---
title: "Alexandr Wang on Building Muse: Under 200 People in Less Than 7 Months, No Coding Grind or Parameter Wars, Meta's Big Bet on Personal AI"
type: "News"
locale: "en"
url: "https://longbridge.com/en/news/301536999.md"
description: "After joining Meta, Alexandr Wang led a team of fewer than 200 people to create the consumer AI product Muse in just seven months. Following setbacks with Llama 4, he determined that Meta's true advantage lay in personal AI rather than coding, advocating for a singular vision to refine the user experience. After Muse's launch, he personally drove marketing with memes and maintained a flat structure with over 200 direct reports to combat bureaucracy, believing that AI can activate individual autonomy"
datetime: "2026-10-12T04:09:06.000Z"
locales:
  - [zh-CN](https://longbridge.com/zh-CN/news/301536999.md)
  - [en](https://longbridge.com/en/news/301536999.md)
  - [zh-HK](https://longbridge.com/zh-HK/news/301536999.md)
generator: "portal-rs"
---

# Alexandr Wang on Building Muse: Under 200 People in Less Than 7 Months, No Coding Grind or Parameter Wars, Meta's Big Bet on Personal AI

Less than a year after Scale AI co-founder Alexandr Wang joined Meta, he led a team of fewer than 200 people to build Muse, the core product representing Meta's bet on consumer-grade AI. Behind this lies a high-stakes gamble on strategic direction, seven months of refinement, and a story about how to make a product truly "break out."

On October 11, Meta's Chief AI Officer Alexandr Wang shared his team's process of building Muse on the well-known US blog "David Senra."

## Llama 4 Setback, Zuckerberg Steps In Personally

In 2025, Meta's Llama 4 failed to meet the company's expectations.

Wang described the context at the time: "Llama 4 was not on the track Meta wanted as a company. This was indeed a disappointment internally, and it happened at a moment when AI was becoming increasingly important."

It was at this juncture that Zuckerberg proactively contacted Scale AI co-founder Wang. The two had previously been introduced through mutual contacts but did not interact frequently, speaking only about once every six months. This time, Zuckerberg sent a message: "Do you have time for a call?"

The initial conversations were unclear. Wang said that during several calls, Zuckerberg kept asking him, "What do you think we should do?" without a clear sense of where things were heading. It was not until they discussed acquisition terms that Wang realized the gravity of the situation. Ultimately, both sides completed the entire transaction within about five to six weeks: **Meta holds a 49% stake in Scale AI, Scale continues to operate independently, and Wang himself, along with a few core members, joined Meta as Chief AI Officer.**

## Why "Personal AI" Instead of Coding

After joining Meta, Wang and Zuckerberg faced a key strategic choice.

In early 2025, **Anthropic's Claude and Opus were far ahead in the programming field, and the industry consensus was that coding agents were the most important direction for AI applications.**

Wang had a different judgment. He said the first thing he told Zuckerberg was: "Meta is actually a very special company because Meta's products represent what people want to do—**Instagram, WhatsApp, Facebook, these are all places where you choose to follow your interests, hobbies, and friends.**"

**In Wang's view, Meta's distribution capability, reaching 3.5 billion daily active users, naturally points toward consumer-end AI rather than scenarios like work or programming.**

> Consumer AI, personal AI—this has always been the area where we believe Meta can create something truly special. Not work, not writing code, but the individual.

This judgment ultimately materialized into Zuckerberg's internal memo released in June 2025, titled "Personal Super Intelligence," which Wang co-drafted with Nat Friedman and others.

## Nat Friedman's "Fear and Ecstasy" Ignites Product Direction

What truly confirmed to the team that the "personal AI agent" direction had disruptive potential was a firsthand experience by Nat Friedman.

Friedman is a key figure at MSL (Meta Super Intelligence Labs) and an early promoter of GitHub Copilot and an early customer of Stripe. Wang said he was the first person within MSL to deeply use OpenClaw (a personal AI agent product).

"His experience, I can only describe as somewhere between fear and ecstasy," Wang said. Friedman told Muse he wanted to drink more water, and the AI subsequently monitored him via security cameras to see if he was drinking, telling him, "Good job."

Friedman then shared a prompt designed for psychological self-analysis. When Wang used it for the first time, he also experienced intense feelings:

> **It was equivalent to compressing three years of psychotherapy into a single moment—it could access my emails, my photos, integrating all information, and I felt very vulnerable during the process.**

In February 2026, Wang and Friedman each wrote memos and sent them to the Meta board. Wang wrote that he believed super agents were "in some ways the endgame for consumer products." Two sentences in Friedman's memo later became core product principles: "the clause is the law" and "trust is a must."

## 7 Months, 200 People, One "Work of Art"

About one to two weeks after the memos were submitted, the team completed the first prototype and demonstrated it to the board. Wang said that many core designs in the final version of Muse already existed in that earliest prototype—including the mascot "Jolly" and an early version of Muse Charm.

But moving from prototype to product was a "long and painful" process.

> Prototypes sometimes work, with moments of magic, but most of the time they don't function properly and constantly crash. We needed to turn it into a highly reliable, meticulously polished consumer product.

This journey lasted seven months and involved fewer than 200 people.

Wang said the team created a spreadsheet containing over a hundred specific behaviors and capabilities, setting "release-blocking" thresholds and "qualified" standards for each item. They checked against this list after every model iteration. "We had many checkpoints; for example, 80 items might be green lights, but 20 were still red lights, and we knew that wasn't good enough," he said. The first checkpoint where all items were green was a specially trained version of Musepark 1.3.

He likened this process to polishing a work of art: "It is a work of art that requires a strong, singular vision and high focus to create a complete and beautiful experience."

In Wang's view, the reason Muse succeeded while other early personal agent products failed lies precisely here.

> **Products from large organizations often become Frankenstein monsters of different people's opinions and visions—every product manager wants to squeeze something in, and the result is a smoothie mixed together. Users can feel it.**

He particularly emphasized that Muse's product vision came primarily from Nat Friedman: "If there is one person whose singular vision ran through the entire Muse, it was him."

## Meme Marketing: Making Meta Cool Again

After Muse's launch, Alexandr Wang began a series of intensive "out-of-control" posts on platform X—creating various memes using the product mascot Jolly (a small character), which sparked massive reposts.

He admitted this was not entirely intentional.

> A meme idea would suddenly pop into my head during meetings, and I felt I had to post it. Sometimes I was eating with colleagues, suddenly thought of one, and posted it.

His logic was that the internet rewards risk, rewards surprise, and rewards things that catch people off guard.

Nat Friedman said in internal meetings during that period: "I thought we would learn a lot about Muse users from this, but in reality, we just learned a lot about what was going on in Alex's head."

Alexandr Wang believes this marketing approach actually drove growth—although he admits it is difficult to track precisely.

He also criticized competitors' advertising strategies:

> Anthropic did massive outdoor advertising, all about themselves. OpenAI's ads just show two people sitting at a table. I don't understand what that means. You have millions of real use cases; your ads just need to show what the product can do for users. It's that simple.

## Over 200 Direct Reports, Opposing Bureaucracy

At Meta, Alexandr Wang has over 200 direct reports—almost all researchers who joined the MSL research team (internally referred to as TBD) report directly to him.

He explained that this is not a management structure, but a statement against bureaucracy.

> These are exceptionally talented people who have proven they can do excellent work. I don't need to micromanage them. They don't need managers; they need a great environment.

He positioned his role at Meta as a "coach," rather than the "founder" role he had at Scale AI:

> At Scale, I had strong opinions on every little thing in the company; the company was the product of my effort and perspective. At Meta, I am thinking more about: How can I create an environment where these talented people can do their best work?

## AI Can Give Everyone "The Adult Life They Dream Of"

Discussing the long-term significance of Muse, Alexandr Wang expressed a broader judgment:

> I truly believe in personal AI agents as a long-term form. Humans naturally have desires and dreams. AI and personal agents will become the mechanisms and bridges to help us realize these dreams.

He believes that most people's ambitions are gradually worn down after entering the workforce:

> We all become zombies in a sense. We lose our autonomy, we lose our ambition.

He wrote in a post:

> The world so far has been shaped by a few fanatics who managed to make their will a reality. But we have never seen a world where the autonomy of every single human being is fully activated.

He believes that the direction Muse represents is a step toward that world.

The full interview is as follows:

> Building Muse, Working with Mark Zuckerberg & The Future of AI | Alexandr Wang, Meta
> 
> October 11, 2026 · David Senra
> 
> Alexandr Wang is Meta's Chief AI Officer and co-founder of Scale AI. In this episode, we discuss the entire seven-month process of building Muse, why Meta is betting on personal AI assistants rather than coding assistants, and how Mark Zuckerberg and Nat Friedman shaped this product. Alexandr also shares how Zuckerberg recruited him after the setback of Meta Llama 4, why he chose to leave Scale after nine years, and how he rebuilt Meta AI labs around small teams, top talent, and minimal bureaucracy. Additionally, we explore how to make a product go viral with memes, the importance of having a singular vision, and why Alexandr believes AI can empower everyone to pursue grander ambitions.
> 
> ## Chapter 1: The Vision Behind Muse
> 
> David: Tell me about the history of Muse.
> 
> Alexandr: First, let's start with the vision when we founded the Meta Super Intelligence Lab (MSL). At the beginning of the lab's establishment, Mark released a memo, and I, Nat Friedman, and several other colleagues were deeply involved in writing this document, titled "Personal Super Intelligence." The core idea was: How do we build powerful AI that truly makes everyone's lives better? How do we create something that elevates the human experience as a whole?
> 
> That memo proposed many ideas, basically foreshadowing Muse's eventual direction, although at the time it was still a very early concept—it was June 2025, and agents had not yet emerged in any substantial sense, and AI that summer was in a completely different stage. But that was the direction we wanted to build, it was MSL's mission, and it was the endpoint we wanted to push everything toward.
> 
> Subsequently, we began building the models, because the end-to-end development process for any frontier model takes several months: you need to build the entire tech stack, create datasets, pre-train the model, then post-train it, and build a complete engineering support system for this—it is a very long end-to-end process.
> 
> ## Chapter 2: Discovering the Potential of Personal AI Agents
> 
> Alexandr (continued): Early this year, this entire "production line" started operating, and the release of Opus 4.5 allowed many people, including us, to see the future of agents, truly bringing the potential of agents to the forefront. It was early this year, and Anthropic (originally referred to as openclaw, uniformly translated to the correct name below) was on the rise.
> 
> My close partner Nat Friedman, I believe, was the first person within MSL to use and try Claude (originally referred to as OpenClaw, referring to Anthropic's Claude). His experience, I would describe as somewhere between "fear" and "ecstasy"—he was truly all-in, completely trusting it.
> 
> He shared some stories, mentioning these experiences in an interview at Stripe Sessions: he told his Claude he wanted to drink more water, so Claude supervised him drinking water through security camera feeds and then praised him for "doing well." He has countless such astonishing stories, making people realize: something quite magical is happening here.
> 
> I remember a sentence Nat said to me, which left the deepest impression: he said, "I feel there is really something here." The judgment of early adopter Nat Friedman is a fairly reliable signal—he was one of Stripe's earliest customers, a pioneer in many open-source developer tools, and participated in creating GitHub Copilot and many other projects.
> 
> So this was a clear signal: in the field of personal agents, there is obviously something quite magical. When you truly trust a personal agent and truly bring it into your life, it can do so much for you.
> 
> After hearing all these amazing stories, I started using it deeply. For several weeks, almost every spare moment, the two of us were conversing with our respective Claudes, immersed in exploring the possibilities of personal agents.
> 
> ## Chapter 3: Experiencing Three Years of Psychotherapy in One AI Session
> 
> David: Was your initial reaction ecstasy or fear?
> 
> Alexandr: At that time, Nat constructed a prompt for comprehensive psychological self-analysis, conducting deep psychological probing in multiple waves. He basically wrote a prompt specifically for this purpose for you to use with Claude. That was also one of the first things I did with my Claude—it was equivalent to compressing about three years of psychotherapy into that single instant.
> 
> David: Wait, I'm curious, why was the first thing you did with Claude this?
> 
> Alexandr: After I downloaded and installed it, Nat was right there, and he said, "You should try this." So I said okay, did as he said—and then that crazy experience happened.
> 
> David: Tell me more about that "crazy experience."
> 
> Alexandr: It was very, very intense. It was indeed like three years of psychotherapy, because it was continuous rounds of deep psychological probing—it had other information about me, could access my emails, my numerous photos, and could look up other information, integrating all of this together. Throughout the process, I felt very vulnerable.
> 
> It was February. The two of us reached the same conclusion: something very special is obviously happening here. We each wrote a memo at the time—and submitted them through relevant channels to the Meta board, as it happened to be board season.
> 
> In my memo, I wrote that I believed on many levels this was the ultimate form of consumer product, with super agents as consumer goods representing a clear endpoint in some sense. Nat wrote two, with very impressive titles—"the clause is the law" and "trust is a must"—these two memos ultimately defined the core philosophy of how we build these products.
> 
> ## Chapter 4: Seven Months to Build Muse
> 
> Alexandr (continued): We reached this conclusion in February: there is something very important here. Subsequently, we formed a team and created a prototype within about one to two weeks, demonstrating it to the board—this was a very fast iteration cycle. Many design decisions in the final Muse product already existed in the first prototype: Jollybot (that cheerful little mascot) was there, and an early version of Muse Charm appeared in the first demonstration. Not all ideas and feature tags were present, but many were already in place.
> 
> From then on—that was February—we launched the product seven months later. These seven months were a truly difficult process: moving from a prototype that worked intermittently, full of magic but mostly malfunctioning and constantly crashing, to a highly reliable, meticulously polished consumer-grade product, a work we were confident delivering to the consumer world. It was a long and quite painful process.
> 
> David: Before entering that phase, I want to go back to one thing: you made a prototype in one to two weeks and showed it to the board—what role did Mark play in this entire process?
> 
> Alexandr: Like all of us, he synchronously experienced that "ecstasy and fear" brought by this product. He mentioned this experience in various interviews: he used it for various household matters, used it with his children, and also applied it to his MMA training—having it watch videos of his training and provide suggestions for improvement.
> 
> During that period, a trend even formed in staff meetings—across a large part of Meta's teams, everyone would share their wonderful stories of using agents during meetings, that was in February.
> 
> Looking back on this entire cycle now, it is indeed interesting, because the Claude phenomenon obviously had a peak and then subsided. It did not last, and I think many people have experienced this feeling: there were those magical moments, but most of the time it didn't work well, various problems arose, and then it slowly faded away.
> 
> So building this product, to a large extent, the core challenge was: how to make it a truly reliable consumer experience, giving you the confidence to put it in front of billions of people eventually, and ensuring that most people would have a great experience.
> 
> David: Why could you do it, while Claude's developers couldn't?
> 
> Alexandr: First, we have great respect for Peter Sainberger (originally Seinberger, should be Sainberger, i.e., OpenClaw/Anthropic founder) and everyone who contributed to Claude. I think Peter is definitely a visionary; the emergence of something like Claude from his beautiful mind is a kind of magic.
> 
> Perhaps the unglamorous reason Muse is so excellent is the continuous polishing and detail sculpting required to refine the overall experience of model plus product to be truly outstanding. We iterated very, very much on the model itself. We aligned the overall model roadmap with the big direction of building personal super intelligence, specifically the direction of personal agents.
> 
> David: Tell me more about this, and how it distinguishes you from other competitors.
> 
> ## Chapter 5: Why Meta Bets on Personal AI
> 
> Alexandr: Going back to that moment is actually quite interesting. If you go back to early this year, February, Opus and Claude were far ahead of everyone in the programming field and experiencing explosive growth. That was the beginning of Anthropic's comprehensive rise, sweeping the entire industry, a very alarming moment: Oh my god, their lead in programming is too large—plus they have Mythos, the lead is very obvious.
> 
> The mainstream view in the industry at the time was: programming agents are the only important product form, or the only important technical capability. The entire industry had a huge inertia driving everyone to follow this use case and compete with Anthropic on programming agents.
> 
> But precisely because of our personal experiences and the beliefs formed around the idea of personal agents, we chose to step back and think. I think the reason we started with personal super intelligence is also because we believe Meta has unique advantages in this direction.
> 
> David: Let's talk about that now.
> 
> Alexandr: The first thing I told Mark—even before the cooperation agreement between Scale and Meta was finalized—was: If we possess AGI or ASI, Meta is actually a very special company.
> 
> One of the promises of AI is: in each of our lives, there are two types of things: "things we want to do" and "things we have to do." The promise of AGI is that we will hardly need to spend time on "things we have to do" anymore, because AI will handle a large amount of such tasks for us, allowing us to spend all our time on "things we want to do." And in my view, Meta's products represent exactly what all of us "want to do": I use Instagram, WhatsApp, or Facebook, these products allow me to follow my interests, hobbies, and pay attention to people I like and friends I care about. Meta's product matrix happens to cover the domain of "things we want to spend time on."
> 
> In the long run, Meta is an incredible winner in all of this. This was one of the earliest beliefs I formed about Meta and super intelligence, and these ideas ultimately merged into the concept of "personal super intelligence."
> 
> You can look at it from several angles: one is the company's genetic heritage—Meta's heritage is connecting people through friendship and family, helping people discover their interests, these are very human things. Second is the business perspective—Meta has 3.5 billion users using its apps every day, the natural domain where Meta is destined to succeed is consumer AI, not work, not programming, but the individual.
> 
> David: Yes, indeed.
> 
> Alexandr: Therefore, personal agents, personal AI, and consumer AI have always been the areas where we believe Meta can make extraordinary contributions.
> 
> Thus, we arrived at early 2026: you had a prototype, gained internal support, everyone knew Mark had approved this, and then you began the seven months of difficult product refinement.
> 
> Yes. At that time, a shadow hung over us: personal agents "came and went"—we always believed this was because the experience was not perfect or refined enough. But we knew this was a problem we had to take seriously during product development.
> 
> We focused one of the primary tasks of our model development roadmap on building personal agent capabilities. We had a long spreadsheet listing over a hundred specific behaviors and capabilities the model must possess to support an ideal personal agent. The process was very detailed. Over time, we established evaluation criteria for all these items, figuring out how to train the model to perform very, very well on each one. We would also regularly review this spreadsheet to identify areas where we were not doing well, ensuring continuous improvement. It was such an extremely detailed process, constantly polishing every corner of the model to enable it to support the entire product.
> 
> David: Is this also part of Meta's culture? This culture of continuous product refinement?
> 
> Alexandr: Yes. A large part of Meta's culture is: once you know what to measure, there are very talented people who optimize these metrics, polish all details, and take these things to the extreme. In the pre-AI era of consumer products, this was reflected in how to optimize conversion rates, how to optimize recommendation rates, and how to optimize every link in the new user onboarding process; in the AI era, at least starting with Muse, this is largely reflected in polishing every detail of the model to perfectly fit product needs.
> 
> ## Chapter 6: Mark Zuckerberg's Obsession with Improving Products
> 
> David: Have you seen Jeremy Stern's in-depth report on Zuck published in Colossus?
> 
> Alexandr: I saw the report. Honestly, it was too long, I didn't finish reading it.
> 
> David: It has 15,000 words. I just invited Jeremy onto the show, he is my favorite author.
> 
> Alexandr: I saw that episode. I think he is currently the best person in the world at writing profiles of entrepreneurs and investors, truly number one globally.
> 
> David: In that report, Zuck talked about his advantages relative to Elon and Sam Altman, saying: My skill set is very low-key and boring, but I am very good at building teams and then continuously, slowly improving products over a long period. Compared to Elon and Sam Altman, I won't lose control of the company, nor do I need to raise any more money, so I can just do this for a very, very long time. When you were talking about the seven months refining Muse, that report kept coming to my mind.
> 
> Alexandr: I think it's completely correct. This requires immense patience—seven months, maybe it doesn't sound too long, but in the AI era, seven months is equivalent to ten years in traditional industries. This is true for all of us. But I really want to give Mark a lot of credit, this requires great patience, restraint, and trust in the entire process.
> 
> You can go back and look at what Mark said in every earnings conference call during this entire period—from February to September—he would mention that we are building personal agents, and we weren't really trying to hide it, we were building and talking about it, but there were heavy doubts within the company during this entire period. The mainstream narrative on Wall Street was: Oh my god, Meta burned so much money, can they actually succeed in AI?
> 
> David: More than that, it was mostly: Zuck can't win AI. This phrase was repeated over and over for a long time. Going back to Jeremy Stern's report—he started talking privately to Zuck's competitors, including a former researcher who used to work under Zuck and now works at a competitor's lab, who said: "I don't want to compete with them, because Mark is like..." I think the quote was "Mark is that damn Terminator"—he just won't stop. This researcher also talked about one of Zuck's superpowers: he is never satisfied with the current state of the product, always thinking about continuous improvement; but he is not the kind of jerk who destroys team chemistry—he is indeed a dictator, but not rude, and does not hurt the team. The report specifically mentioned that team cohesion is extremely important to him.
> 
> Alexandr: I think these are all true. One of the most magical things about Muse is that we spent a considerable amount of time polishing all the details before finally launching the product. I think the product category of personal agents is a razor's edge category: if it can reliably create these magical experiences, it has the potential to become one of the greatest consumer products ever. But you are facing models, facing non-deterministic things, facing agents that might fail or be unreliable—once users have an unreliable experience, they will think it is garbage.
> 
> This is something that requires bringing product quality and value to a certain critical point to truly resonate widely.
> 
> ## Chapter 7: How Meta Determined Muse Was Ready for Launch
> 
> David: I want to talk more about this process. I am curious about how many resources were invested, how large the team was, and how many people were involved. Do you have data, for example: I use Muse, nine times it's a great experience, two times it's inconsistent or bad, and then I churn?
> 
> Alexandr: You can see this situation with almost all other agent products since the beginning of this year—almost everyone churned, the reason being unreliability.
> 
> We conducted extensive user testing for Muse, organizing many rounds of tests with brand new users downloading and using it from scratch, observing what the process felt like. A core premise of Muse—I think this is where it truly lands—is that it is completely different from the consumer AI products most people have experienced: it is not a chatbot, it is an agent that can do things only agents can do. For developers, this might be old news, as they have been using agents for a while; but for most people, just this leap requires explanation and very thoughtful guidance.
> 
> We also indeed conducted tests of different model versions and continuous A/B testing. The product experience is extremely sensitive to model quality—extremely sensitive to the reliability of the model completing tasks, communicating with users, and proactively confirming when guidance is needed.
> 
> David: But how did you know it was ready for launch?
> 
> Alexandr: Part of it is very quantitative: going back to that spreadsheet with over a hundred rows, we set "release thresholds" and "good standards" for each row. We had many checkpoints; 80 rows passed with green lights, but 20 rows were still red—we knew this wasn't good enough. Until we had the first model checkpoint where all rows passed with green lights—that was a specially trained Musepark 1.3 version.
> 
> This is the very quantitative part. There is also the experiential part: we tried this model and could feel that it was much better than any previous iteration.
> 
> David: Going back to the seven months you mentioned, how many people were working on this, combining model work and product work?
> 
> Alexandr: Less than 200 people in total. By large company standards, this is actually very few.
> 
> Looking back, I think this is quite remarkable: the entire executive team, everyone knew this might be one of Meta's most important products this year, but organizationally we had that restraint—keeping a truly small and focused team, not letting it swell into a project involving half the company, because we knew this was a singular work of art. It requires strong conviction and focus to condense it into a complete and beautiful experience.
> 
> David: Tell me more about this. Why do you call it a work of art, and say it requires a singular vision or conviction?
> 
> ## Chapter 8: Why Great Products Require a Singular Vision
> 
> Alexandr: I think there is a common criticism of products from large organizations, and this problem is now starting to spread in many AI labs: products become a Frankenstein monster—almost like a Cronenberg-style hodgepodge—pieced together from a bunch of different people's perspectives, visions, and beliefs. This is a bit like PM hell, where every product manager has something they want to squeeze into the product, and eventually it all mixes together into a smoothie mixed with various things.
> 
> As users and consumers, you can feel this difference—you can feel that a product was made by many different people, each with their own goals, and then awkwardly pieced together. Muse had to be the complete opposite: only one clear conviction, one complete and unified experience we present to the world.
> 
> David: I like what you said there. It's purely coincidental—I don't know when this episode will be released, but the time we recorded this episode happens to be the 15th anniversary of Steve Jobs' death. Speaking of consumer products with a singular vision, you naturally think of Steve first. I have read all books about Steve and Apple's history, and done about 15 to 20 episodes about them on my other podcast, Founders. My favorite phrase to describe Steve's influence is: Apple is Steve Jobs living ten thousand lives. The product embodied his taste, his perspective, his conviction—what he wanted to see in the world is what the end user got. In the case of Muse, Nat Friedman deserves credit—his taste and aesthetic perception ultimately decided and shaped the face of this product.
> 
> David: Do you publicly discuss Muse's user data?
> 
> Alexandr: We said "millions," and I also retweeted charts posted by others implying that Muse is the fastest-growing consumer AI application in history.
> 
> David: But you won't give specific numbers.
> 
> Alexandr: We said "millions."
> 
> David: Yeah, millions can also be hundreds of millions, that's also millions. So you just didn't say the modifier before it.
> 
> ## Chapter 9: Using Memes to Make Meta Cool Again
> 
> David: I want to talk about your status on X—I say this with goodwill—you've been a bit "unstable" recently. Since Muse's launch, your public tweeting style has changed significantly, is this a marketing strategy? What is this about?
> 
> Alexandr: Several things are happening simultaneously.
> 
> David: This is definitely more interesting than what you did at Scale.
> 
> Alexandr: Definitely interesting. The past few weeks have been the happiest time of my career, bar none. Several things are happening:
> 
> First, we felt it was important to find a way for Muse to break through. It is an era where AI products, startups, new versions, and new features are constantly bombarding everyone, for most people it is just an information flood. We held large team meetings specifically to discuss this: we need to break through—Meta has strong distribution capabilities that can ensure everyone knows about this, but this is not enough to make it truly capture the world's attention in an interesting, culturally infectious way.
> 
> Coincidentally, I know some friends in life who are obsessed with internet memes and internet culture—I affectionately call them "my neurotic friends"—they think about memes and the internet all day long.
> 
> David: I know some people obsessed with Muse, they are also obsessed with your X timeline, they are convinced you have a hundred Muses helping you think of memes.
> 
> Alexandr: Actually, every tweet on the X timeline is written by myself. For a period, I really found it hard to concentrate during meetings because meme ideas kept popping into my head, and I couldn't resist posting them. Once I was having dinner with colleagues, halfway through I suddenly said, "Wait, I got one!" and then posted another meme. It felt a bit like being an engineer solving bugs back in the day—you can feel that intuition of "this will blow up."
> 
> Second, I think it's important: how do we make Meta cool again? I don't know why, but despite owning unparalleled cultural software like Instagram and Facebook, Meta hasn't been cool for quite a long time. So I wanted to make Meta cool, fun, and a bit playful.
> 
> As for that "letting go" feeling, it happened naturally over time. When Muse first launched, I was mainly sharing various interesting use cases, retweeting dozens of people's Muse usage scenarios. Later, for a period, I posted a few tweets without thinking much, and those turned out to perform very well.
> 
> David: Have you seen the rule that "the longer you think, the worse the effect"?
> 
> Alexandr: It fits perfectly. Some of my tweets I really didn't think much about, I didn't even look at my phone after posting, and a few hours later I looked, "Wow, this one blew up." Then I remembered something about the internet—although not intentional, I recalled: the internet rewards risk, rewards surprise, rewards things people don't expect. So I let go.
> 
> Many of my colleagues were a bit surprised by this. Nat had a sentence, which he said in several meetings during that entire week of posting memes: "I thought we would learn a lot about Muse users from this, learn how people use Muse, but in reality we just learned a lot about what's in Alex's brain."
> 
> That period was really happy, I felt like I was tapping into my years of accumulated internet knowledge. I feel this actually had a real impact on growth—it's a bit funny, although we don't have precise tracking data, it's hard to be sure.
> 
> David: I definitely think it had an impact. I hear about it everywhere, and the people I hear about have all downloaded the app and are using it now.
> 
> Continuing what you said earlier—you have distribution capabilities, but that doesn't mean you can make people truly come back and make the product truly sticky. So how did you achieve that?
> 
> Alexandr: I think the key is: how do you get users to the "wow, so it can do this" moment as early as possible? This product is naturally suited for this, because if someone has never used an AI agent, many of the things it can do will surprise you.
> 
> Another key is: how to capture stories of people having "wow" moments, and use these stories to spread the product wider.
> 
> David: But how did you use these stories? You weren't running ads, right?
> 
> Alexandr: We did run some ads. But honestly, at the beginning it was just me constantly retweeting interesting cases—as long as I saw someone doing something interesting with Muse, I would quote-retweet it.
> 
> I never understood why Anthropic and OpenAI's ads were so poor. Remember Anthropic did massive outdoor advertising, all talking about themselves; OpenAI was the same, the ad just had an icon (which I think most people wouldn't recognize as their logo), and then wrote "ChatGPT," with a picture of two people sitting at a table. I was wondering what this meant? You have so many use cases, the ads just need to show the benefits of the product, it's that simple.
> 
> David: You have millions of ready-made materials, just use them to run ads. It's interesting that such a basic error is repeatedly made.
> 
> Alexandr: A question we repeatedly returned to when building Muse and doing other things at Meta is: why would others care? They don't care about you, they don't care about your company. Everyone acts out of self-interest.
> 
> There is a very old book called "Scientific Advertising," written in the early 20th century by Hopkins, a top copywriter who was one of the founders of the most profitable advertising agency in the 20s and 30s. He said in the book: this is not just about advertising, but about human nature that does not change. His core point is: your ads talk about yourself, about your product, saying "I have this thing, come buy it"—but people are not interested in these, they are all self-interested. You only need to do one thing: start by showing what this can do for them. Those who absorbed these ideas written in the early 20th century built huge companies and advertising empires in the 50s, 60s, 70s, and even today.
> 
> Alexandr: Yes, completely correct. Several key insights are: First, most people in the world don't care about OpenAI or some research, they don't care about these companies or our research, they care about what this can do for them. We also have a gift, which is that mascot—that little cheerful character.
> 
> I remember the tweet that made me realize "this has potential": on my drive to work, I used Muse to generate an image of a little guy carrying a briefcase, and then tweeted: "MFW (My Face When) going to work on Monday." At that moment I felt this was one of my best tweets recently, and then I realized I could put him into all kinds of situations.
> 
> I also realized how to create surprise moments completely different from any other product marketing you've seen. I don't remember the last time someone used bold memes to market a product, I'm not even sure if anyone has done this. But I felt: wow, this is something we can do with this little guy, it's interesting, it will spark curiosity in ordinary people.
> 
> ## Chapter 10: Building Trust with AI Agents
> 
> David: Okay, but this might get them to know the product, even download the app. But after the app is installed on the phone, how do you keep them using it?
> 
> Alexandr: Users who truly fall in love with the app go through a cycle—we observed this pattern as early as February: you first give it a small task, it does it, and you think, "Wow, it did it." Then you give it a slightly larger task, it does it, and you think, "Wow, it can do that, and this." Then you give it a slightly more complex problem. It's a bit like a trust-fall game with an AI agent—at first you don't trust it much, but you are curious, so you give it a small try, then slightly larger, then larger.
> 
> When I talk to college students, some are already in a state of complete trust-fall with AI: they just record voice notes, dump all their thoughts to AI, and trust AI to organize all ideas, giving them clear plans and action items. I think the key for successful Muse users is to get them into this trust-fall process, first using it for one thing, then a bigger thing, then bigger, and succeeding every time.
> 
> David: I am curious to what extent this is driven by word-of-mouth. Obviously, after talking to you I will download this app anyway. But what really sparked my interest was a screenshot a friend sent me, showing what he was doing with Muse. He left his ID card on a FlexJet plane, and then he said to Muse: I don't want to participate in this process, you figure out how to get this ID card back, my ID is there, I am here, help me figure out all the steps in between. Then Muse arranged a courier, retrieved the ID card, passed through his security system, and delivered it to his office. This is the best marketing: just three or four screenshots, back and forth, he says "I don't want to participate, you handle everything," and then Muse handled it step by step.
> 
> Alexandr: Yes. Modern product marketing is largely driven by screenshots. A modern consumer product must be able to speak in screenshots—screenshots must be shareable, whether through group chats, word-of-mouth, or social media. I think this has always been an important driver. It wasn't deliberately designed, but I feel the presence of the little guy in the screenshot is a big deal—having Jolly or your Muse image appear in the screenshot immediately conveys "this is a Muse screenshot," which is the best marketing you can have.
> 
> ## Chapter 11: The Future of Human Ambition
> 
> David: So do you use Muse all day yourself? What are your personal usage habits?
> 
> Alexandr: I mainly set up many workflows and use it between meetings. I am supposed to listen carefully and actively interact with others at work, so in actual use, I set up a bunch of workflows, using it as a "second brain."
> 
> David: You mentioned this in the article you were writing, which was short, and had 5.5 million views last time I checked. Your description of Muse is very good—"if everyone had a second mind beyond themselves." I think philosophically, this is one of the most interesting questions about "what it means to be human when you have powerful AI"—what that relationship is exactly, different people have different imaginations. The most worrying scenario is of course AI rising above us, and we just passively follow. But before you continue, what is your own view?
> 
> Alexandr: I truly believe in Muse, and in personal agents as a long-term product form. Humans naturally have desires, wishes, and dreams, this is human nature. I think AI and personal agents, like Muse, will become the mechanisms and bridges for us to realize these dreams, to continuously realize dreams. And then humans will dream bigger, bigger, bigger. Children will dream of having their own galaxies, and then Muse helps them realize those dreams.
> 
> Bezos wrote a sentence in a shareholder letter, roughly meaning: the most beautiful quality of consumers is that they are always "beautifully dissatisfied" with existing choices. I think this is a model for our relationship with AI. There are other models—I think AI will increasingly participate in the process of scientific discovery, embedding deeper and deeper into various parts of the economy. But I do believe this form will exist long-term: everyone has an AI supporting them to pursue what they want.
> 
> David: So you see AI as something that can remove "life's burdens," right? Those things we don't want to do. You said in that article that the world is full of gatekeepers and obstacles, and Muse can help you bypass these, no longer consuming more of your time, allowing you to truly focus on what you want to do.
> 
> Alexandr: I think Muse can give everyone the adult life they dreamed of in childhood. Every child has grand ideas and dreams, when they imagine their lives, they imagine becoming astronauts, saving the earth, changing the world—they think big.
> 
> But for various reasons, when people finish school, enter the workforce, and work in a position for a period of time, those hopes, ambitions, and ability to dream are slowly drained away.
> 
> David: Not drained, but beaten out of them. Life beats them out of people. As adults, I think we all underestimate how much we are like walking dead. We lose our agency, we lose— I think "ambition" is a good word. The promise of AI in a broad sense, and the promise that products like Muse can give, is to allow this state to continue from childhood: having big dreams when young, using AI to help realize them, then dreaming bigger, realizing again, dreaming bigger, experiencing a rising sense of agency throughout life—rather than agency being abundant and then crushed.
> 
> Alexandr: What you just said reminds me, you previously mentioned that unsettling AI psychological review you and Nat experienced, with waves of deep psychological probing. What if it were reversed? Not that, but really giving you more confidence—letting you understand that the world is malleable, and with enough energy and drive to pursue, you can really change the world around you.
> 
> I think this is the trait exhibited by the most powerful entrepreneurs: they dream big, realize it, then dream bigger, perhaps cultivating in a direction for ten years, realize it, then dream bigger. Elon is obviously a typical example, and so is Mark. Many iconic entrepreneurs have such personal experiences. I think this should become the norm for everyone.
> 
> The tragic part is, for most people: you have bigger dreams and ideas, enter the corporate workplace, slowly become a walking dead, and at some point you want to pursue dreams again, but it is really too difficult, because you have so many commitments—perhaps you have a family, perhaps various ties—you are almost trapped in that "walking dead" state, hard to escape.
> 
> There is one more thing about Muse, before we talk about Scale AI and your decision to cooperate with Meta, I want to say first. You wrote in that article: "The world so far has been shaped by a few fanatics who managed to make their will a reality in some way, but we have never seen what humanity looks like when the agency of every individual is fully activated."
> 
> I think the world I described—where every individual's agency is continuously strengthened throughout their life, and has the ability to realize their wishes and dreams—will become crazy in a very beautiful way. It will be interesting, full of artistry and coolness, walking into different corners of the world, it will be very diverse, simply unimaginable: billions of people suddenly having the ability to create extraordinary things due to abundant intelligent resources. I think this is part of the promise. In my mind, it is a bit like interdimensional cable TV in "Rick and Morty," manifesting in the human world in some way—that would be truly remarkable.
> 
> David: A line from that animation has been echoing in my mind: the universe eats smart people.
> 
> ## Chapter 12: The Call from Mark Zuckerberg That Changed Everything
> 
> David: Okay, I want to trace back. I am personally curious: you founded Scale AI, were running Scale AI, and then one day Mark Zuckerberg contacted you, saying—I assume he said—"Hey, can we have a call?" Can you tell me how this happened?
> 
> Alexandr: I think the message was roughly: "Hey, do you have time for a call?"
> 
> David: Did you have any prior relationship with Mark? Did you spend time together?
> 
> Alexandr: I have a friend named Alex Schultz, actually I've known him since Scale was about one year old, we met in San Francisco. He is now Meta's Chief Data Officer, and was a long-time executive at Meta. He introduced me to Mark around 2021 or 2022, when Mark was becoming increasingly interested in AI, and Scale was also doing a lot of work in the AI field.
> 
> I remember it took about a full year from the introduction to the first meeting being scheduled. Why so long? Mark's calendar is obviously extremely busy, he has countless things to handle, and he is very good at prioritizing—he spends a lot of time on truly important things, sometimes even too much time. The only way he can do this, while also being a good father and partner, is to ruthlessly prioritize.
> 
> Later we met a few more times, I would ask him for advice, getting advice from Mark Zuckerberg is of course an incredible thing. I remember at that time, I asked him questions on WhatsApp, and he would reply with very long messages, I was confused at the time: how does he have time to write such long replies? Later I realized, he just writes extremely fast, capable of quickly writing rigorous and detailed replies to a large number of WhatsApp messages, this is an impressive ability of his.
> 
> Meta is an important customer of Scale, before this event, we communicated about once every six months, this was the frequency of our interaction. Then came that call, initially it was not clear where it would ultimately lead, because the first conversation was just—after Llama 4—basically asking: what do you think we should do?
> 
> David: Why emphasize that this conversation happened after Llama 4, is this important?
> 
> Alexandr: Llama 4 was not on the track Meta hoped for. That was indeed a setback internally, and it was at a time when AI was obviously becoming increasingly important—AI would obviously be crucial to Meta's future.
> 
> The initial few calls were telephone exchanges, he was somewhat seeking opinions: what do you think we should do? Where do you think we should focus?
> 
> David: This aligns with what I've heard privately about him—he has this amazing network of world-class entrepreneur advisors, not exactly the Meta board, although some people on the board are involved. I've heard he constantly seeks advice from these people, messaging them: "This is the situation I face, what would you do?" Doing this over and over again. I think this is a very remarkable quality, because it requires humility—to continue doing this after achieving those incredible accomplishments.
> 
> Alexandr: From his first contact to our announcement of cooperation, it was only about five or six weeks, quite fast.
> 
> David: This is also a commonality among world-class entrepreneurs. We had Jonathan Ross, founder of Grok, talking publicly about that $20 billion cooperation with Nvidia for the first time on my show, he said: "From Jensen's first call to the money hitting the account, it only took three weeks."
> 
> Alexandr: That's impressive. I had many ideas about Meta's AI strategy at the time. One of the earliest ideas was what I said before: Meta is genetically focused on what people "want to do," and in this great AI transformation, it is actually one of the most uniquely positioned companies. I conveyed a large number of ideas to Mark, and at the same time, was having another parallel conversation: "Perhaps we can consider if there is a possibility of cooperation."
> 
> David: So when he said this, it was completely unexpected for you.
> 
> Alexandr: Completely unexpected. The whole thing was like a series of crazy events. At the time I was thinking: "Oh, this is nice, but impossible..."
> 
> David: When they say "we have to find a way to cooperate," you know what this means, right? It's acquisition. They always say this, "we have to find a way to cooperate."
> 
> Alexandr: I probably didn't have this awareness at the time. I was thinking: "Maybe he is just testing the waters, just saying it, but does this really make logical sense?"
> 
> David: Why did you think it didn't make logical sense at the time?
> 
> Alexandr: Meta's iconic acquisitions are Instagram and WhatsApp, both very clear product-logic-driven decisions. And Scale is a deep enterprise-level, government-level sales company, a completely different business model from Meta. So the logic I thought of at the time was: if Meta wanted to enter those fields, that might make sense, but this was not that situation—Mark was obviously mainly interested in getting the Llama project and the overall AI project on track, he wanted talent.
> 
> David: Yes, he wanted talent. The whole thing was shrouded in some ambiguity and uncertainty.
> 
> Alexandr: When did it become clear? When we started talking numbers, I thought: "Oh, okay, so it's real." That's when things became clear.
> 
> ## Chapter 13: Leaving Scale AI, Joining Meta
> 
> David: At first you were still thinking "does this make sense," and then you continued talking, how quickly did he convince you this was a good path?
> 
> Alexandr: Part of the reason is that the deal structure itself was fascinating. The plan we ultimately landed on was: Meta invests and holds a 49% stake in Scale, Scale continues to operate independently, and I and a few others join Meta.
> 
> I was a bit incredulous about this plan at the time, because it was a very peculiar deal, but I think this is really a "triple-win" design: Scale's shareholders got a good result, Scale continues to exist, and I truly believe Scale's best days are still ahead.
> 
> David: I am curious, could he change your mind in one phone call? Or was it more like "Mark, I need to think for a few days, this is too crazy"?
> 
> Alexandr: It was several weeks of thinking: "Is this real? Does this make sense? If it makes sense, how do I feel about this?" The dominant emotion was: this is too crazy. Then as I started to seriously consider whether to make this decision, that itself was not simple. Colleagues inside Scale, our investors, there were many important conversations to have.
> 
> Giving up or handing over your "child" is of course very difficult. How many years did you work at Scale?
> 
> From founding to... nine years. And you are still young, not yet 30, this is almost one-third of your life.
> 
> Paul Graham has a saying that has been around for a long time: if you view the company as your life's work, your behavior will be completely different. I did sincerely treat Scale as my life's work, throughout the entire time I worked on Scale. It was a difficult emotional process.
> 
> But what ultimately made me think it through was a combination of several aspects: this is really a good result for all parties, and at the same time I saw the potential of what could happen at Meta. And at that time, Meta indeed looked a bit like a "mess" in terms of AI, possibly not the most attractive place to work in AI—but this precisely triggered my entrepreneurial instinct: you can basically rebuild it, like re-founding this lab.
> 
> David: Yes. Precisely because there is so much work to do, it clears space for you to build many things from scratch, establish the right principles, build the right culture, creating conditions for building something extraordinary.
> 
> ## Chapter 14: Rebuilding Meta AI Lab from Scratch
> 
> David: Did you and Mark discuss it this way? Like "we need to press the reset button here"?
> 
> Alexandr: He didn't look like he wanted minor tweaks. His attitude was: "This path doesn't work, tear it down completely and start over." It was indeed so. The entire conversation was evolving. I firmly believe one thing—which Elon has also proven at different stages of his career—is: you only need an extremely small, extremely powerful, highly technical, very flat team, achieve this, and you can achieve very, very large results very, very quickly.
> 
> This is something I am very confident in, especially in the AI field, this is the right way—especially given we face the necessity of moving fast. This was one of the things I talked about a lot with Mark early on, and I think he was excited about it, ultimately becoming one of the core pillars of our overall strategy: small and flat, highly technical teams, extremely high talent density, and how fast we can move under these premises.
> 
> David: From the outside, Mark's approach looks like "going all in"—that's how I understand him. Have you read the book "The Mind of Napoleon"?
> 
> Alexandr: No.
> 
> David: It was published in 1957. I found it because of an interview—I think it was Tyler Cowen interviewing Sam Altman, asking him what the most important book he read that year was, he said "The Mind of Napoleon." This book is hard to find, 300 pages entirely of Napoleon's own words, organized by theme. In the book he repeatedly emphasizes: hesitation is fatal. But he also says, before taking action, he conducts a lot of deep consideration, examining from every angle to ensure the right decision is made. I feel an echo of Mark's way of operating: doing a lot of deep consideration, ensuring you are on the right path, but once a decision is made, going all out without hesitation. I feel this echo, right at this moment—he comes to you, and then you start rebuilding the entire organization.
> 
> Alexandr: I think he has truly internalized the belief that "failure doesn't matter, what matters is how big you win when you win." This underlying logic allows him to act truly fearlessly in many situations, because he very calmly assumes a large amount of risk.
> 
> David: He has a wonderful quote: in a constantly changing world, what is the biggest risk? Doing nothing.
> 
> Alexandr: Yes, he has indeed deeply internalized this. Of course, there exists a parallel world where Meta looks more like Google, slower, more bureaucratic, looking very different as a company. But precisely because he has so deeply internalized these lessons, it keeps the entire organism dynamic, vital, and alive.
> 
> ## Chapter 15: From Founder to Coach
> 
> David: You were the founder of Scale, running the company hands-on; now you are rebuilding this lab, working with Mark—how is your way of working different in these two environments?
> 
> Alexandr: At Scale, I deeply internalized a concept—although I can't remember where this advice came from—which is: if you want to be a good founder, you must be competent in every role. I had very strong opinions on how every little thing in the company should operate. This created a company that was largely the product of my personal effort and personal perspective, this style permeated every aspect of the company, but it also made me a bottleneck in many important links. This has pros and cons for a company.
> 
> I think, if you are a first-time entrepreneur, you should absolutely adopt this method, because the probability of success is much higher than not doing so. Scale was my first company.
> 
> David: Right.
> 
> Alexandr: But at Meta, the situation is completely different, because we must accomplish an extreme amount of things in an extremely short time. I don't have the luxury to "be competent in every role." I wasn't an AI researcher originally, nor a product designer, nor did I possess various professional capabilities crucial to the organization's success.
> 
> So at Meta, I adopted a completely different philosophy—the core is: how do we create an environment that can attract top outstanding talent from various fields? Recruiting the best, most excellent people is the top priority.
> 
> David: Why do you think you can achieve this at Meta? Is it just because resources are more abundant?
> 
> Alexandr: I think the reason MSL was extremely attractive in its early stages was that it offered an opportunity to start from scratch, building a lab with an extremely small team, and we had a clear direction—personal super intelligence. This direction resonated with many people. I think everyone working in AI hopes it truly helps ordinary people, hopes it means something to their mothers, grandmothers, grandfathers, hopes it has practical significance for every person in the world.
> 
> At the same time, because we are doing this leveraging Meta's resources, we are not limited by compute power, not limited by infrastructure, not limited by distribution channels—these are precisely the most common bottlenecks for startups. But this is also a real opportunity to leave your mark, exert profound influence, and shape things with your own hands. I think this is quite exciting for many people.
> 
> David: So this is why Mark is willing to spend seemingly astronomical sums to recruit a few top talents—because he knows these talents will in turn attract more top talents. But you need this starting point.
> 
> Alexandr: Completely correct. Another point that might be underestimated: the stock price increases of OpenAI and Anthropic are very large, so if you consider how much these people could earn by staying there, it is also a considerable figure. For many talents, compensation might ultimately be roughly equal. Of course, we often ignore how much the stocks of employees and researchers who joined these labs early have appreciated, and what their current compensation levels are.
> 
> There are only two beliefs behind our founding of MSL: First, recruit the smartest, most excellent people; Second, personal super intelligence—how to create an environment at Meta and MSL where everyone can do their best work.
> 
> At Meta, what I do is more like traditional management work. The core question I have been thinking about is: how to create an environment where these talented, extremely special people can fully unleash their talents? I play more of a coach role, whereas at Scale, it was completely different—I was the "author," the one who needed to impose my perspective on everything we did.
> 
> David: Tell me more about this "coaching feel."
> 
> Alexandr: I think my job is, first, to help establish the North Star and forward direction—that is, personal super intelligence, winning the consumer AI race, building AI that is meaningful to everyone; second, to discover special talent within the team, and then give them the environment and resources to fully unleash their talents.
> 
> A mistake large companies often make is: excellent talent is clearly there, but doesn't get the opportunity to flex their muscles, so they often choose to leave, going to smaller companies or startups.
> 
> David: You previously mentioned Jeff Bezos' shareholder letter. This is also a point he repeatedly emphasizes, in the books I've read about him and in his own shareholder letters—if excellent people cannot build things, they will leave.
> 
> Alexandr: Indeed. AI is a very special thing. The word "research" now seems to have lost its original meaning, but by definition, humans do not yet know the capability boundaries of these models, nor do we know what can be made with them. They are truly unknown entities we do not yet fully understand. So this is really science, really research—we are exploring the limits of this technology, exploring what can be built with it, this process requires a lot of trial and error, requiring smart people to burst forth with extraordinary insights.
> 
> This is not a process that can be advanced with extremely high-pressure, extremely intensive work methods. People need space, time, and resources to do their best work. I think about this in many dimensions. For example, Nat Friedman, he is absolutely brilliant—as I said before, if Muse has one person's core vision running through it, it is him. Now it is important to put him in a position where he can fully express all of this. We also have some outstanding researchers with very exciting ideas—how to create an environment where they can explore freely? Most attempts will not succeed, but some of them will achieve astonishing results, and these successful cases are enough to prove that all investments are worth it.
> 
> There are two types of problems, or two modes: one is called diamond mining, the other is building skyscrapers, depending on the return distribution of what you are doing. Some fields have a super power-law distribution—most of what you do has almost no value, but a few things might be worth a million times more than everything else combined. Venture capital is a typical diamond mining industry—everything is about identifying those special ideas and carrying them through to the end.
> 
> This is easy to say, just go to your 19th birthday party and see.
> 
> ## Chapter 16: Future Billionaires at Alexandr's 19th Birthday Party
> 
> David: Ha, this is truly incredible. Some listeners don't know what I was just talking about—can you briefly say who was at your 19th birthday party, and why investing in that table of people would be a good deal?
> 
> Alexandr: This is really a crazy thing. At my 19th birthday party, I was still a freshman at MIT, it was January, and I was doing a winter internship at a trading company called HRT. We had about a dozen people, twelve or so, almost all Harvard or MIT students, most of whom had known each other before through various math, science, or computer competitions.
> 
> Among them was Jeffrey Yan, who later went on to found Hyperliquid, very successful; Scott Woo later founded Cognition, also very successful; another person was Jesse Zang, who founded Decagon, also quite successful; another was Vicky Yei, who I think is a very outstanding researcher, now working at Anthropic. Everyone was extremely excellent.
> 
> I recently chatted with the founder of HRT—HRT itself has now developed very impressively, incredibly so.
> 
> David: What does HRT do?
> 
> Alexandr: HRT is a proprietary trading firm, trading with its own money. In the second quarter of 2020, they set record figures of $11 billion in revenue and about $8 billion in profit. The company has fewer than a thousand employees, truly breathtaking. I recently met with the founder of HRT, they said they were considering making some private investments, and I said: "Just invest in those interns, give seed round checks to all the interns."
> 
> I grew up participating in math, science, and computer competitions and Olympiads, meeting various people from across the US in this circle. Watching those people who competed together back then become so successful now really has a surreal feeling.
> 
> ## Chapter 17: Diamond Mining vs. Building Skyscrapers
> 
> David: It must be exciting to see what happens to your group in the next few decades. I interrupted you earlier—is there anything else you want to add about diamond mining and skyscrapers?
> 
> Alexandr: The core is: is its return a power-law distribution—where the value of a few things is a million times the investment—or is it more like building a skyscraper? I would categorize data labeling or data business as the latter, Amazon Prime delivery is also a similar example: returns are relatively linear, you must establish a process and execute it to the extreme.
> 
> If you are building a skyscraper, it is a highly operational business, everything is about squeezing every bit of efficiency out of the process—finding ways to improve it by another 2%, another 1%, another 0.5%, constantly optimizing.
> 
> And diamond mining is completely different, it's like saying: "Alright, everyone go explore, try all the crazy ideas," and then you just nurture those ideas that look promising.
> 
> David: Is this the feeling of what you are doing at Meta now?
> 
> Alexandr: Yes, this is the feeling I have at Meta. And building skyscrapers, generally speaking, was the feeling I had at Scale. This is a fundamental paradigm difference.
> 
> ## Chapter 18: Why Alexandr Has Over 200 Direct Reports
> 
> David: Returning to the differences in management style—your management style at Meta and Scale is fundamentally different. Why do you have so many direct reports? Seems like over 100? What is the specific number?
> 
> Alexandr: Over 200.
> 
> David: How can that be? Why?
> 
> Alexandr: Going back to the founding concept of MSL—what kind of lab did we want to build? The core idea was: how to maintain extremely high talent density? How to let these excellent people do their best work here? How to build an organization centered on technology, where technical people can truly unleash their potential, even complete their life's work?
> 
> One of the design principles around this is: extreme flatness. In fact, all researchers who joined the TBD team report directly to me. This is more of a manifesto—TBD is a symbol of anti-bureaucracy. We recruit talented people who have clearly proven in their careers that they can do excellent work. I don't need to manage them in detail, they are fully capable of completing excellent work autonomously.
> 
> But we do need technical leadership. Some people are responsible for setting clear technical directions for their respective groups. We have a "Pod" structure—each Pod has a technical leader who has the final say in technical decisions. So technical leadership exists, but from the perspective of "people management" or bureaucracy, we are anti-bureaucratic.
> 
> On the other hand, we do make many collective decisions. This is an extremely talent-dense team, everyone has brilliant insights, and we hope to find the path forward that everyone collectively believes is most correct through debate and discussion. Everyone in the lab has done excellent things—we hope to listen to everyone's opinions.
> 
> This is of course a very unconventional management style. If I say I am the best manager for over 200 people, that is obviously lying. But the point is—they don't need managers at all, they need a good environment.
> 
> They are all extremely smart, have super strong abilities, they just need space to display themselves.
> 
> David: So maybe not "coach." You are more like a guardian of a good environment, allowing them to do great work.
> 
> Alexandr: Yes, exactly.
> 
> David: Alex, today was really wonderful, thank you for taking the time.
> 
> Alexandr: Thank you.

### Related Stocks

- [META.US](https://longbridge.com/en/quote/META.US.md)
- [FBL.US](https://longbridge.com/en/quote/FBL.US.md)
- [FBX.US](https://longbridge.com/en/quote/FBX.US.md)
- [FBY.US](https://longbridge.com/en/quote/FBY.US.md)
- [FBYY.US](https://longbridge.com/en/quote/FBYY.US.md)
- [MEIB.US](https://longbridge.com/en/quote/MEIB.US.md)
- [METD.US](https://longbridge.com/en/quote/METD.US.md)
- [METG.US](https://longbridge.com/en/quote/METG.US.md)
- [METQ.US](https://longbridge.com/en/quote/METQ.US.md)
- [METU.US](https://longbridge.com/en/quote/METU.US.md)
- [METW.US](https://longbridge.com/en/quote/METW.US.md)
- [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md)

## Related News & Research

- [Meta's 'Rapture' could send wealthy employees into early retirement](https://longbridge.com/en/news/305575521.md)
- [Brazil electoral court fines Meta, sees no deliberate election bias](https://longbridge.com/en/news/305164398.md)
- [Meta's $27B Off-Balance-Sheet AI Deal Reminds Steve Eisman of Enron— Joseph Carlson Disagrees](https://longbridge.com/en/news/305341988.md)
- [Muse can now place calls to U.S. businesses — if you pay](https://longbridge.com/en/news/305378857.md)
- [Meta fined $887M by Italian Data Protection Authority, Bloomberg says](https://longbridge.com/en/news/305607893.md)

---
> **Disclaimer: This article is for reference only and does not constitute any investment advice.**