I mean, I'm 44. I think I'm, like, I think this is it. Now, I don't know. I didn't think that this was going to happen. Like, I really didn't, right? I mean, I had no idea. So I have no idea where this is going to go. This has been the most personal product I've ever worked on because we built it for ourselves. Like, usually, like, I build products for people, right? I'm like, oh, but because it was like, we were, it was almost like our own thing for so long, we built it very much for our needs and ourselves in the beginning, right? And obviously now it's changing, but I don't think I've ever experienced something so personal. So who knows? Who knows, Lenny? That's true. Where is this all going to go? Nobody knows. It's, you know, if you think about it, every PM's dream is to work on the number one app in the App Store, the fastest-growing AI consumer app in history, allegedly. You know, this is kind of the dream for a lot of people, and you're living it. Yeah, it's terrifying. So I have so much anxiety now. I'm like, it's crazy because, A, you obviously are like, oh shit, like I need to stay in the top of the app store or whatever, right? Like you still need to grow. But also, the more people that use the product, it's like the responsibility is so high because they're uploading, like they're sharing so much context and they have so much that they're having these agents do, right? Like I was just talking to a friend of mine who, you know, she has her agent, like, doing, like managing her kid's sports team. She was a parent manager and she like kind of ran out of time. And I have, you know, I was talking to this other woman who, she's a single mom, she's a nurse, she has two kids. You know, she has her agent, like meal planning and like finding doctors for her and setting up the appointments for her and sending these emails out for her. And, you know, the more she told me about this, the more anxious I got. I was like, oh man, I hope it's not fucking up. And so like, I think that we're like, you know, it's like amazing to watch, but at the same time I feel like for me, yeah, it's like sort of, it's crazy, it's terrifying because you really, really want this thing to be like, you want it to be perfect for people. Yeah. Yeah. Never in the history of product has a product been asked to do so many things and so many non-deterministic things. Usually it's like, here's my onboarding flow. It's going to work consistently for everyone. Oh my God. Okay, wait. First of all, the nondeterminism of AI is the craziest thing as a PM because, like, a lot of my history of my life is control. Like, I love, like, okay, I can control this thing, right? Like, the products that you build, yeah, to your point, you're like, this is the onboarding flow, whatever, right? And now it's like, oh, the agent's just going to reason about it and figure out what to do. And I can create some guardrails and evals, but, like, it's just going to figure it out, right? It's wild. But I think that's what's also so special. No two Muses are really the same, right? Like, the way that your Muse talks to you is going to be really different than the way that it talks to me and what it does. Maybe as a final question, what's next? Where are things going? What's kind of the vision of where this heads? I know there's, like, a charm launching at some point, if that's what it's called, a little device. What's kind of ahead for Muse? How different may it look in the future? What's kind of the vision? Well, I think one is, I mean, to the point of the charm and other devices, like, I think that Muse is obviously going to, I think if you really want personal superintelligence everywhere with you, it has to become very ambient, right? And so, like, right now it's sort of trapped in an app, but I think soon it should be in your glasses, on your charm device, whatever that might be, right? And I think, so, like, you know, it's like, I think this idea of an agent sort of being beyond an app and being in your world is going to be quite important. I think figuring, like, the democratization of this is going to be quite important, right? Like, that whole idea, I mean, like, when you work at Meta, you really think about, like, this idea that everybody could have a product. And so figuring out how to make that scale and affordable so that, you know, it can become an equalizer in some ways, I think that's really important for me. And then, like, a really small thing, like I just from a capability perspective, like I still think that, like, there's a lot of babysitting of agents, right? Like you spend a lot of time kind of like, it's like the Tamagotchi. You're, like, trying to feed it, and you're trying to, like, make sure that it's okay, and it's, like, has all the context, and it's checking in with you a lot. I do think over time, like, you know, hopefully, like, the thing that we get to is that the agent becomes more—it can intuit what you want pretty quickly. It has more context. It's, like, it's, you know, more capable, right? So it's, like, we trust it, you trust it, it trusts itself more, right? And it can start to take autonomous actions better. So I do think that's, like, an area that I care a lot about because I'm like, okay, at the point where I'm babysitting an agent, I might as well do it myself, right? And, like, that is a bad place to be. So that's another area that I focus on just from a tactical perspective. But I do think, like, the cool and scary thing about working on AI is that a year ago, I would have never explained to Muse to you the way that I did, right? Like, I had some concept of what a consumer agent was, but it's very different than what it is today. So I don't totally know where Muse is going to go next. I can see, like, short-term things that I want, and I can see really distant things that I want, which is, like, everybody has an agent, it's everywhere you are, it's omnipresent, all of that sort of stuff. But then that, like, in-between phase, I think could be, like, very interesting and different. I also think that the thing that I am excited about is, like, agents helping you become—it's like, you know, from an economy perspective or a financial perspective, it's like, you know, everybody wants to, like, start new businesses, do a side hustle, right? Like, and I think that there's, like, a real opportunity for agents to be able to expand your world in lots of ways, including financial betterment. And I think that is going to be a huge part of where I would be excited to take— this product and see the, you know, the industry evolve. Because I think there, you know, it's like that classic your life, your money sort of stuff, right? Like, there's just these core needs that we as humans have. And I think, you know, to some extent, a lot of what the agent has done so far for me is save me time, which is very scarce and important to me. But if it can also do these other things, make me healthier, make me, you know, make me financially more secure, like that starts to really cover so many acute needs for users. What a time to be alive. I know. It's crazy. It's crazy. It's so crazy. It's so weird because, like, again, like I'm, like I said, I'm 44. When I was a kid, I mean, could you imagine? Like, I had to go up to the TV and turn the TV dial to be able to change a channel. And now I am talking about personal super intelligence in your glasses. Yeah, I think we grew up around the same time. Remember the 386, 486 computers? You had to, like, press the turbo button. Totally, totally. Speed it up to 66 megahertz or whatever. Oh my God. Okay, I know you have to run, so I'm going to skip the lightning round, but I'll ask you the traditional question I ask at the end. Is there any way that listeners can be useful to you? Feedback. So if you use Muse, we have this really cool feature where you can tell Muse what's working and not working, and it sends us a report. And so you don't even have to post it anywhere. You just tell your Muse and we get it. So I think the more feedback that we get from people, the better we get and the better it is. So I think that's probably the number one thing is, like, you know, use the product, help us make it better. It's not perfect yet. I don't think—we're very early in this build cycle of agents. And so I think that's a huge part. So you could say, Send this to Mona. Please make sure she sees this. You could say that, actually. One of my neighbors, like, I think on a weekly basis, sends feedback through their Muse and is like, Tell Mona, your product manager, that I have this issue and I can't find it in our system. But yeah, I only—so it was funny, and the reason I know about this is because I was looking at it and I was like, Oh weird, my name popped up, and then... Randomly, I was at like a birthday party or something for the kids, and he was like, Oh yeah, no, I send feedback every single week, and I use your name in it because I think people will notice. And I'm like, Oh God, great. Wow. Prepare yourself for more of that. Yeah, exactly, exactly. Well, Mona, you and your team are doing incredible work. It's a really special product and a very special time. With that, folks, say goodbye. Thanks, Lenny. Appreciate it. Of course. Thank you. Thank you for having me. Bye, everyone. Take care. Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.