Meta’s launch of the personal AI agent Muse in September pushed the company beyond the standard chatbot model and into software that can operate apps, arrange parts of a user’s schedule, and handle day-to-day tasks. The product quickly climbed the U.S. App Store rankings and became one of the first major outputs tied to Meta Superintelligence Labs, or MSL.
In a recent interview with Founders host David Senra, Meta Chief AI Officer Alexandr Wang gave his first extended account of how Muse moved from idea to prototype to release over seven months. He also explained why he chose to leave Scale AI, the company he founded and led for nine years, to join Meta at a moment when some outside observers had viewed it as trailing in the AI race.
Wang described a series of experiences that shaped the team’s thinking. Meta board member Nat Friedman had OpenClaw monitor whether he was actually drinking enough water. Wang, in turn, let an agent analyze his emails, photos, and other personal information in a deep psychological exercise that he said felt like compressing three years of therapy into a single interaction. Those moments helped persuade the team that a personal AI agent could become one of the most important consumer products in the category, and possibly what Wang called the “ultimate consumer product.”
How Meta arrived at its personal superintelligence thesis
According to Wang, MSL had already settled on “Personal Superintelligence” as a long-term goal when the lab was established in 2025. He tied that idea back to a public memo from Mark Zuckerberg that centered on a basic question: how to make powerful AI improve ordinary people’s lives, rather than simply making work more efficient.
The sharper turning point came in early 2026. By then, Anthropic’s Claude Opus 4.5 had shown strong capability in software development agents, and OpenClaw was beginning to draw serious attention in tech circles. Wang said Friedman was one of the earliest people inside MSL to experiment with OpenClaw, and that usage left a lasting impression on him.
Friedman had asked OpenClaw to help him build a habit of drinking more water. The agent then used camera footage to check whether he had actually done it, and even praised him after he met the goal. Wang said Friedman’s reaction fell somewhere between delight and fear. Once Friedman started trusting the system, the team began to see just how deeply a personal agent could reach into daily life compared with an ordinary chatbot.
Friedman later suggested that Wang try a set of prompts for deeper psychological analysis, leading OpenClaw through multiple rounds of introspection. Wang said the agent had access to his emails, photos, and other personal information, combined those sources, and then surfaced a string of probing psychological questions. He described the session as intense and exposing, closer to a compressed multi-year therapy process than a conventional AI interaction.
Those experiments led Wang and Friedman to a shared conclusion: once AI stops answering one-off prompts and starts understanding a user over time, holding context, and taking action, a different kind of consumer product becomes possible. By February 2026, both men had written internal memos to Meta’s board. Wang argued that a personal super agent could become the “ultimate consumer product,” while Friedman submitted memos titled The Claw s the Law and Trust Is a Must, stressing that the value of a personal agent is inseparable from user trust.
A prototype in one to two weeks, then seven months of refinement
Once Meta committed to the direction, the team built the first Muse prototype in roughly one to two weeks and showed it to the board. Wang said that early version already contained parts of the final product’s core identity, including Jolly, which later became Muse’s signature character, along with the beginnings of the interaction design.
Zuckerberg was deeply involved in testing. Wang said Meta’s CEO used the agent to help with household tasks and even had the system watch videos of his mixed martial arts training so it could analyze his movements and offer suggestions for improvement.
Meta still chose not to ship quickly. Wang said the early prototype could occasionally pull off striking tasks, but most of the time it was not stable enough. He pointed to that gap as one reason the early OpenClaw wave did not last. Wang said he has strong respect for OpenClaw founder Peter Steinberger and views the technical vision as a genuine breakthrough. Still, he drew a clear distinction between building an agent that billions of mainstream consumers can use with confidence and shipping an open-source tool that developers can tune on their own.
That meant Muse went into a seven-month polishing phase. The team created a spreadsheet listing more than 100 concrete capabilities and behavioral requirements, then evaluated whether the model could execute tasks, understand user intent, ask for confirmation at the right time, and complete operations reliably. Each requirement had its own evaluation method and release threshold. If the model performed well on 80 items but failed the other 20, Wang said, the product would not ship.
Only after the specially trained Muse Spark 1.3 model cleared all of the key evaluations did the team decide it was ready for release. Wang said the entire Muse effort, covering both model training and product development, involved fewer than 200 people. By Meta standards, that was a relatively lean group.
He argued that a product like this cannot become a patchwork of features defended by different product managers. It needs one coherent vision. Wang added that the final Muse experience reflects Nat Friedman’s taste and judgment to a large degree.
Why Meta did not go all-in on coding agents
Wang said Muse also reveals a strategic difference between Meta and companies such as OpenAI and Anthropic. In early 2026, Anthropic had built a clear lead in coding-focused models. The Claude Opus line and rapid gains in coding agents pushed much of the industry toward the view that software development was the most important AI application, perhaps even the only one worth fully chasing.
He acknowledged that Anthropic’s lead put real pressure on competitors, including Meta. Even so, Meta chose not to direct all of its resources toward coding agents. Instead, it turned its main product focus toward the personal agent category.
Wang linked that choice to Meta’s core business. He said human life can be split into two broad classes of activity: the work people have to do and the things they actually want to do. The promise of AGI, in his telling, is that AI should absorb more of the first category so that people can spend more time on the second.
Instagram, Facebook, and WhatsApp already sit around relationships, personal interests, and lived experience. With an app ecosystem used daily by about 3.5 billion people, Wang said Meta has a distribution advantage in personal AI that would be difficult for other labs to reproduce.
That is why, in his view, the real market where Meta can build an edge is not enterprise workflow software and not software development, but personal consumer life. Muse was designed from the start not as a chatbot with more features, but as a system that can take over real tasks from a user’s everyday life.
Meta’s own description says Muse can use a dedicated Muse Secure VM to operate a browser, use applications, and handle tasks such as sending emails, arranging travel, filling out forms, and helping a user plan longer-term personal goals.
Trust, not chat frequency, sits at the center of the product
Wang said the biggest difference between an agent product and a traditional app is that users have to build trust in stages. On the first try, most people will hand over only a very small task. If the agent completes it successfully, they may try something slightly more complicated. If the second and third attempts also work, they start handing over more important responsibilities.
He described that dynamic as an ongoing “trust fall.” The real retention loop for Muse, he said, is not getting people to open the app every day for conversation. It is letting them keep discovering that there are many parts of life they no longer need to manage personally.
Senra shared an example from a friend who left identification documents on a private plane and then asked Muse to determine where the items were, arrange a courier pickup, and get the documents sent to his office. In Wang’s telling, the user’s core demand was to stay out of the coordination process altogether, and Muse ultimately handled a multi-step pickup and delivery chain.
That, Wang argued, is the basic distinction between a personal agent and a chatbot. The user is no longer asking how to do something. The user is handing the responsibility for getting it done to the system.
That kind of autonomy also brings privacy and trust concerns. Reuters reported in September that Meta had used human contractors to help complete parts of phone calls while testing Muse’s phone concierge feature, triggering internal concerns around disclosure and privacy protection. Meta later paused the related test feature.
Meme-driven marketing and the push to make Meta feel culturally relevant again
Wang also spoke about his much different public style on X after Muse launched. Over the past few weeks, he has posted memes frequently, including material with adult humor and absurdist touches, a sharp break from the more restrained social media posture usually associated with senior executives at large tech companies.
He said the reasoning was straightforward: he wants Meta to feel interesting again and regain cultural relevance. In an AI market where new models, new products, and new features appear every day, Wang said a large existing user base does not automatically make a launch the center of online conversation.
He began by reposting real Muse user examples that showed the agent completing different tasks. He then realized that Jolly, the Muse mascot, was well suited to meme formats. That gave him a way to catch attention quickly and create a visual cue that users could tie back to the product the moment they saw a screenshot.
Wang said he was not heavily relying on Muse agents to generate those memes automatically. Many of the posts, he said, came to him while in meetings or at meals and were published by him directly. He also found that some of the less overthought posts worked better than carefully planned promotional material.
That, to him, was another reminder that internet culture often rewards surprise, risk, and oddness. He said AI companies often make a marketing mistake by repeatedly promoting their brand, model names, and research results without clearly showing what the product can actually do for a user. A screenshot showing Muse successfully helping someone through a complex task may be more persuasive than a polished brand ad.
At the same time, Wang said it is hard to measure precisely how much growth his personal posts generate. He noted that Meta has publicly said Muse has millions of users, but did not offer a more exact figure in the interview and did not directly confirm Senra’s assumptions about the pace of that growth.
From Scale AI to Meta after Llama 4 disappointed internally
Later in the interview, Wang retraced the path that led him to work more closely with Zuckerberg. He said the two first met around 2021 or 2022 through Meta executive Alex Schultz. Scale AI was already an important Meta supplier at the time, and Wang occasionally sought Zuckerberg’s advice on entrepreneurship and company management.
The real inflection point came after the release of Llama 4. Wang said the project’s development did not meet Meta’s internal expectations and was viewed inside the company as a disappointing outcome. At the same time, AI was becoming more important to Meta’s future business, which led Zuckerberg to seek outside views more actively.
Wang recalled that Zuckerberg initially called him to ask what he thought of Meta’s AI strategy and what the company should focus on next. Wang said his core advice was that Meta should use its existing consumer ecosystem and point its AI efforts toward personal life, instead of simply chasing enterprise markets and coding tools. From there, the two sides gradually moved into discussions about a possible partnership.
At first, Wang said, it was hard for him to imagine how a Meta-Scale AI deal would fit together. Scale AI’s business was centered on enterprises, governments, and data services, which did not match the logic behind Meta’s past acquisitions of consumer products such as Instagram and WhatsApp. As the talks developed, though, he came to believe Zuckerberg was less interested in Scale AI’s commercial business than in rebuilding Meta’s AI research organization by bringing in talent.
The outcome was a deal in which Meta acquired about a 49% stake in Scale AI, while Wang and part of his team joined Meta and Scale AI continued operating independently. Wang said the whole process, from Zuckerberg’s first outreach to the public announcement, took only about five to six weeks. For someone who had spent nine years building Scale AI, he said, it was not an easy decision.
Still, Wang argued that Meta’s position at the time created a rare opening: the chance to rebuild an AI research lab and rethink talent, organizational culture, and technical direction at the same time.
More than 200 researchers reporting directly to Wang
Wang said his management style at Meta is very different from the one he used at Scale AI. At Scale, he believed a founder should be able to understand and participate in every part of the company, which led him to hold strong personal views across product, operations, and day-to-day details. That gave Scale AI a founder-shaped culture, but he also came to see himself as a bottleneck as the company grew.
At Meta, he concluded that frontier AI research cannot be run the same way. He is not a specialist in every technical field, and he cannot personally decide every research direction. His focus has shifted instead to recruiting top talent, setting a shared vision, and building an environment where researchers can do their best work.
He said one internal MSL research organization known as TBD has more than 200 researchers reporting directly to him. The structure is unusually flat, though not leaderless. Different research pods still exist, and technical leads decide the specific direction inside each pod. On the people-management side, however, the goal is to minimize conventional hierarchy and bureaucratic process. Wang described that as an anti-bureaucratic design.
He acknowledged that he cannot be the best direct manager for 200-plus people individually. But he also said those researchers are already capable of working independently at a very high level, which changes what management should do. In his view, leaders should supply resources and the right environment rather than monitor every detail.
To explain the distinction, Wang used the contrast between “digging for diamonds” and “building skyscrapers.” Scale AI’s historical data-labeling business looked more like skyscraper construction, where value comes from process optimization, efficiency, and scaled operations. Frontier AI research, by contrast, is like digging for diamonds: most experiments may not produce much, but a small number of breakthroughs can create value far beyond their cost.
That is why, he said, AI research organizations cannot be managed purely through traditional operating-efficiency metrics. Top researchers need enough room to pursue high-risk ideas that may or may not work, because some of those ideas can produce outsized breakthroughs.