Alexandr Wang recounts the creation of Muse, Meta’s personal AI agent, from its origins in a “Personal Super Intelligence” vision through a seven-month refinement sprint, the unconventional acquisition of Scale AI, and his shift from founder-mode micromanagement to stewarding a flat, high-talent-density lab at Meta.
The vision: Personal Super Intelligence as Meta’s unique AI bet
In June 2025, before agents were viable, Mark Zuckerberg, Nat Friedman, Wang, and others wrote a memo called “Personal Super Intelligence” arguing Meta should build AI that elevates daily life rather than just coding or work tasks.
The thesis: Meta’s heritage (connecting people, surfacing interests) and distribution (3.5B daily users) make it uniquely positioned to win consumer AI — the “things we want to do” — while competitors chase enterprise and coding.
This memo became the north star for Meta Super Intelligence Labs (MSL) and foreshadowed Muse’s eventual form factor.
The OpenClaw spark: euphoria, terror, and a psychological audit
In early 2025, Nat Friedman tried OpenClaw (an open-source agent framework) and described the experience as “somewhere between terrifying and euphoric” — e.g., it watched security footage to verify he drank water and praised him.
Friedman’s conviction mattered: he was an early Stripe customer, helped create GitHub Copilot, and has a track record of spotting developer-tool inflection points early.
Wang tried OpenClaw himself, starting with a Friedman-built prompt that ran a multi-round “full psychoanalysis” using his email, photos, and other data — “the equivalent of 3 years of therapy compressed into one moment.”
Both wrote memos to the Meta board in February 2025: Wang argued the “super agent” is the final consumer product; Friedman wrote “The Claw, The Law, and Trust Is a Must,” which shaped product principles.
Prototype to product: seven months of sanding edges
Within 1–2 weeks of the February memos, the team built a working prototype containing the Jolly mascot, early Muse Charm, and core tabs; they demoed it to the board.
The subsequent seven months (February–September 2025) focused on turning a fragile, sometimes-magical prototype into a reliable consumer product for billions.
Zuckerberg experienced the same euphoria/terror cycle personally — using it for MMA video analysis, home tasks, and staff-meeting show-and-tells — and backed the long refinement despite Wall Street skepticism.
Wang credits Zuckerberg’s patience and “Terminator” persistence: he internalizes that the biggest risk is not taking risk, and he empties the clip once convinced.
Model development: a spreadsheet of 100+ behaviors, evals, and checkpoints
The team enumerated 100+ specific behaviors the model needed for a personal agent (e.g., tool use, memory, communication style, ambiguity handling).
They built evals for every row, ran continuous reviews, and iterated the model (Musespark 1.3) until every row hit a “green” launch threshold.
Product experience proved extremely sensitive to model quality: small regressions broke user trust; reliability was the gate to retention.
Team size stayed under 200 people (model + product) — intentionally small to preserve a single point of view and avoid “Frankenstein” design-by-committee.
Marketing that breaks through: memes, screenshots, and the Jolly mascot
Wang leaned into “unhinged” X posting — memes, racy humor, and the Jolly character — to cut through the torrential AI-release noise and make Meta “cool” again.
Organic screenshots became the best ads: users sharing Muse completing complex multi-step tasks (e.g., retrieving a lost ID via courier) with “I don’t want to be involved, you handle it.”
The Jolly mascot in every screenshot acted as instant brand recognition; Wang’s meme instinct came from years of internet immersion and a circle of “schizophrenic, terminally online” friends.
Retention hinged on a “trust fall” loop: users start with tiny tasks, see them work, then escalate to larger delegations — every success deepens trust.
The Scale AI deal: a five-week, win-win-win refounding of the lab
Zuckerberg cold-messaged Wang in 2025 (after a year of scheduling a first meeting in 2021–22) asking for advice post-Llama 4 disappointment; Wang sent extensive strategy thoughts.
The conversation evolved into a deal: Meta invests for 49% of Scale, Scale continues independently, Wang and a few others join Meta to rebuild the AI lab.
Wang initially thought the logic didn’t fit (Scale = enterprise/government sales; Meta = consumer) but came around over weeks: great deal for Scale shareholders, Scale’s best days ahead, and a rare chance to refound a lab with Meta’s resources but startup speed.
Zuckerberg’s deliberation-then-decisiveness mirrors Napoleon: study every angle, then commit fully and move fast.
Management shift: from founder-bottleneck to environment steward
At Scale (first company, 9 years), Wang internalized “do every job” — his taste permeated everything, but he became the bottleneck.
At MSL, the assignment demanded speed and breadth he couldn’t cover alone (not a researcher, not a designer), so he flipped to “coach/stewward”: set the north star (Personal Super Intelligence), hire brilliant people, give them a flat, bureaucracy-free environment.
Structure: 200+ direct reports (all researchers in TBD group report to him); technical leads run pods and tie-break decisions; group debates replace top-down mandates.
Philosophy: brilliant people don’t need managers, they need room to cook — big companies usually fail by disempowering talent; AI research is “diamond mining” (power-law payoffs), not “skyscraper building” (linear optimization), so the org must nurture high-variance exploration.
The diamond-mining analogy and a legendary 19th birthday party
Wang contrasts “diamond mining” (VC, research: most bets fail, a few return 1M×) with “skyscraper building” (data annotation, logistics: squeeze every percent of efficiency).
MSL is diamond mining; Scale was skyscraper building — different mental models, different management.
At 19 (MIT freshman, HRT trading-firm intern), Wang’s birthday party included 10–12 peers from math/CS Olympiads: Jeffrey Yan (Hyperliquid), Scott Wu (Cognition), Jesse Zhang (Decagon), Vicky Yi (Anthropic researcher) — all now exceptional founders/researchers; HRT itself now does ~$11B revenue / $8B profit with <1,000 people.
The endgame: agency escalators for everyone
Wang believes Muse and personal agents can give every person the “adulthood they dreamed of in childhood” — an escalator of increasing agency instead of the typical crush into “zombiehood.”
The world has been shaped by a few fanatics who made their wants real; abundant intelligence could switch on full agency for billions, yielding a wildly diverse, artistic, “Rick and Morty interdimensional cable” future.
Personal conviction: AI as a second mind that handles the “kludge of life” (gatekeepers, obstacles, chores) so humans dream bigger, accomplish more, and keep the childhood ambition loop alive.
