Building Muse, Working with Mark Zuckerberg & The Future of AI | Alexandr Wang, Meta

Summary · 5 min read instead of 1 h 12 m

Alexandr Wang is the chief AI officer of Meta and cofounder of Scale AI, which he ran for nine years from founding. He joined Meta after Mark Zuckerberg reached out following the Llama 4 setback, in a deal where Meta took 49% of Scale, and now leads Meta Super Intelligence Labs, which built the Muse personal agent.

Wang argues that personal agents, not coding agents, were the right bet for Meta, because its products represent what people want to do and it reaches three and a half billion people daily. He describes how Muse went from a prototype built in one to two weeks to launch seven months later, through a spreadsheet of more than 100 model behaviors with launch-blocking thresholds and a team of under 200 people. He credits Nat Friedman's taste as the single point of view behind the product and says his own meme posting on X likely helped growth. Wang recounts how Zuckerberg recruited him and how he is rebuilding Meta's AI lab around high talent density and anti-bureaucracy, acting more as a coach than the auteur he was at Scale. He believes AI agents can let everyone keep the ambition they had as children.

Minute by minute

1. 0:00

   Muse grew from a 2025 memo. Wang says Muse traces back to the Personal Super Intelligence memo Mark Zuckerberg published in June 2025, which Wang and Nat Friedman worked on. At the start of 2026, Opus 4.5 and OpenClaw showed the potential of agents. Friedman was the first in MSL to try OpenClaw; his agent watched security footage to check he was drinking water, and Wang treated Friedman's early-adopter instincts as a clear signal.

2. 3:30

   A prototype in two weeks, launch in seven months. Wang's first OpenClaw use was Friedman's psychoanalysis prompt, which he likens to "three years of therapy" using his email and photos. In February both wrote memos to the Meta board; Wang called the super agent a consumer endpoint and Friedman wrote "trust is a must". The team built a prototype in one to two weeks, then spent seven months making an unreliable demo dependable. Zuckerberg used it at home, with his kids and for MMA training tips.

3. 9:07

   Meta bet on personal, not coding. Wang says industry wisdom in February held that coding agents were the only form factor that mattered, with Anthropic far ahead. He told Zuckerberg early that Meta's apps represent what people want to do, while AI will absorb what they have to do. Combined with distribution to three and a half billion daily users, he argues consumer and personal AI was where Meta could do something special.

4. 12:30

   Sanding the model against a spreadsheet. Wang says the team listed more than 100 specific behaviors a personal agent model needed, built evals for each and held reviews to fix weak spots. He links this to Meta's culture of optimizing whatever it measures, from onboarding and referral rates in the pre-AI era to model details for Muse now.

5. 16:44

   Patience through seven months of doubt. Senra cites Jeremy Stern's Colossus profile, where Zuckerberg says his skill is building teams and improving products slowly over a long period. Wang credits Zuckerberg's patience and restraint, noting Wall Street doubted Meta's AI spending throughout. He says personal agents sit on a knife's edge: reliable magic makes a great product, while unreliable experiences make users think it is trash.

6. 21:45

   Numerical thresholds decided the launch. Each spreadsheet row had a launch-blocking threshold; many checkpoints were green on 80 rows and red on 20. The first all-green checkpoint was a specially trained version of MewSpark 1.3, and testers also felt it was clearly better. Wang says fewer than 200 people worked across model and product, despite executives seeing Muse as one of Meta's most important launches of the year.

7. 23:19

   Great products need one point of view. Wang calls Muse "a single work of art" and criticizes large-organization products as amalgamations where every PM jams in features. He says Friedman's taste shaped what Muse became. Senra compares this to Steve Jobs on the anniversary of his death. Wang says Meta has only disclosed millions of users, while he retweeted graphs implying it is the fastest growing consumer AI app.

8. 25:50

   Memes to break through and make Meta cool. Wang says distribution alone could not make Muse break through the stream of AI launches, and he also wanted to make Meta cool again. He wrote the memes himself, drawing on friends who think about the internet all day, and found that "the internet rewards risk" and surprise. Racy memes of the Jolly mascot drove intrigue. He thinks it affected growth, though Meta lacks tracking to prove it.

9. 34:30

   Retention comes from a trust fall. Wang says engaged users give Muse slightly bigger tasks each time it succeeds, building trust step by step. Senra describes a friend whose Muse retrieved an ID left on a Flexjet and couriered it to his office, shared as screenshots. Wang argues modern consumer products must work as screenshots, and the mascot in each one marks it as Muse.

10. 39:15

    AI can restore childhood ambition. Wang uses Muse as a second brain through workflows. He believes personal agents will be the bridge to people's wants, saying Muse can give everyone "the adulthood they dreamed of in childhood". He argues most adults become zombies who lose agency, and imagines billions of people with abundant intelligence pursuing ever bigger dreams, as Elon Musk and Zuckerberg do.

11. 45:11

    Zuckerberg's call after Llama 4. Alex Schultz introduced Wang to Zuckerberg around 2021 or 2022; the first meeting took a year to schedule. After Llama 4, which Wang calls an internal disappointment, Zuckerberg called for advice. The deal took five to six weeks: Meta invested and owns 49% of Scale, which continues, and Wang and a few others joined Meta. Leaving his life's work of nine years was hard, but Meta looking like "damaged goods" in AI appealed to his entrepreneur side.

12. 55:10

    Rebuilding the lab as a coach. Wang says small, flat, highly technical teams with high talent density became a pillar of MSL's strategy. At Scale he tried to do every job and became a bottleneck; at Meta he focuses on hiring top people and giving them room. He says Meta's resources and mission attracted researchers, and pay was often net neutral given OpenAI and Anthropic stock gains.

13. 1:05:16

    Diamond mining versus skyscrapers. Wang contrasts power-law work like research and VC with linear, operational businesses like data annotation, saying Scale was skyscrapers and Meta is diamond mining. His 19th birthday party during an HRT internship included Jeffrey Yan of Hyperliquid, Scott Wu of Cognition and Jesse Zhang of Decagon. He says HRT made $11 billion revenue and $8 billion profit in Q2 2026 with under 1,000 people.

14. 1:09:12

    More than 200 direct reports. Wang says all researchers in the TBD group report directly to him as a statement against bureaucracy. Pods have technical leads who tie-break technical decisions, and the lab debates major choices as a group. He admits he is not the best manager to more than 200 people, but argues they need a great environment and "the room to cook" rather than managers.

Published Oct 11, 2026.