Podcast: Alexandr Wang
Compiled by: Houlang Evolution Planet
In a recent podcast interview, Alexandr Wang, founder of Scale AI and head of Meta Superintelligence Labs (MSL), detailed the origins of Muse, his experience working with Mark Zuckerberg, the deal between Scale AI and Meta, and his vision for the future of personal superintelligence. The core points can be summarized as follows: Muse was born from a belief in "personal superintelligence"; its success comes from extreme refinement of details and a unified taste; and Meta's unique advantages make it the best place to realize this vision. (Related reading: Meta AI head Alexandr Wang: Why I want to build Muse)
The following is a compilation of Alexandr Wang's views:
1. We build for the model six months from now, not today's
The first thing to discuss is the vision we had at the very beginning—Meta Superintelligence Labs. When we started working on this, Mark (Zuckerberg) released a memo, and Nat Friedman and some other people and I also worked together. The core concept was "personal superintelligence." The whole idea was: how do we build powerful AI that truly improves all of our lives? How do we create something that can genuinely elevate the human experience as a whole?
In that memo, we talked about many ideas, and I think many of them hinted at what Muse would eventually become, but at the time we were just gesturing at it, very, very early, probably around June 2025. At that time AI was still at a completely different stage, and agents had not yet participated in any meaningful way. But that was exactly the direction we wanted to push toward. That is MSL's mission.
Then we started building the model, because the process of developing these frontier models takes many months. You have to build the entire stack, generate the datasets you need, do pretraining and post-training, and build a whole technical stack for the model. This is a very long end-to-end process. At the beginning of this year, the release of Claude 4.5 gave us a chance to see what the future of agents might look like. It really brought that potential, pushing agents to the forefront.
2. Both terrifying and exciting: Nat Friedman's OpenClaw experience
Nat Friedman and I are close, and I think he was the first person within MSL to use OpenClaw and try to use it. He went through something I can only describe as "both terrifying and exciting." I think he really went all in and trusted it completely.
He told some stories. He told OpenClaw he wanted to drink more water, and then OpenClaw wanted to check on him in terms of safety and visuals, make sure he drank water, and then tell him good job. Crazy stories like many others. It also makes you realize that something quite magical is happening here.
Early adopters like Nat Friedman should count as a pretty good indicator. Because he was one of the earliest customers, many open-source software developers were involved early on, and he also helped create GitHub Copilot and many other things. That was like a clear signal—personal agents clearly have some kind of magic. When you truly trust a personal agent and truly bring it into your life, that is the ability to do too many things.
After hearing those bizarre stories, I started using it deeply. I think for several weeks, both of us were a bit like this—whenever we had free time, we would talk to OpenClaw, immersing ourselves in the role that a personal agent might play.
Nat designed a prompt that basically has you conduct a comprehensive psychoanalysis of yourself, going through multiple waves of deep psychological inquiry. The first thing I did was run my own OpenClaw. That was roughly equivalent to three years of psychotherapy, all compressed into that moment. It was incredibly intense, like abyss-like rounds of psychological probing, because it had other information about me. It could access my email, access many of my photos, search for other information, and integrate everything. During that process, I was very vulnerable. This happened in February.
3. It's like the ultimate consumer product: one memo changed everything
I think both of us reached a similar consensus: something very special is clearly happening here. At the time we both wrote memos. I wrote a memo about these two issues, and both were submitted through Meta forums. Because for whatever reason, it was board season. I wrote a memo about what I felt—in many ways, this is like the ultimate consumer product. As far as consumer products go, super agents are in many ways a clear endpoint.
Then Nat wrote two. One impressive clause was that law and trust are necessary, and ultimately define that this is largely the core of what we are building with these products. So we reached this conclusion. This happened in February. There is something very important in this.
Then we formed the team, and we built a prototype, probably in a week, maybe two. I presented it to the board, and the iteration speed was very fast. Many of the decisions that ultimately appeared in the Muse product already existed in the first prototype. The charm of Muse was initially embodied in that first encounter. That was obviously in February, and seven months later, we launched this product.
The adoption process was really, really difficult. The prototype sometimes worked, and it was magical. But most of the time it didn't work, and it often broke. From there to something highly reliable, like a finely crafted consumer product, that was a long and painful process.
Mark at the time seemed to experience some kind of euphoria, and like all of us, felt fear toward this product. He had used some things at home, he used it with his kids, and he also used it on himself. In his MMA training, he watches videos of himself doing MMA fights, and it can give him some suggestions for improvement. Like some employee meetings, this has in some way become a phenomenon. Most Meta teams have these employee meetings, discussing the crazy things employees have experienced.
4. This is not programming, this is personal
Why were we able to do this when public developer recruitment could not? First, I think we should respect Peter Steinberger and everyone who made huge contributions to OpenClaw. I think Peter is absolutely a visionary. The process by which he created OpenClaw in that beautiful mind of his is simply amazing.
I think perhaps what makes Muse so magical are the less glamorous parts—the hard work, polishing away all the details, and ultimately forming the combination. The model and product experience are really excellent, and we iterated many times on the model itself. Our goal is to steer the overall model roadmap toward personal superintelligence.
If we go back to the beginning of the year, that is, February, Claude was far ahead in programming, far ahead of everyone else, and their technology was advancing by leaps and bounds. This was in some sense the beginning of a certain kind of human activity, acquisitions across the industry, and that was a truly terrifying moment. I think overall it was like, "Wow, they are really advanced in programming." They had Mythos, and they were very clear that they were ahead.
I think it was conventional wisdom, and people in the industry believed that coding agents were the only form, the key factor, perhaps the only technology—did that matter? Or was it just the only capability? That mattered, for advanced AI. There was an incredible feeling that some people in the industry like to crowd into that use case and compete with humans on coding agents.
Because of our experience and our beliefs, I think it took shape around the concept of personal agents. The reason we pursue personal activity is, first, because we believe superintelligence is exactly that. Meta has a unique advantage to do this.
I think one of the first things to consider is something I told Mark even before the whole scale thing. The first thing I told him was that I think Meta is actually a very special company. If we have AGI or ASI or whatever you want to call it. Because part of the promise of AI is that for each of us, there are all kinds of things we want to do and have to do. And the promise of agility is that we will spend almost no time doing the things we have to do, because a lot of that work is about to be done by AI, and for us, we can spend all our time on the things we want to do.For me, Meta's products represent what we all want to do. I follow my interests. I pay attention to all the people I like, and my friends and the people I care about like these products. The Meta team encapsulates the kind of area where we want to spend our time. I think in the long run, Meta is absolutely the winner in this game. That was almost one of the first things I believed about superintelligence and superintelligence. Eventually, a lot of these ideas got incorporated into it, this concept of personal superintelligence.
Meta's mission is to connect people with their friends and family and help them discover their interests. These are very human things. Then, you can also look at this from a business perspective. Meta has 3.5 billion users using the app every day. And the natural area where Meta would have succeeded is consumer AI. That didn't work. This isn't programming, this is personal. So personal that, in a broad sense, agents, personal AI, and consumer AI have always been something Meta has always felt it was meant to do something about, something it could do that was very special.
5. Fewer than 200 people, but everyone knows this is the most important product
We focused on setting out a model development roadmap, and one of the top priorities or key focus areas at the time was building a personal agent system. We had a lot of very long spreadsheets, hundreds of different—not thousands, but over a hundred—for example, specific behaviors and the capabilities we knew the model needed to have in order to build the ideal personal agent. It felt very, very detailed.
Then, over time, we built evaluation systems for all of these, and we figured out ways to train the model to be really, really good at every single thing. We would do some reviews, like looking at that spreadsheet, and then we would identify where we were lacking and make sure we kept improving them. It was like a very meticulous process, requiring slow, continuous improvement and polishing all the edges of the model smooth, so that it could power the product.
This is Meta culture. A lot of the thinking in Meta culture is that once you know what you're measuring, there are a lot of brilliant people there, and you can optimize, and for these things, you can send all the details, and you can make these things really great. I think it's a bit like the era of AI consumer products, where a lot of things revolve around how to optimize conversion rates, how to optimize recommendation rates, how to optimize every part of the new user registration flow. How do you make everything so perfect? And then I think in the AI era, a lot of this work, especially the work from Muse, shows up in the following form: polishing every detail of the model so that it fits the product requirements perfectly.
The resources invested in this project, model making and product making, totaled fewer than 200 people. That sounds like a lot, but in a big company it isn't. Actually much smaller. In retrospect, it feels pretty cool, just thinking about it, we knew this was going to be the most important product, the entire executive team, everyone knew it. This is very likely one of Meta's most important product launches this year. But we stayed restrained organizationally. I think it was important to make sure it was like a very small, clearly targeted team doing this, and that it didn't rapidly balloon the way it is now. Because we knew it was a unique work of art, and it needed a distinct point of view and a high degree of focus, ultimately blending all of this into a beautiful experience.
A lot of AI labs end up becoming some kind of Frankenstein's monster, almost like a Cronenberg fusion, bringing together the perspectives and visions and beliefs of many different people. In a sense, this is an afternoon nightmare, because almost every afternoon these things are trying to get into the product, and then all these mixtures together, like a milkshake. People feel it, users and consumers like you can feel it when the product happens to meet expectations. It feels like a lot of different people have worked on a lot of things, and obviously different voices, different people, they just cram everything together. Muse has to be the exact opposite, that is, it has to have only one clear point of view. We strive to provide a coherent and consistent experience.
In Muse's case, Nat Friedman deserves a lot of credit. His taste and sensibility ultimately determined and shaped the final form of the product. We said millions. We hinted that Muse is the fastest-growing consumer app in history, the best consumer AI app in history.
6. The internet rewards risk-taking: marketing with memes
Since Muse launched, my public tweet style has changed noticeably. There are several parts to what's happening here. This is definitely more fun than when you were building Scale AI. It has to be. The past few weeks have been some of the most fun things I've done. My career is steadily developing, and I think a few things are happening.
First, we felt it was important to find a way to do this—inspiration erupting, like a flood. We're back to this flood-like stream of updates. Like most people, they're always like this, like being bombarded, new products and startups, launches and things and models, here's another model, here's another product, and another feature. Like this downpour, this stream. We had a few big meetings, and the team discussed this together. It felt like we needed a breakthrough, and obviously like the goal, we have a lot of distribution channels, like we can make sure everyone knows about this, but that still isn't enough to make them like it. Breakthrough, in a fun, culturally rich, or engaging world, is worthy of our attention.
Many times in my life I've met and become friends with people who like making memes, and then I have these people I fondly—you could call them my schizophrenic friends, but the people who like them, they think about all kinds of memes on the internet all day, and they understand that. I know some people who are obsessed with Muse, and they're also obsessed with you. The X feed, they're convinced you have hundreds of Muse goddesses making memes ready to use. Turns out you just have a bunch of schizophrenic, internet-addicted friends.
For a while, like a handful of people, I felt at the time that it was really hard to concentrate in meetings. Because I might suddenly think of a meme, and then I'd think, "Oh my god, I have to say this." I was with a friend of mine at the time. I was trying to have dinner with a colleague. And then at dinner, I was like, wait, I figured it out, and then I posted it from a different account than my own. At the time I almost felt like an engineer solving a program bug. Again, because it feels like you can sense that something is about to happen. Yes, this is a problem we want to break through.
There's one more thing I think is also very important, and that is how do we make Meta Cool again? Because for whatever reason, despite having these incredible cultural software products, like the history of Instagram and Facebook. The Meta narrative was once uncool, so I wanted to create something that was both cool and fun, and a little funny, and then that kind of insane feeling. What happened was kind of natural like this. Over time, when Muse first appeared, I was basically just focused on promoting all the interesting use cases I saw. At the time I was like retweeting like dozens of interesting cases I saw of people using Muse. And then, somehow, just because I tweeted too much. There were a few tweets where I was like: I didn't think much about it at the time, I just posted directly. And then, some of those tweets performed really well.
I didn't look at my phone for hours, and then I was like, "Oh, wow. That went really well." And then I suddenly remembered something about the internet. Although this wasn't intentional, I remember the internet rewards risk-taking behavior, and also rewards behavior that likes to take risks, surprises, and things people don't expect.
I think a lot of my colleagues were somewhat surprised by something. Nat said, during the week when I posted all the memes, he kept saying this. He said it in a lot of meetings. He said at the time: "You know, I thought we were going to learn things. I thought we were going to learn a lot about Muse users, about the ways people use Muse, but actually, we've only just begun to learn a lot about Alex." Very interesting.
I feel like I used all my internet knowledge, and after years of effort, I was finally able to immerse myself in those moments. And I think that's true. I don't know, we didn't track the information. It's hard to say, but I do think it actually had an impact on growth, which is kind of funny. I always hear people bring it up.I think the question at the time was, how do we turn this into—how do we turn this breakthrough into a product? I think there were a few parts to it, but I think one of them was: how do you give people the same experience? Like, how do you get someone to wow users with the product early on? And this product, I think, naturally lends itself to that because I think it's AI. It does a lot of things. Like, if someone has never used an AI agent, there are a lot of things about it that will surprise you. So I think that's one of the big reasons. I think another part of it was just like: how do you take all these stories of people having that "wow" moment? Use those to help make the product shine, to help make the product take off.
We ran some ads. I think at first it was just because I would repost them, and there were too many. Every time I saw someone post some interesting news comment, I would repost and repost. I don't understand why Anthropic and OpenAI's ads are so bad, because remember when Anthropic did a massive outdoor advertising campaign? Everything was about them, and the same with Chet. It was like they would love that—icons, and I don't even think people would recognize it as a logo, it just said Clex, with two people sitting at a table. I was like, what does this even mean? It's like you have so many of these use cases. Your ad just needs to show the product's strengths. That's it. You have millions of these. Run those ads. It's so basic. The funny thing is, people always make the same mistake.
I think we ultimately came back to this. A lot of it in construction news and the construction industry in general, I feel like, why would anyone care about this? They don't care about you or your company at all. They only care about everyone's own self-interest. It's like I have a very old book in front of me, it's called Scientific Advertising. It was written in the early 20th century. It's so valuable that the advertisers who made the most money, the agency founders, around the 1920s and 1930s. This book was written by a guy named Hopkins, who was the best copywriter. But essentially it's not about advertising, it's about human nature, and human nature doesn't change. His point is that your ads are actually all about you. Your ads are about your product, your ads are like: hey, I have a thing, buy it! People don't care about that, they only care about themselves. If you just do the work, you just start with that. Some people took these ideas, written in the early 20th century, and they eventually grew into massive industries, corporate advertising agency foundations, in the 50s, 60s, and 70s, and still today.
Let's say the number one thing is that most people in the world don't care at all about OpenAI or research or you know other companies or related research, or even us or our research, etc. What they care about is what it can give them. We received a gift—the mascot, a happy little guy. I remember that tweet, when I realized it, I was thinking, "Oh, wow. This is mileage." I was driving to work. On my way, I like to generate, I saw a guy holding a briefcase, and then I tweeted "my expression," I'm going to work on Monday. I remember thinking this tweet was probably the best one I've ever posted. A while later, I realized you could put him in so many different situations.
One thing I like, I gradually realized, we were thinking about this the whole time during this process: how do you create moments of surprise? This is completely different from any other product marketing you've seen. I don't really remember the last time someone like you came close, using lowbrow memes to promote a product. I actually don't know if anyone has done it before, but I kind of realized, wow, this seems like something we can do with this little guy. This is pretty fun. This easily sparks curiosity in ordinary people. But it might raise their awareness, and it might even get them to download the app. But how do you keep them engaged? Once the app is on their phone?
There's a loop, people, we really like this app, even like in February, we loved it then. You give it a little something, and it does it. You think, "Oh, wow. It actually did that." Then you try to give it a little more. It feels like it needs a little more, and then it does it, and you think, "Oh, wow. It can do this. It can do this first. It can do the second thing." Then you push it a little further, one question. It's kind of like trust fall, almost, your interaction with an AI agent, where at first you might not trust it much, but you're also kind of curious. So you take a gentle bite. Then you change it to something a little bigger, a little bigger, a little bigger.
Of course, like when I talk to people, college students or whatever, some of the ways people use AI, it's like a complete trust crisis. They just love recording voices, notes, and brains, dumping them onto AI. They send things off, and then they trust AI, like organizing all the thoughts and giving them a clear plan, letting them know clearly what to do, etc. I think for successful Muse users, it's really like getting them involved in this trust fall process, using it for slightly bigger and bigger and bigger things, and then letting it work, each one of those times is that kind of situation.
I found my favorite, favorite, I quoted some lines while reading the book "Zero to One." The quote says: "The most powerful pattern I've noticed is that successful people find value in unexpected places." They achieve this by thinking about business from first principles, not formulas.
I wonder how much of this is the result of word of mouth. But what really caught my interest was that a friend of mine sent me some screenshots of how he used Muse. Like when you download the app, it doesn't really tell you whether you'll like it and what you can do with it. It's kind of like an open-ended question, like an open-ended text box. But I have a friend who runs a hundred-billion-dollar company, he left his ID on FlexJet, and then he went over, and just said, hmm, I want him to tell me I don't want to be involved in this, and in the process, you have to figure out how to get this ID, like go get the ID, it's right here, and then where I am, and then the numbers go out, all the steps, all the way through, and then like it was arranged in advance. The courier had been there, somehow I got in touch with the security guard at his office. Then he went upstairs to his office as usual. At the time it felt like, oh my god, this is the best marketing ever! Only three or four screenshots.
I know, this is actually one of those things, I don't know how it's possible, kind of like intentional, but definitely modern. Product marketing relies heavily on screenshots, and I think if you're launching a modern consumer product, it has to show up in screenshots. Like screenshots, it has to fly. That is, whether through group chats, word of mouth, or online spread. I think this played a big role in it. This wasn't intentional, and it wasn't artificially orchestrated. But I think the fact that the little guy appears in screenshots feels like a big deal, because I feel like having a happy or like your Muse goddess in a screenshot immediately signals that this is a Muse screenshot. This is simply the best marketing you could think of.
Seven, a personal agent is like a long-term partner
I mainly like to use it to set up a lot of workflows, and then I, I think most of the time it's meetings. But so actually, like I think in my work, I should focus, and be quite engaged, and interact with people. I set up a bunch of workflows, and I really like feeling like I'm being treated as a second brain.
When you have a powerful technology like AI, what does it mean to be human? What is this relationship? Different people have different views on what it actually looks like. I think the fear case is obviously: AI is above us, and we're like them doing what they're doing. But in other worlds, this relationship is different.
I really believe in Muse. I think a personal agent is like a long-term partner. The form factor is just human nature, we have desires, cravings, and dreams. I feel like AI and personal agents like Muse will become the mechanism for achieving this, and the bridge that enables us to accomplish these things and continue to accomplish these things. And then just like humans, they will dream bigger and bigger and bigger. Kids will dream of having their own galaxies and things like that. And the Muse goddess will help them do that.Bezos wrote that consumers always give people a feeling of being "beautiful but unhappy," or rather, they always carry a feeling of being "beautiful but unsatisfied." They have many choices. It's like a model for our relationship with AI. I think there are other models, such as what I think AI will begin to do, more and more of the scientific discovery process. I think AI will increasingly permeate every area of the economy. But I believe in this module, the long-term use of this form factor: everyone has an AI that can support them and meet their needs.
So you see AI as the process of clearing away all the dirt and impurities in life, right? The things we don't want to do. I think the world is full of gatekeepers and obstacles, like Muse, that can bypass this problem. For you, no expenses, more of your time there, so you can actually focus on the things you want. I think Muse can let everyone have the adult life they dreamed of as a child. Everyone had big ideas and dreams as a child, when they liked to imagine their lives, and then, you know, act it out. They imagined they could live like an astronaut, or they could save the Earth, or they could change the world, or, you know, whatever it was. It's like they aimed high, and then for, you know, there are many reasons why people do this. After finishing school, they enter the workforce. By the time they get to their destination, you know, after working for a while, that's it. All their hopes, ambitions, and ability to dream are drained away. Their lives are worn down to nothing.
I think as adults, we don't like it, like having to realize how much all of us are like the walking dead. We lose our autonomy, we lose what I think "ambition" describes very aptly. One of the promises is AI in a broad sense, and I think products like Muse are like that, cultivating this habit from childhood. Like when you're still a child, harboring big dreams, using AI to help make those dreams come true. Then, dream bigger, make the dream come true, dream bigger, make the dream come true. It feels like experiencing a ladder of institutional upgrading, an increase in autonomy throughout life, and a bit like, you originally had a lot of autonomy, and then that autonomy was completely destroyed.
Of course, the most impressive entrepreneurs have all shown big dreams. I think they are like dreamers. They achieved that goal, and then had even bigger dreams. Maybe it takes them ten years to accomplish it. But you know, if they achieve that goal, they will have even bigger dreams. Elon Musk is obviously a good example. I think Mark is a good example. I think for everyone, there is a version that applies to all people. I think the tragedy is that, I think for most people, you enter the corporate world with bigger dreams and lofty ideals. You sort of become a zombie. Maybe at some point you want to chase your dreams, but you find it really hard, because you have a lot of responsibilities. Maybe you have a family, maybe something else. It feels like you're almost trapped in a zombie-like state.
I think the promise, the world I described, everyone has this like an institution growing throughout their life, and they have the ability to fulfill their wishes and dreams. I think that world looks crazy, but also beautiful. I think it would be very interesting, it would be a feast of art, and it would be cool. You would enjoy entering different pockets, and the world would be very diverse. Imagining what that would actually look like is a bit crazy. Billions of people have, you know, suddenly, due to abundant intelligence, like abundant resources, creating miracles. I think this is part of the promise. I think it's like Rick and Morty's interdimensional cable, but, like, somehow manifested as human reality. It feels like it could be really great.
8. The experience of meeting Zuckerberg
What I'm personally more curious about is that you founded Scale AI, you run Scale AI, and then one day Mark Zuckerberg reached out and said, "Hey, I want to talk to you." I have a friend, this person named Alex Schultz, whom I've known for many years. It probably started about a year after Scale was founded. I met him in San Francisco. He is currently Meta's Chief Data Officer, but for a long time served as a Meta executive lieutenant, and he introduced me to Mark. I'd say it was probably around 2021 or 2022. Talk about AI, because I think at that time I was becoming more and more interested in AI. At Scale, we did a lot of work in AI. My memory is that it seemed to take a year to arrange the first meeting. Like from the introduction, from the meeting being held to finally setting a date, I think it probably took a full year.
Mark obviously has a super complicated schedule, and he seems to have countless things to deal with, all different. And he is very good at prioritizing. Like he will spend a lot of time on the things that really matter. In some ways, it feels like spending too much time on the things that really matter. That's the only way he can do it. Like being a good dad, being a good partner, and all these other things are like ruthlessly prioritizing.
I spoke with him several times. Not long after that, I got his advice, as a founder, seeking advice from Mark Zuckerberg. I actually remember at the time he wrote these, like I would ask him a question, and he would message on WhatsApp, and then write very long replies. I was really very confused at the time. I was thinking, where does he find the time to write such long replies? But I now realize he likes to really, like, write quickly and rigorously, and like a lot of WhatsApp messages. He can quickly write a very rigorous, very long reply. This is a very remarkable skill of his.
Meta is an important customer at scale. I'd say, like before that, we probably contacted each other about every six months or so. Maybe that was the rhythm of our interaction. Then, I think he called, I think it was actually, at first it wasn't very obvious, like the endpoint was originally supposed to be, because the feeling of the first conversation at the time was like, you know, this was after Llama 4, and it was a bit like asking what do you think we should do?
The trajectory of Llama 4 did not match what Meta had hoped for. I think deep down, it was definitely a disappointment. And the situation at the time was also that AI was clearly becoming more and more important. It was obvious that AI was crucial to Meta's future.
This is some of what I heard privately about him. He has this crazy tendency. It's not the board, like a Meta board, although there are also some people on the board who have this kind of relationship with him. But he has a group of world-class players around him, entrepreneurs, and I talked with some of them, like you don't understand, he seems to be seeking advice, constantly. He's like chatting with us, telling us about his recent life or this work, would you do this? He does this over and over again. I think this is indeed an impressive quality, because it obviously requires humility, and continuing to do it even after you know, you know, all the incredible achievements he has made.
The whole timeline from the first time, it was about five or six weeks after we announced the agreement before we got in touch with them. It felt like it ended very quickly.
I had a lot of thoughts at the time, like if I chose, I would think about Meta's AI strategy. One of the first things I said was what I said before, which is like, hey, I actually think because Meta is like from a legacy protection perspective, focusing on what people want to do is what really matters. Among the companies that have this incredible automation technology, one occupies a very special position. I had a lot of thoughts, and I told him all of them in one go. And then that was kind of like a conversation. At the same time, like "oh, right, maybe we should..." we should also maybe consider whether there is a way to cooperate.
When he said that, I thought, hmm, you know what they mean when they say that, right? They want to buy you. It's always the same thing. But they say, it's like, "oh, we have to find a way to cooperate." I think at the time I hadn't received that kind of guidance. One thing I clearly felt at the time was: "Yes, this may be completely meaningless." Like my personal thought at the time was, "Yeah, you know, maybe he's just, you know, teasing it or like throwing things out." Some people did say that, but did what he said really make sense?I mean, obviously like a deal ended up getting done, so, you know, what do I know? But at the time, we didn't know what was happening. Obviously, like the iconic Meta acquisitions include Instagram and WhatsApp, and those were very successful acquisitions. Obviously you understand product logic-driven decisions. Scale is like, you know, a deep enterprise government sales, like, completely different business than, you know, Meta sales. So the logic I thought was reasonable at the time was: if Meta wants to go deep on these things. And then I thought, you know, you know, maybe from an industrial logic perspective, that would make sense, but it wasn't. That's because I think Mark was obviously primarily interested in getting the Llama program and his overall AI program, he wanted talent. Yeah, he wanted talent.
The whole time, things were somewhat murky, unclear. You know, a lot of times that's how it is. After the money hit, you know, when we started talking about numbers, I was like, "Oh, wow. Okay. Alright." That's when things became clear. At first, you think, "Oh, this makes sense." Then you keep talking.
Nine. This is really a strange deal
Part of it felt like, "Wow, this is just fascinating." Deal structure. "We ended up finalizing this deal where Meta invested to own 49% of Scale, Scale continues to scale. A few people and I joined Meta." At the time I was somewhat in disbelief that this is really a strange deal. But I think one of them really is kind of like this. Very interesting win-win, so to speak, win. Scale shareholders benefited from it. Because all Scale shareholders got enormous benefits, benefited greatly. Scale is still scaling, and I firmly believe Scale's golden era has yet to come.
For several weeks it felt like "this is crazy." Is this real? Does this make sense? If it does make sense, how do I feel about it? The main emotion was more like this is just crazy. Then as I thought about it more, like, oh, I actually have to make a decision whether or not to do this. That's how it was. There were a lot of conversations there, a massive number of people, and our investors. Felt like there were a lot of interesting conversations on that side. I mean, obviously, it's really hard to like it, to let go or give up your baby.
It's been 9 years since founding, and you're still very young. You're not even 30 yet. So that's equivalent to a third of your life. Paul Graham said for a long time something like: if you view your company as your life's work, you'll operate differently. I did seriously think about it, Scale is my life's work. I was working on it all along. That was a difficult emotional process. Then, I think what ultimately broke me was a combination of the following factors: wow, this is like win-win, win-win, like everything, um, this is a good deal for all parties, and that's it. I saw the potential Meta could bring. I think the Meta perception of AI definitely holds at this point. In many ways, it seemed like a defective product, and maybe that's not the most attractive place to work on AI, but I think that thing kind of triggered my emotions. Entrepreneur I thought at the time was that you could basically rebuild it, this is like a rebuild of the lab. I think this was one of those things that turned out to be quite interesting, which is like because you know, obviously there's a lot of work to do. It really cleared the way for me, enabling me to build a lot of things from scratch, establish the right principles, build the right culture. It created the conditions for us to create amazing things.
Is that how you and Mark discussed it? It was like, we have to start over. Obviously, if we continue down this path, it's like we'll end up throwing everything away. Our existing resources need to be redeveloped. Completely different, comparable level of talent. He didn't seem to want that kind of small adjustment. He was like, "Oh, this just doesn't work. This doesn't work. Then let's tear it down to the foundation and rebuild." Exactly. It's an evolving process, conversation. I think one thing I strongly believe in. I think Elon has like proven this. At different points in his career, you're like a small, incredibly, highly technical team, very flat organizational structure, and if you do that, you can succeed, a lot of things happen very, very fast. So that's something I strongly believe in, I firmly believe in, in AI, that's the right approach, especially if you have to. You know, what we had to do at the time was move fast. I think early on that was one of the things Mark and I talked about often, and I think he was excited about it too, and ultimately it became one of the pillars. Our overall strategy, for example, small team size, lean staff, high technical level, extremely high talent density. If you have these materials, how fast can you move?
From an outside perspective, it seems like just Mark. Like empty shell casings in a magazine, that's how I thought about it, right? "Mind and Napoleon" was published in 1957. I found it because I saw this interview. I think Tyler Cowen interviewed Sam Altman. I think it was 2018 when he asked him, what's the most important book you've read this year? He said, this is Napoleon's thoughts. So it was hard to find. I bought this book. The whole book is 300 pages, all Napoleon's own words, organized by theme. He repeatedly emphasized that indecision is fatal. But he said, before taking any action, you need to think it through carefully. He would study it from all angles to make sure he made the right decision. I think Mark's way of doing things is a bit like imitating someone. He'll think it through carefully, make sure he's on the right path. But once he makes a decision, he goes all in. He doesn't like to fire all the bullets in the magazine. That's how I felt about what was happening at the time. That's when he'll come get you, and then you start rebuilding the entire organization. I think he's truly understood that failure doesn't matter. It matters, like when you win, and how much you win by. I think this overall philosophy enables him to operate. I think in many situations, I can truly be fearless. Because I think he's very willing to take risks. He has that great line, as if in a constantly changing world. All along, the biggest risk is taking no action. I think he's deeply internalized this. Like there definitely exist all kinds of versions of the world where the goal looks more like Google, in some sense, like more slow-moving and more bureaucratic and looking like a very different company, but I think he's able to do this because he's deeply internalized some of these lessons, keeping things alive, evolving, and vibrant.
Ten. Over 200 direct reports, but I'm not the best manager
One thing I really internalized, I don't even remember where this advice came from, its origin, but it's one of those ideas. Like if you want to be a great founder, you have to be able to do every job. He has very strong opinions on how, like everyone. Small companies should happen, and should work. This leads companies to become very much like—product I try and I like my perspective, and it likes to express it, it's reflected in every aspect, company but it also made me a bottleneck, in a very big way. This has pros and cons for the company.
I think indeed, I feel like, you know, if you're starting a company, it's your first time starting, you'll definitely like to take the approach, because the likelihood of this happening is much higher. To succeed than if you don't take it. Scale your first company. Scale is the first company I founded. But at Meta, the situation is completely different, because we had to, get so much done in such a short amount of time. It's like I didn't have the luxury of doing all the work in the lab. I'm certainly not an AI researcher. I'm not really a product designer either. I'm not, you know, I'm not all these critical things, um, for an organization to succeed.
So, really like this style, so I adopted it, around the philosophy—how do we create an environment like this? We will treat hiring outstanding talent as a top priority. Hire people who are as brilliant and admirable as they are, who are top-tier in every field.I think one of the things is that at the founding moment of MSL, it was very, very charismatic and also very appealing. Is there an opportunity to build a lab from scratch? The team is very, very small. And we have a clear vision for the direction of personal superintelligence. I think it resonates, and I think everyone working in AI, like these people, wants to help people, just like ultimately hoping it will be incredibly meaningful for their mom or grandma, or for their grandpa, hoping it can become something meaningful for everyone in the world.
Because we are using Meta's resources to do this work, we have the opportunity to avoid being constrained by compute bottlenecks. We won't be dragged down by things like infrastructure. We won't be blocked by distribution channels. It's like we won't be bound by the bottlenecks that startups often encounter. But this is truly an opportunity to really build something you might like to leave your mark on, have a lot of influence, and truly shape it. I think this is very exciting for many people. So this is why he chose to spend a crazy amount of money, which seems like a huge amount, to get a few top talents, because then he knew. These top talents, in turn, will recruit other top talents. But you need to use this as a starting point.
I think one thing about that may be underestimated. Part of it is like, I mean, both OpenAI and Anthropic's stock prices have risen a lot. So it just happens to be that way. You know, if you look at how many of these people there are, if they keep their eyes open or if humanity stays this way, what would they do? It feels like a lot of money. So I think for many talented people, the end result is just like this, kid or just staying neutral in many ways. Hmm, overall, but obviously I think this is similar to how it is many times. We won't, many of us won't think about the early situation, early employees, or early researchers at these labs, who have a lot of stock, stock value appreciation, you know what their competitors are like.
We started entirely from a belief like this: having talent recruitment like most smart people is a difficult problem. Personal superintelligence. Hmm, these two problems, how do we create an environment where people can perform at their best? Meta data and MSL. What I do is more of so-called "traditional management." Because these are things I often think about: how do I create an environment? Where all these brilliant people, excellence and very, very special everything can fully develop and express their talents to the greatest extent.
My thinking is more like playing the role of a coach in some ways. It feels very different after scaling, like I became Auri, like I'm the one who has to express his views. In everything we do.
I think my job is something like helping set something like the North Star and something like directionality, where we want to go, which again is personal superintelligence, succeeding in the consumer AI space, for example building AI that is meaningful to everyone, and then discovering special talents within the team, and then providing people with the environment and resources so they can express that ability to the maximum extent.
I think one mistake big companies often make is that there are, like, excellent people in most big companies, but they just don't have the power to do what they are capable of, so they often leave for smaller companies or startups, or whatever it is.
You mentioned Jeff Bezos's shareholder letters before. He would repeat this sentence. Also, in all the books I've read about him, and in the shareholder letters, it says that if you have excellent people but they can't build, they will leave. Yes, exactly. AI is like, it's like a very special thing. I feel like you know the word "research," it feels like it doesn't mean much, but like by definition, humans don't know the limits, what these models are capable of and what you can do with them? It's like they are mysterious artifacts we don't understand, from a supernatural world. So it really is scientific research. It's like we are exploring the limits, what you can do with this technology, what can be built? This is a process that requires many attempts and failures. This requires people with extraordinary insight like geniuses. It's like one of those things, you can't view this process as a very intense and stimulating process. It's like people need space and time, resources to let them perform at their best like you.
And I think about problems from multiple dimensions, right? I think Nat Friedman is simply amazing. And as I mentioned before, like a lot of Muse's work. If there is a unique, distinctive quality in the Muse goddess, then that is her. This is now, what matters to him is to be in a position where he can fully express all these ideas. In addition, we also have some very excellent researchers. They have many, many very exciting ideas. How do we create an environment where they can explore these things? And like most, in the end it won't succeed. Some approaches will achieve very good results. If these methods work, then all the investment is worth it.
So there are two kinds, so to speak, this is a problem, and the metaphor is like mining diamonds or building skyscrapers, it depends on the distribution of returns. One of the things you are doing is. So some areas have very strong power distribution. It's like a lot of what you do is useless, but some of what you do may actually be like literally a million times more valuable, better than everything. I think venture capital is like diamond mining, a great industry that can give people a feeling. The key lies in whether you can identify the special thing. Putting ideas into practice, this is easy to do.
Just look at when you go to your 19th birthday party. My 19th birthday party. I was a freshman at MIT. It was in January. I was doing a winter internship, a work internship program. I won an internship at a trading company called HRT. We had about 10 people, 12 people, something like that. Almost all of them were Harvard or MIT students. Most of us had already met each other through various math, science, or computer science competitions. Some of the others included Jeffrey Yan, who later founded Hyperliquid and achieved great success. Scott Wu later founded Cognition, very successful. Another person was Jesse Zang, who founded Decagon, also quite successful. And another was Vicky Yei, I think she is an outstanding researcher working at Humanotic. Everyone was very talented.
I did talk with him recently, the founder of HRT, and HRT itself is now like crushing it. Unbelievable. HRT is a proprietary trading firm. So in the second quarter of 2020 they were trading with their own money. Revenue reached 11 billion dollars, and I think 8 billion dollars in profit was a record high. I think this company (HRT) has fewer than a thousand employees. I mean, this is simply unbelievable. But I recently met with the founders of HRT, and they were talking about, "Oh yes, we started making some private investments." I was like, "Just invest in the intern pipeline. All interns can get seed funding." That would be great.
I grew up doing math and science and computer science competitions and olympiads, and met all kinds of people from across the United States. Doing this kind of thing. Seeing people who used to compete together now being so successful feels unreal.
Basically, it's like whether it has power-law returns. Some things, like paying, are simply like paying a million times. Investment or more like building skyscrapers and that kind of thing? I would describe data annotation or the data business as more like this, just like Amazon Prime delivery is another example of fairly linear returns. You just need to grow. One makes you feel very, it's really great when doing this. If you are going to build a skyscraper, then it's like a very standardized business, everything revolves around extracting profit. Every ounce of efficiency, extracting from this process, you just observe it, and then figure out I can make it about 2%. This method is more efficient, while another method is 1% more efficient. That way efficiency can increase by 0.5%, and you just try to squeeze it. Every optimization, an ounce extracted from it. And then there is diamond mining, where you feel something like, "Hey everyone, like you know, just go for it, try all your crazy ideas." It's like you are carefully nurturing those ideas that seem promising. This is what I feel I am doing at Meta.Overall, I think at the time I was working on server setup on a large scale. This was a big paradigm. Going back to the original vision, it's like going back to one of the original few arguments. MSL, and what we wanted to do, like rebuilding this lab. How do you have a very, very high talent density? How do you make it so that for a lot of these incredibly talented people, this is the best place for them to bring out their best abilities? And how do you build an organization like this, where the technical focus is, and the technical people are all working to the best of their abilities, and, you know, they have the opportunity to truly be doing their life's work.
One of the design principles is to make it very flat. You know, actually, all the researchers we hired, this group called TBD, they report directly to me. This is more of an anti-bureaucracy statement. Like TBD, it is simply the embodiment of anti-bureaucracy. The key is that we want to hire very excellent people. You have all clearly proven that you can do very outstanding work. You have done similar things before in your careers. So, I don't need to micromanage you in order for you to do excellent work. You can do it yourself. We need technical leadership. So, some of you will be responsible for setting a clear technical direction, your teams, we have this kind of pod-like structure, we have all kinds of pods and various technical pod leads. So, there is already a technical leadership architecture established here. But from the so-called people management perspective or bureaucratic perspective, we are just anti-bureaucracy. And then on the other hand, we do often make collective decisions. It's like we want to debate, because it's like a team rich in talent, bringing together many brilliant people. We want to have this kind of conversation, so that everyone can participate in the discussion. What do we think is the right path? Everyone in the lab has made outstanding contributions. We would love to hear all of these opinions. We want to find the best way forward that everyone agrees on.
So this is indeed a very unconventional way of managing a team. If I said I was the best manager, then I would be lying. More than 200 people. But the key is that they don't seem to need managers, they need a good working environment. They are all great. They are all very capable. They just need room to cook.
So, I may not be a coach. I am more like a guardian of a good environment, so that they can do great work.
Conclusion
This conversation reveals the complete story behind Muse: from a memo about "personal superintelligence," to the prototype in February, to seven months of extreme polishing, and finally becoming one of Meta's most important products this year. Alexandr Wang repeatedly emphasized: unified taste, extreme attention to detail, a small and focused team, and an anti-bureaucratic organizational structure are the keys to Muse's success. And Meta's unique advantages—3.5 billion users, a consumer AI DNA, and Zuckerberg's patience and trust—made all of this possible.
As Alexandr said: "Muse can let everyone have the adult life they dreamed of as a child." This may be the most moving promise of personal superintelligence.