Meta built Muse in 7 months with under 200 people

1. Memo before agents — ▶ 0:09

   When Meta Superintelligence Labs started, Mark Zuckerberg published a memo on Personal Super Intelligence. Wang, Nat Friedman and others worked on it closely with him. It was June 2025, before agents had happened in any meaningful way. The aim was powerful AI that lifts the human experience broadly, and it already gestured at what Muse became. Then came the slow part. A frontier model takes many months end to end: building the stack, producing datasets, pre-training and post-training. The vision had to wait on a production line nobody could shortcut.

2. Terrified and euphoric — ▶ 1:32

   Opus 4.5 and OpenClaw, at the start of the year, brought agents' potential to the foreground. Nat Friedman was first in MSL to use OpenClaw, and described the experience as "somewhere between terrifying and euphoric". His agent watched him on security footage to check he drank water. Wang ran Friedman's multi-wave psychoanalysis prompt on an agent with access to his email and photos, calling it the equivalent of 3 years of therapy. In February both wrote memos to the board. Wang's called the super agent the "final consumer product".

3. Prototype in two weeks — ▶ 5:40

   The teams built a prototype in 1-2 weeks and demoed it to the board in February. Many decisions that shipped in Muse were already there, including the mascot Jolly. Launch came 7 months later, and the gap was the hard part. The prototype sometimes worked magically but mostly didn't and kept breaking. Turning it into a highly reliable, crafted consumer product was what Wang calls a lengthy and quite painful process. The speed of the idea mattered far less than the grind that followed.

4. Why OpenClaw faded — ▶ 7:42

   OpenClaw peaked and then disappeared. People got magic moments, but for the most part it didn't work well, and interest died down. Wang calls Peter Steinberger a visionary and credits OpenClaw's contributors. Muse's difference was the unglamorous part: grinding out and sanding the details of the combined model-plus-product experience. The team ran many iterations on the model itself. The target was an experience reliable enough to put in front of billions of people, most of whom would have an amazing time.

5. Betting against coding — ▶ 9:16

   In February, Claude and Opus were way ahead on coding, and Anthropic's lead looked like a takeover. Conventional wisdom said coding agents were the only capability that mattered, so the industry crowded in. Wang's counter is that Meta's apps represent what people want to do, as opposed to have to do, and AGI shifts time toward the former. With 3.5 billion daily users, consumer AI was always Meta's natural zone. "It's not work. It's not coding. It's personal." He made this case to Zuckerberg before the Scale deal.

6. The 100-row spreadsheet — ▶ 12:43

   Model development was aimed at personal agents using a spreadsheet of more than 100 specific behaviors and capabilities the ideal agent needs. Each row got evals, targeted training, and regular reviews of the weakest rows. Wang ties this to Meta's pre-AI culture of optimizing conversion, referral rates and every step of onboarding. Once you know what you're measuring, talented people can sand out the details. In the AI era that discipline moved from the funnel into the model itself, tuned to be perfect for the product.

7. Patience under doubt — ▶ 17:41

   Zuckerberg showed patience and restraint from February through September. Earnings calls alluded to personal agents, and Meta wasn't hiding it. Inside the company there was doubt, and the dominant Wall Street story was that Meta was burning money and might not pull anything off in AI. Wang says the product rested on a knife's edge. If it reliably delivers magic it could be among the greatest consumer products ever, but with nondeterministic models, unreliable experiences make people think "it's trash". Seven months is a long time in AI.

8. Launch gate — ▶ 22:00

   Readiness was partly numerical. Each spreadsheet row had a launch-blocking threshold and a higher threshold for good. Many checkpoints were green on 80 rows and red on 20, which wasn't good enough. The first checkpoint green across the board was a specially trained version of the model. It was also experiential: the team could tell it was far better than anything before. Product quality is very sensitive to model quality, including how reliably it communicates and checks back with users. The whole effort involved under 200 people across model and product.

9. One point of view — ▶ 23:38

   Everyone knew Muse would be among Meta's most important products that year, yet the team stayed small and focused. Wang calls it a single work of art needing one strong point of view. Larger organizations, and now many AI labs, ship a Frankenstein: every PM jams in their own idea until it blends into a smoothie users can feel. Nat Friedman's taste and sensibilities dictated what Muse became. The host compares this to Apple as "Steve Jobs with 10,000 lives". Wang gives only "millions" for usage, and says Muse is likely the fastest-growing consumer AI app ever.

10. Memes as breakthrough — ▶ 26:00

    Distribution wasn't enough, since users face a torrential stream of AI launches. The team agreed they needed a cultural breakthrough and wanted to make Meta cool again. Wang's memes, built around the Jolly mascot, started as retweets of use cases. He found some quick, unplanned tweets did best, remembering that "the internet rewards risk" and surprise. Friedman joked they were learning about Alex's mind, not Muse users. Wang has no tracking but believes the posts genuinely helped growth. The mascot gave him endless situations, such as a briefcase-carrying Jolly captioned for Monday morning.

11. Sell the benefit — ▶ 32:21

    The recurring question in building Muse was why someone would care. Most people don't care about OpenAI, the research, or Meta's research. They care what the product does for them. The host argues Anthropic and ChatGPT ads fail by being about themselves and showing the logo or two people at a desk. His fix is to show the benefit and run the millions of real use cases. He cites Hopkins's Scientific Advertising, the early-1900s book on human nature. Wang agrees and sees Muse's wow-moment onboarding as the product doing that work.

12. Trust fall and screenshots — ▶ 34:40

    Retention comes from a trust fall. Users give the agent a small task, see it work, then give a bigger one, repeatedly. Successful Muse users follow that path, and each step has to work every time. Wang also thinks modern consumer products must "work on the screenshot". Muse screenshots circulate through group chats, and Jolly in the frame identifies them instantly. The host's example is a friend running a $100 billion company whose ID was left on a Flexjet. Muse arranged a courier and retrieval with him saying he didn't want to be involved.

13. Escalator of agency — ▶ 39:53

    Wang believes the personal agent is a long-term form factor because humans have wants and dreams, and AI becomes the bridge to accomplishing them. He argues adults become zombies as ambition is beaten out of them after school and work. Muse should give everyone "the adulthood they dreamed of in childhood". The aim is an escalator of agency rather than a crushed one. Until now, a small number of fanatics shaped the world. He envisions billions with abundant intelligence, an interdimensional-cable world made real, which he calls crazy in a very good way.

14. Zuckerberg's call — ▶ 45:19

    After Llama 4, which Wang calls an internal disappointment, Zuckerberg messaged asking for a call. The first conversations sought advice on what Meta should do. Wang assumed a deal made little sense, since Scale sold to enterprises and governments, and Zuckerberg wanted talent. It became clear once numbers came up. From first contact to announcement took 5-6 weeks. Meta owns 49% of Scale, which continues, and Wang joined Meta after 9 years. Meta's AI looked like "damaged goods", but that triggered his entrepreneur side. He saw a chance at a refounding.

15. Flat lab, 200 reports — ▶ 58:14

    Wang says that at Scale he tried to do every job, which made him the bottleneck. At Meta he acts more like a coach, hiring brilliant people and setting the north star. He says AI research is closer to diamond mining, nurturing crazy ideas with power-law payoffs, whereas Scale was skyscraper-building, squeezing efficiency out of linear payoffs. TBD researchers report directly to him, more than 200 of them, with pods and technical leads tie-breaking. He admits he isn't the best manager of 200+ people, but says they "need the room to cook".

Consumer AI, not coding, is where Meta can win. That is Alexandr Wang's bet, and he placed it when the rest of the industry was betting the other way. In February, Claude and Opus were way ahead on coding and exploding. Anthropic's lead looked like a takeover. The conventional wisdom held that coding agents were the only capability that mattered for advanced AI, and the pull to crowd into that use case was strong. Wang calls it a very scary moment. Meta did not follow. It aimed its model roadmap at personal agents instead, on the conviction that this was the ground it was uniquely positioned to take.

The logic is about what AI frees people to do. Wang gave Mark Zuckerberg this framing before the Scale deal closed. AGI shrinks the things we have to do and leaves us the things we want to do. Meta's apps sit in the second category. Wang follows his interests, his hobbies and the people he cares about through Instagram, WhatsApp and Facebook, and says "Meta's products represent the things that we all want to do." Add distribution to 3.5 billion people a day and consumer AI is the natural zone. The June 2025 memo on personal superintelligence set this out before agents had happened in any meaningful way.

A thesis is not a product, and the proof arrived as a shock. Opus 4.5 and OpenClaw, at the start of this year, made the potential of agents visible. Nat Friedman was the first person in Meta Superintelligence Labs to try OpenClaw. He called it somewhere between terrifying and euphoric. His agent watched him on security footage to make sure he drank water. Wang then ran Friedman's psychoanalysis prompt on an agent that had his email and photos. It felt like 3 years of therapy compressed into one moment. Both men wrote memos to the board in February. Wang's said the super agent looked like the final consumer product. A prototype followed within 1-2 weeks, and many of its decisions survived into the shipped product.

Then the hard part started. The prototype sometimes worked magically but mostly didn't, and it kept breaking. OpenClaw had followed the same arc. It peaked, gave people magic moments that mostly didn't hold, and disappeared. Wang respects its creator, Peter Steinberger, and calls him a visionary. But he locates Muse's advantage in the unglamorous work. The team ground out and sanded down the details of the combined model and product. Agents, he says, rest on a knife's edge. If they deliver reliably, they could be among the greatest consumer products ever. If they don't, people decide the product is trash.

The mechanism was a spreadsheet. It listed more than 100 specific behaviors and capabilities the ideal personal agent needs. Each row got evals, targeted training, and regular reviews of the weakest rows. Each row also had two thresholds, one that blocked launch and a higher one that counted as good. Many checkpoints were green on 80 rows and red on 20. That was not good enough. The first checkpoint green across the board was a specially trained model, and the team could also feel it was far better than anything before. Wang ties this to Meta's pre-AI habit of optimizing conversion, referral rates and onboarding flows. The same discipline now runs on the model. The product is highly sensitive to model quality, including how reliably the agent communicates and checks back with users.

The organization around that work was deliberately small. Under 200 people covered model and product combined. Everyone knew this could be among Meta's most important launches of the year, and the executive team still held back from letting the project balloon. Wang calls Muse a single work of art. Products from large organizations, and increasingly AI labs, turn into a Frankenstein, with every PM jamming in a pet idea until the result is a smoothie users can feel. Muse needed one cohesive experience. Friedman's taste and sensibilities dictated what it became. Zuckerberg's role was patience. From February to September, Meta alluded to personal agents on earnings calls while Wall Street asked whether it could pull anything off in AI. Wang credits him with restraint and trust in the process.

Reliability earns retention, but it doesn't earn attention. Users face a torrential stream of model and product launches, and distribution alone doesn't break through. So the team decided it needed a cultural moment, and Wang became the experiment. He started by retweeting use cases. Then he found that quick, barely considered posts did best, and remembered that the internet rewards risk and surprise. The mascot, Jolly, gave him endless situations to put in front of people. Friedman joked that the team was learning about Alex's mind rather than about Muse users. Wang has no tracking, but he believes the posts genuinely helped growth.

The deeper marketing principle is simpler. Ask why someone would care. Most people don't care about OpenAI, its research, or Meta's. They care what the product does for them. That is why Wang thinks the product sells itself once a user reaches a wow moment early. The retention mechanism he describes is a trust fall. Users hand the agent a small task, watch it work, then hand it something bigger. Each step has to work every time. Modern consumer products also have to work as a screenshot, shared in group chats and online, and a mascot in the frame makes any screenshot recognizably Muse.

What Wang thinks all this adds up to is bigger than a product. Adults, he says, are zombies. Ambition gets beaten out of people by school, jobs and commitments. A personal agent should keep a child's escalator of agency running instead of crushing it. Until now, a small number of fanatics have shaped the world. He pictures billions of people with abundant intelligence, producing a world that is artistic, diverse and crazy in a very good way.

The lab that built it was itself a rebuild. After Llama 4, which was an internal disappointment, Zuckerberg called Wang and asked for advice. Wang assumed a deal made little sense, since Scale sold to enterprises and governments and Meta clearly wanted talent. It became clear once numbers were discussed. Meta took 49% of Scale, Scale continued, and Wang joined Meta. He saw a lab that looked like damaged goods and an opportunity to refound it. His model was a small, flat, highly technical team with high talent density. Researchers report directly to him, more than 200 of them, with pods and technical leads breaking ties. He describes research as diamond mining, where a few payoffs can be a million times larger than the rest. Scale, by contrast, was skyscraper-building, where efficiency gains are linear.

The obvious objection is that none of this is proven where it counts. Wang says Muse has millions of users, and then concedes that millions could be hundreds of millions. He offers no retention or usage numbers. He cannot measure whether his memes mattered. OpenClaw also generated wow moments and shareable stories before it faded. Muse's answer to that precedent is reliability, but the evidence offered is the launch gate and user testing, not long-term behavior. Wang is also candid about the management model. He says he would be lying if he said he was the best manager of 200-plus people, and that they don't need managers, they need room to cook. He describes the lab as diamond mining, yet the Muse process he praises looks like the skyscraper side of his own distinction: a spreadsheet and thresholds, polished until green.

The open question is whether the trust fall holds at the scale Meta wants. A product that has to be right every time for billions of people is a harder test than one that has to be magical for millions.