Building Muse: Seven Months to Earn Users’ Trust

Original source: David Senra interviews Alexandr Wang about the creation of Muse. Episode page.

Guest: Alexandr Wang, Meta’s chief AI officer and co-founder of Scale AI. Host: David Senra. Published 2026-10-11. Duration: 01:12:17.

This article draws on the complete English automatic captions and checks relevant details against Meta and Scale publications. Internal evaluations, user figures, and personal experiences remain attributed to the speakers. Product ambitions and editorial interpretations are identified separately.

The early Muse prototype took roughly 1 to 2 weeks to build. Turning it into a product for ordinary consumers involved about 7 months of refinement.

Alexandr Wang returns repeatedly to that interval in his conversation with David Senra. The prototype could occasionally accomplish something exciting, but it also failed and broke frequently. An impressive demonstration left a practical question unanswered: could a person comfortably hand it the next task?

Meta announced Muse on 2026-09-08. Wang’s account connects the work behind that launch to the choices facing a personal-agent team: which behaviors to train, how trust develops, who shapes the product, and how the research organization keeps exploring.

A useful way to read the interview is to trace user trust back to individual tasks.

Why Meta centered personal agents

By February 2026, Wang recalls, he and Nat Friedman were spending substantial time with OpenClaw. Their experiments suggested that an agent could connect personal context, information, and action to help someone with recurring parts of life.

They wrote internal memos and demonstrated an early prototype to the board. Mark Zuckerberg was experimenting too, including having an agent review videos of his mixed martial arts training and suggest improvements. These are early experiences as recounted by Wang.

Wang describes coding agents as the main competitive pressure at the time. In his assessment, Claude was progressing quickly, and the industry had a strong incentive to crowd into the same category. Meta’s interest in personal agents rested on another set of advantages.

Instagram, WhatsApp, and Facebook connect people to interests, friends, and family. Wang cites roughly three and a half billion daily users across those apps. His ambition is for AI to absorb more obligatory work so that people can spend time on activities they choose.

That audience belongs to Meta’s apps; it is not a Muse user count. A focus on personal AI also does not establish that Meta abandoned coding models. The official Muse Spark 1.3 announcement discusses both agentic and coding capabilities.

His explanation places the people and situations Meta already understands at the center of product strategy. Model development then has specific needs to serve. Interview: the personal AI strategy.

Breaking reliability into more than 100 behaviors

During development, the team maintained a long spreadsheet of the behaviors and capabilities an ideal personal agent would need.

Wang briefly says thousands, then corrects himself to more than 100. The team built evaluations for the entries and worked through successive training and model iterations. Product requirements became part of the model roadmap. Interview: the behavior checklist.

The concerns he describes can be grouped into completing tasks reliably, communicating clearly during the work, and returning to the user when guidance is needed. That grouping is an editorial synthesis; the complete internal checklist is not public.

Each row had a launch-blocking threshold and a standard for good performance. Wang illustrates the problem with 80 green rows and 20 red ones: the product still would not be ready. Those are counts of checklist entries, not a published task-success rate.

He says the first checkpoint to pass every row was a specially trained version of Muse Spark 1.3. Alongside those evaluations, the team repeatedly observed new users installing and trying the product and compared experiences across model versions.

Readiness depended on both measurable behaviors and the experience of using the product. The interview does not disclose the full test set, individual thresholds, or long-term success rates. It does describe a practical development method: identify failures precisely enough to know what the next model iteration should improve. Interview: the launch criteria.

Trust depends on the next task working too

Wang describes a recurring pattern among enthusiastic Muse users. Someone assigns a small task, sees it work, and tries a slightly larger one. The scope expands as earlier experiences hold up.

That trust can be fragile. A larger commitment can also leave a larger mess for the user if the agent fails. An inference from his account is that execution, communication, and requests for guidance must remain dependable together for delegation to continue. Interview: earning trust gradually.

Senra offers a story about a friend who left identification on a private aircraft and asked Muse to arrange its return to the office. In Senra’s account, the agent coordinated a courier and the subsequent handoff. His friend shared screenshots of the exchange.

The story makes the benefit easy to understand, but it remains a secondhand anecdote. The interview does not present a complete action log or demonstrate that every user can reliably accomplish comparable tasks. Interview: retrieving the identification.

Meta’s product announcement also describes user-controlled service permissions and confirmation before sensitive actions such as sending email or making purchases. Those are the vendor’s stated mechanisms, rather than an independent security assessment. Growing trust should retain boundaries the user can control.

Fewer than 200 people and a coherent product

Wang says fewer than 200 people worked across the model and product effort during Muse’s development.

His emphasis is on focus. A high-profile project can attract contributions from many departments, each bringing its own goals, features, and interfaces. The resulting experience can feel fragmented to the person using it.

Muse needed a consistent view of whom it served, how it interacted with people, and what belonged in the product. Wang credits Nat Friedman’s taste and decisions with shaping much of that experience.

In Wang’s account, Zuckerberg also helped by allowing time for continued improvement. In a fiercely competitive AI market, waiting for reliability to improve is itself a resource decision.

A coherent product direction requires sustained choices about what belongs. Team size is one organizational condition; the interview does not establish an ideal headcount for every AI product. Interview: team size and product coherence.

Before a download, the benefit has to make sense

Even extensive distribution leaves an unanswered question: why would someone pause to understand another new product?

Wang describes a constant stream of AI models, products, and features competing for attention. Announcing another release was unlikely to be enough. He shared large numbers of user examples and personally wrote memes involving the Muse mascot.

He sees particular value in screenshots. A friend shares a conversation showing an irritating task getting handled, and the recipient can immediately understand the benefit. A recognizable mascot helps identify the product in the same image that shows the result.

The identification story illustrates that mechanism. Someone can understand a practical inconvenience being resolved without first learning about models, tools, or execution environments.

Wang nevertheless says they lack tracking that would establish how much growth his posts produced. He believes the activity helped, and Senra adds observations from people around him. Neither supplies a causal measurement.

The user figures need similar care. The interview confirms only a broad claim of millions of users. Wang mentions reposting third-party growth charts, but provides no precise active-user count, measurement window, or retention series. The growth ranking in the video’s title should therefore remain an unverified comparison. Interview: growth claims and promotion.

Moving from Scale changed the management assignment

In the second half of the interview, Senra asks why Wang left Scale.

Wang recalls Zuckerberg asking for advice after Llama 4 fell short of Meta’s expectations. What initially resembled an exchange between founders developed into discussions about working together. He estimates roughly 5 to 6 weeks from the initial outreach to the deal announcement. Interview: the early conversations with Zuckerberg.

Wang describes the arrangement as Meta owning 49% of Scale, Scale continuing to operate, and him joining Meta with a small number of colleagues. Scale’s official announcement confirms a minority investment, Wang’s move, and Scale’s continued independence. The precise 49% figure here comes from the interview.

The decision followed about 9 years building Scale. Wang says he had regarded it as his life’s work, making departure from day-to-day leadership emotionally difficult. The opportunity to rebuild Meta’s AI organization, including its culture and working methods, helped draw him toward the new role. Interview: the deal and leaving Scale.

At Scale, he had tried to understand and participate in almost every job, extending his judgment throughout the company. That approach gave the organization a distinct character while also making him a bottleneck.

At Meta Superintelligence Labs, or MSL, research, design, and product work required expertise he could not personally substitute for. He describes the new assignment as recruiting exceptional people, setting direction, and providing the resources and conditions for them to perform.

Research needs exploration and technical decisions need owners

Wang uses two kinds of work to explain the difference: building skyscrapers and finding diamonds.

The first offers relatively continuous returns. Improving a process improves its cost, speed, or quality. He puts data annotation businesses and large-scale delivery in this category, where the challenge is to keep optimizing work whose basic method is understood.

The second tolerates many unsuccessful attempts because a few discoveries can be exceptionally valuable. He places frontier AI research closer to that pattern. Researchers need time and resources to test ideas whose usefulness is still unknown. Interview: different distributions of returns.

That reasoning also appears in his organizational choices. Wang says he has more than 200 direct reports and describes a very flat people-management structure for the TBD research group.

This figure has a different scope from the earlier team of fewer than 200. One concerns his reporting relationships in the research organization; the other concerns the model and product effort behind Muse during the development period. They should not be treated as interchangeable headcounts.

Technical leadership still exists. Researchers work in pods, and technical leads set direction and resolve technical decisions when needed. Group discussion coexists with responsibility for choices.

Wang acknowledges that he cannot be an ideal conventional manager to so many people. He wants already accomplished researchers to have room and support. The model depends on high talent density; the interview does not provide a universal management formula. Interview: direct reports and technical pods.

Taken together, his comments suggest that product coherence and research freedom can coexist. A product needs clear choices, while research needs room to explore questions whose answers are unknown. Management has to recognize which kind of work it is supporting.

The long-term ambition is greater human agency

Wang hopes personal AI can help people retain their ambitions. Obligations, obstacles, and routine work accumulate, pushing desired projects aside. An agent that takes on some of that burden could help people continue turning intentions into action.

This is his vision for the future. The interview presents no long-term study establishing that Muse already gives users more free time, helps them achieve more goals, or durably increases their agency. Interview: personal agents and human ambition.

The practical lessons are closer at hand: identify the capabilities a delegated task requires, study the failures, maintain a coherent product, and preserve room for research to find new answers.

A personal agent earns a place in everyday life by repeatedly completing entrusted work, preserving the user’s control, and giving that person a reason to delegate the next task.