Creating Muse: The Fastest-Growing AI Product Since ChatGPT | Alexandr Wang, Meta
Summary
Muse began with Meta Superintelligence Labs’ June 2025 “Personal Superintelligence” thesis: build powerful AI that “lifts the human experience” by helping people pursue what they want, not merely automating work. Wang argues Meta is uniquely positioned for that consumer future because Instagram, WhatsApp, and Facebook already map people’s interests and relationships, while the company reaches 3.5 billion daily users. The strategic bet was therefore personal—not work, coding, or enterprise AI.
The catalytic moment came in early 2026, when Opus 4.5 and OpenClaw showed Wang and Nat Friedman what personal agents might become. Friedman’s agent monitored whether he drank enough water through security footage; Wang let his own agent conduct what felt like “3 years of therapy” using email, photos, and successive psychological probes. Those “somewhere between terrifying and euphoric” experiences led Wang to describe the super-agent as, in many ways, “the final consumer product” and led the team to build a prototype within one or two weeks.
Meta then spent seven months turning an occasionally magical prototype into a reliable consumer product—a period Senra calls roughly a decade in conventional technology time. The team enumerated more than 100 required behaviors, built evaluations for each, and refused to launch while even 20 rows remained below threshold. Its first green-across-the-board checkpoint was a specially trained Musepark 1.3, reflecting Wang’s view that the product experience is extraordinarily sensitive to model quality, communication, and knowing when to ask the user for guidance.
Muse was built by fewer than 200 people because Meta treated it as a “single work of art,” not a committee-produced feature bundle. Wang says large-company products often become a “Cronenberg amalgamation” or “smoothie” of conflicting product-manager agendas; Muse instead carried one cohesive point of view, with Friedman’s taste especially influential. Zuckerberg’s contribution was patience: he tolerated seven months of doubt and a Wall Street narrative that Meta might not pull anything off in AI.
Wang would disclose only “millions” of users, while pointing to graphs he had amplified that implied Muse was the fastest-growing consumer AI app of all time. Distribution created awareness, but growth depended on an early wow moment and an escalating “trust fall”: users hand the agent a small task, see it succeed, and gradually entrust larger problems. The clearest word-of-mouth specimen was a CEO who told the agent to recover an ID left on a Flexjet without involving him; screenshots showed it arranging the courier and navigating the recovery to his office.
Muse’s breakout marketing combined concrete user benefits, screenshot-native design, and Wang’s increasingly risky meme campaign. His feed was personally written, not generated by “a hundred Muses”; he wanted both Muse and Meta to break through a “torrential stream” of AI launches and become culturally interesting again. The lesson he rediscovered was that “the internet rewards risk” and surprise, while the Jolly mascot made every shared screenshot instantly recognizable as a Muse interaction.
For investors, Muse is also part of Zuckerberg’s post-Llama 4 reset: talent density, consumer distribution, and patient execution combined in a small, flat lab. Meta’s 49% investment in Scale let Scale continue while Wang and several colleagues joined Meta; Wang accepted after nine years because the struggling AI program offered a genuine “refounding” opportunity. He now treats some AI research as diamond mining—many failed experiments, but rare outcomes worth a million times more—while more than 200 direct reports, pod-level technical leads, and an anti-bureaucratic structure give researchers “room to cook.”
Deep dive
1. Muse began as a thesis about what humans should do with superintelligence
Wang traces Muse to the June 2025 “Personal Superintelligence” memo developed with Zuckerberg, Nat Friedman, and others at Meta Superintelligence Labs. The mission was broader than an assistant or chatbot: build powerful AI that “actually makes all of our lives better” and “lifts the human experience broadly speaking.”
The underlying division was between things people want to do and things they have to do. Wang expects advanced AI to absorb much of the latter, leaving humans more time for relationships, interests, hobbies, and ambitions—the terrain Meta’s existing products already occupy.
That philosophical fit also creates a business advantage. Meta has 3.5 billion people using its apps daily, a heritage in connecting friends and family, and consumer distribution at enormous scale; Wang therefore saw personal agents and consumer AI as the company’s natural arena.
2. OpenClaw supplied the terrifying, euphoric proof of concept
Early in 2026, Opus 4.5 pushed agents to the foreground while OpenClaw offered a rough glimpse of personal agency. Friedman became its earliest champion inside MSL, having an experience Wang describes as “somewhere between terrifying and euphoric.”
Friedman trusted the agent deeply enough to let it monitor security footage after he said he wanted to drink more water; it checked whether he complied and congratulated him. For Wang, Friedman’s conviction carried weight because he had been early to Stripe, open-source developer tools, and GitHub Copilot: “I think there really is something here.”
Wang’s initiation was a Friedman-authored psychological-probing prompt. With access to his email, photos, and other personal information, the agent ran successive rounds that felt like “3 years of therapy” compressed into one moment. Wang says he was “very vulnerable through that process.”
Both men wrote February memos after immersing themselves in agents during every spare moment. Wang framed the super-agent as, in many ways, “the final consumer product”; Friedman’s memorable titles were “The Claw Is the Law” and “Trust Is a Must,” which helped define the product’s core principles.
3. A two-week prototype became a seven-month reliability campaign
Meta assembled the first prototype in roughly one or two weeks and showed it to the board. Several durable elements were already present: Jollybot, an early Muse Charm, and many of the eventual interaction ideas, even though the full set of concepts and tabs had not yet arrived.
Zuckerberg was experiencing the same euphoria and terror. He used an agent around his home, with his children, and during MMA training, asking it to watch footage and suggest improvements; staff meetings began featuring stories about the strangest or most useful agent experiences.
The hard work was not demonstrating magic but making it dependable. Wang says OpenClaw’s phenomenon peaked and faded because users encountered flashes of brilliance amid frequent breakage; Meta spent seven months moving from a prototype that “sometimes worked and was magical but most of the time didn’t work” to something suitable for billions.
4. Meta chose personal agents while competitors crowded into coding
Wang recalls that Opus and Claude were “way ahead in coding” at the start of the year, creating what looked like an Anthropic takeover. Industry consensus increasingly treated coding agents as the only meaningful AI form factor, capability, or competitive battleground.
Meta resisted that gravitational pull because its own agent experiences had created conviction elsewhere. Rather than chase the same developer use case, it aimed its model roadmap at personal superintelligence and the specific behaviors a personal agent would require.
Senra compresses the distinction sharply: “It’s not work. It’s not coding. It’s personal.” Wang agrees; Meta’s consumer heritage and distribution gave it a plausible right to win that would have been weaker in enterprise software or developer tools.
5. Musepark 1.3 crossed a launch bar defined behavior by behavior
The model team maintained spreadsheets covering more than 100 specific capabilities and behaviors. It built evaluations for each, reviewed deficiencies repeatedly, and trained toward them one by one—a “very meticulous process” of sanding down the model’s rough edges.
Wang connects this to Meta’s older product discipline. The company once optimized conversion, referrals, and every step of onboarding; with Muse, that instinct shifted toward the underlying model—task completion, communication, reliability, and knowing when to return to the user for guidance.
Every spreadsheet row had a launch-blocking threshold and a higher definition of “good.” Checkpoints that were green on 80 rows but red on 20 remained unacceptable; the first model checkpoint to clear the whole sheet was a specially trained Musepark 1.3.
Numerical gates were paired with experiential judgment. The team ran repeated user tests, model-version A/B tests, and first-time onboarding sessions because consumers had to understand that Muse was “not a chatbot”—it could act, but that unfamiliar capability needed careful explanation.
6. A small team protected Muse from large-company product entropy
Fewer than 200 people worked across Muse’s model and product layers. Wang concedes that sounds substantial, but within a company Meta’s size it reflected organizational restraint around what executives expected to be one of the year’s most important launches.
The governing idea was that Muse constituted a “single work of art.” Wang feared the large-organization failure mode: a “Frankenstein,” “Cronenberg amalgamation,” or “smoothie” in which each product manager jams a separate agenda into one incoherent surface.
Friedman’s taste supplied much of the unifying point of view. Senra likens that authorship to the description of Apple as “Steve Jobs with 10,000 lives”—a product transmitting one person’s sensibility rather than exposing the org chart behind it.
Zuckerberg’s role was patience and restraint. “Seven months in AI time” felt like a decade while Wall Street questioned whether Meta could pull anything off in AI, yet he continued publicly alluding to personal agents and trusted the team to wait for quality.
7. Reliability determined whether magic compounded or churned
Personal agents sat “on a knife’s edge”: reliable magic could create one of the greatest consumer products ever, while nondeterministic failures would make users conclude the entire category was trash. Wang says OpenClaw’s exciting moments were not enough to sustain the phenomenon once users encountered its limitations.
Muse’s product experience was highly sensitive to the underlying model. It needed not just to complete a task, but to communicate about it appropriately, recognize uncertainty, request input when necessary, and do all of that consistently enough for users to risk a larger assignment next time.
This explains the unglamorous difference Wang draws between invention and commercialization. He credits Peter Steinberger and OpenClaw’s contributors as visionaries; Meta’s contribution was “really grinding out and sanding out all of the details” across model and product until the combined experience held together.
8. The launch broke through by making Meta culturally surprising again
Wang would disclose only that Muse had “millions” of users. When Senra notes that could mean hundreds of millions, Wang still declines specificity, pointing instead to graphs he retweeted that would imply Muse was the fastest-growing consumer AI app of all time.
Meta’s distribution could ensure awareness, but Wang believed awareness alone could not cut through the “torrential stream” of models, products, startups, and features. The team explicitly discussed how Muse could become culturally interesting—and how the campaign might make Meta “cool again.”
Wang says he personally wrote every post in the supposedly unhinged feed. Memes interrupted meetings and dinners like bugs demanding to be solved; after casual posts unexpectedly performed well, he remembered that “the internet rewards risk” and surprise, then leaned further into that register.
Friedman joked that the team expected to learn about Muse users but was instead “learning a lot about Alex’s mind.” Wang believes the campaign genuinely mattered; Senra’s own inbound evidence was that people discussing the memes had also downloaded and begun using the app.
9. Product marketing worked when the user benefit fit in a screenshot
The recurring question was why someone should care. Wang’s answer is that most people do not care about Meta, OpenAI, Anthropic, or research for its own sake; they care about what a product can do for them.
The product therefore promoted real use cases rather than institutional branding. Wang initially retweeted dozens of interesting interactions, while Jolly—the small mascot—became a flexible cultural object that could carry a briefcase, inhabit memes, and communicate a lighter personality.
Wang says Meta ran a few ads, but he was struck by how little benefit-led content appeared in Anthropic’s and OpenAI’s campaigns. Modern consumer products, he argues, must “work on the screenshot.” Jolly’s presence made shared conversations immediately identifiable as Muse, turning screenshots in group chats and social feeds into recognizable distribution alongside the few ads they ran.
The strongest example came from Senra’s friend, who left his ID on a Flexjet and told the agent he wanted no involvement in recovering it. The agent worked through the courier, office security, and retrieval; three or four screenshots conveyed the product more effectively than an abstract campaign.
10. Retention follows an escalating trust fall
Wang describes adoption as a sequence: give the agent a small task, watch it succeed, then offer something slightly larger. Each completed step expands the user’s belief about what is possible until the relationship becomes “this trust fall” with an AI agent.
Senra notes that college students sometimes make a full trust fall immediately: they record voice notes, brain-dump into AI, and trust it to organize their thoughts into plans and actions. That shows where broader consumer behavior might head once reliability is assumed.
The practical objective is therefore not generic engagement but an early wow moment followed by uninterrupted successful escalation. If the agent works “every one of those times,” curiosity becomes trust, and trust expands both task size and product stickiness.
11. Wang frames personal agents as a second mind that restores human agency
Wang mainly uses the product for recurring workflows and as a “second brain,” since his job still requires presence in meetings. His larger framing asks what happens “if everyone had a second mind beyond their own”—an AI that supports a person’s wants, dreams, and unfinished intentions.
Senra suggests that the product could give people “the adulthood they dreamed of in childhood,” describing children who imagine becoming astronauts, saving the planet, or changing the world before work and obligations erode those ambitions. Wang responds that adults are “zombies,” having lost agency and ambition because “the life has been beaten out of them.”
Wang’s promise is an “escalator of agency”: accomplish one ambition, dream bigger, and repeat. He sees iconic entrepreneurs such as Elon and Mark living this cycle already; personal AI could make some version available to everyone rather than only unusually fanatical founders.
Senra supplies the widest speculative framing: history has been shaped by the few fanatics who made their wants real, but humanity has never seen “every single person’s agency fully switched on.” He imagines billions gaining abundant intelligence and producing a strange, diverse, artistic world—his analogy is “Rick and Morty Interdimensional Cable” manifested in reality. Wang responds with a line from the show: “The universe eats smart people.”
12. Muse emerged from Zuckerberg’s rapid refounding of Meta’s AI lab
Wang’s earlier relationship with Zuckerberg consisted of occasional conversations after an introduction around 2021 or 2022; he thinks their first meeting may have taken a year to schedule. After Llama 4 disappointed Meta internally and AI’s strategic importance intensified, Zuckerberg called to ask what the company should do.
The entire path from that call to the announced Scale transaction took five or six weeks. Scale’s enterprise and government business did not resemble Meta’s iconic product acquisitions, so Wang initially doubted the industrial logic; the real objective became clearer when the discussion turned to talent.
The resulting structure left Meta owning 49% of Scale while Scale continued independently and Wang plus several colleagues joined Meta. Wang calls it a “win-win-win”: shareholders benefited, Scale retained a future, and Meta gained people who could help rebuild its AI effort.
Leaving remained emotionally difficult after nine years—roughly a third of Wang’s life—and he had genuinely treated Scale as his life’s work. What changed the decision was the chance to refound a program that looked like “damaged goods,” establish new principles, and build much of the lab from scratch.
13. The new lab is organized for diamond mining, not bureaucracy
Wang’s founding organizational thesis was a small, flat, highly technical team with extreme talent density. Meta could offer a startup-like chance to shape the institution without startup bottlenecks in compute, infrastructure, or distribution, all oriented around the clear north star of personal superintelligence.
Recruiting was not simply about offering an obviously superior cash package. Wang says the stock-price appreciation at OpenAI and Anthropic meant that staying could already be very lucrative, making Meta’s offer net neutral in compensation for many people. The attraction was the chance to shape a new lab and do meaningful work with Meta’s resources.
His own role changed from author to coach. At Scale, believing a founder should do every job gave the company his point of view but made him a bottleneck; at Meta, he instead recruits exceptional people, identifies unusual talent, sets direction, and creates conditions for “the greatest expression of their talent.”
The operating analogy is diamond mining versus skyscraper construction. Senra contrasts frontier research with the operational optimization of data businesses; Wang says that is the difference between what he felt he was doing at Scale and what he is doing at Meta. In some AI areas, most ideas fail but one result “can be literally a million times more valuable than everything else.”
Wang illustrates the power-law idea with his 19th birthday: 10 or 12 Harvard and MIT students, including Jeff Yan of Hyperliquid, Scott Wu of Cognition, Jesse Zhang of Decagon, and Vicky Ye of Anthropic. He jokes that HRT, where he was interning, should simply invest in its interns.
More than 200 people effectively report to Wang, with pod-level technical leads resolving technical choices. He admits he is not the ideal conventional manager for that many people; direct reporting is an anti-bureaucratic statement that proven researchers “don’t need managers, they need a great environment” and “room to cook.”