Creating Muse: The Fastest-Growing AI Product Since ChatGPT | Alexandr Wang, Meta
In a Nutshell
Alexandr Wang founded Meta Super Intelligence Labs to build personal super-intelligence agents, achieving breakthrough with Muse after seven months of obsessive model refinement and product iteration. The core insight is that personal agents succeed or fail on reliability—Meta won by systematically fixing every failure mode across 100+ tracked capabilities rather than relying on the magic moments that doomed OpenClaw. Meta's unique advantage stems from Zuckerberg's long-term patience, a small focused team maintaining one cohesive vision, and Meta's distribution to billions of users who want AI for life rather than work.
These notes were generated by AI and may contain inaccuracies.
When Meta Super Intelligence Labs was founded, Mark Zuckerberg published a memo titled "Personal Super Intelligence" developed in collaboration with Alexandr Wang, Nat Friedman, and others. The memo outlined the goal of building powerful AI that lifts the human experience broadly, rather than simply advancing technical capabilities. This was published in June 2025, before agents had emerged in any meaningful way and when AI was in a significantly earlier state.
Developing frontier models requires an end-to-end process spanning many months, including building the entire stack, producing datasets, pre-training, and post-training the model. At the start of 2026, the team was in active production mode when external developments shifted focus.
Opus 4.5 brought agent potential to the foreground, coinciding with OpenClaw's emergence. Nat Friedman was the first within MSL to work with OpenClaw and described the experience as somewhere between terrifying and euphoric. He trusted the agent entirely, including allowing it to monitor security footage to verify he was drinking water as requested. Friedman shared stories of this experience during a Stripe Sessions interview.
Friedman's background made his endorsement significant: he was one of Stripe's first customers, an early participant in open source development, and helped create GitHub Copilot. This background provided a clear signal that personal agents represented something genuinely transformative.
Both Wang and Friedman spent weeks immersing themselves in OpenClaw, dedicating spare moments to exploring what personal agents could become. Friedman had created a prompt for comprehensive self-psychoanalysis involving multiple waves of deep psychological probing.
Wang's first major interaction with his own OpenClaw was this psychoanalysis exercise, which he described as the equivalent of three years of therapy compressed into a single intense session. The agent had access to his email, photos, and other personal information, enabling it to pull together comprehensive insights. Wang felt very vulnerable during this February process.
By February, both Wang and Friedman reached the conclusion that personal agents represented something genuinely special. They wrote memos sent to the Meta board. Wang's memo argued that the super agent was the final consumer product and a clear endpoint for consumer products. Friedman wrote two memos titled "The Clause" and "Trust is a Must," which defined core principles for how the products would be built.
Within one to two weeks of these memos, the team built a prototype and demoed it to the board. Many decisions that made it into the final Muse product were present in this first prototype, including the Jollybot and an early version of the Muse Charm. Not all tabs and ideas were present, but a substantial portion of the core vision was already there.
Mark Zuckerberg experienced the same euphoria-terror cycle with the prototype. He used it for home tasks, with his children, and during MMA training where it analyzed videos of his sessions and provided improvement tips. Staff meetings became venues for sharing crazy agent experiences, a practice that spread throughout the Meta team during this February period.
The OpenClaw phenomenon peaked and then disappeared. While users experienced magic moments, the product largely didn't work reliably, leading to its decline. This informed Muse's development focus: transforming the unreliable prototype into a highly reliable consumer product that could confidently be presented to billions of people.
Wang emphasized respect for Peter Seinberger and OpenClaw's creators, describing whatever happened in Seinberger's mind as magical. However, Muse's differentiation came from the unglamorous work of grinding out details to make the combined model-plus-product experience exceptional. The team conducted many iterations on the model itself.
The model development roadmap was aimed toward personal super intelligence broadly, with personal agents as a specific focus. This represented a deliberate strategic choice. At the start of 2026, Opus and Claude led in coding capabilities, with Anthropic experiencing an industry takeover moment. The conventional industry wisdom held that coding agents were the only form factor that mattered for advanced AI.
Wang's early conversations with Zuckerberg before the Scale and Meta deal emphasized that Meta was uniquely positioned for AGI or ASI. The promise of AGI is that AI will handle things people have to do, freeing them for things they want to do. Meta's products—Instagram, WhatsApp, Facebook—represent the realm of what people want to spend time doing: following interests, hobbies, friends, and people they care about.
From a business perspective, Meta has distribution to three and a half billion daily active users. The natural area for Meta's success is consumer AI rather than work or coding applications.
The team maintained awareness that personal agents had previously failed due to imperfect, unpolished experiences. The model development roadmap prioritized building toward personal agents with long spreadsheets tracking hundreds of specific behaviors and capabilities the model needed. The team built evaluations for all these capabilities, trained the model systematically, and conducted regular reviews to identify and improve weak areas.
Meta's culture emphasizes that once metrics are defined, talented people can optimize and sand out details to make products exceptional. In the pre-AI consumer product era, this meant optimizing conversion, referral rates, and onboarding flows. In the AI era, particularly with Muse, this manifested as sanding out every detail of the model to perfect it for the product.
A Jeremy Stern profile in Colossus discussed Zuckerberg's advantage relative to competitors: his unsexy skill set of building teams and improving products over long periods without losing company control or needing additional fundraising. This approach enabled the seven-month refinement period for Muse.
The patience required was immense. Seven months in AI time equates to roughly a decade in non-AI time. Wang credits Zuckerberg with providing patience, restraint, and trust in the process. During earnings calls from February through September, Zuckerberg alluded to building personal agents without hiding the effort, despite Wall Street narratives questioning Meta's AI investments.
Off-record conversations with competitors, including a former Meta researcher now at another lab, revealed perceptions of Zuckerberg as the "f***ing Terminator"—someone who will not stop. His superpower is never being satisfied with product state and constantly wanting improvement, while maintaining team cohesion as an excessively important priority.
Personal agents rest on a knife's edge: if they reliably deliver magical experiences, they have potential to be among the greatest consumer products ever created. However, models are nondeterministic, agents can break and be unreliable. Unreliable experiences lead users to dismiss the product entirely. Getting Muse to a quality level where it resonated with downloaders required extensive iteration.
The team conducted many rounds of user testing with new people downloading and using Muse from scratch. The product is fundamentally different from typical consumer AI—it is an agent, not a chatbot, requiring thoughtful onboarding to explain this distinction. Product experience proved very sensitive to model quality, specifically how reliably models execute tasks, communicate with users, and check back for guidance when needed.
Launch readiness was determined through both numerical and experiential criteria. The team had established launch-blocking thresholds for each of the 100+ tracked capabilities. Earlier checkpoints showed green across 80 rows but red on 20 rows, which was insufficient. The first model checkpoint that was green across all rows was a specially trained version of Musepark 1.3. Experientially, the team could tell this model was substantially better than previous iterations.
Between model work and product work, under 200 people worked on Muse total. Despite being recognized as potentially the most important product Meta would ship that year, organizational restraint kept the team small and focused. The product was treated as a single work of art requiring strong point of view and focus rather than ballooning into a large organizational effort.
Larger organizations and AI labs often create products that become a Frankenstein amalgamation of different people's points of view, visions, and beliefs. This creates a kind of PM hell where every product manager tries to jam their own thing into the product, resulting in a smoothie-like blend that users can feel as disconnected and unfocused. Muse was designed to be the exact opposite with one clear point of view and one cohesive experience.
Nat Friedman deserves significant credit for Muse's development, as his taste and sensibilities ultimately dictated and shaped what the product became. This single point of view approach mirrors Steve Jobs' impact on Apple, where products reflected his personal taste, perspective, and point of view.
Meta has publicly stated that Muse has millions of users. Alexandr Wang has retweeted graphs from others implying Muse is the fastest-growing consumer AI app of all time. The company has avoided providing specific user numbers beyond the broad "millions" designation.
The increased activity on X since Muse's launch stems from several factors. First, the team recognized the need to break through a torrential stream of AI product updates and releases that bombard people constantly. Large team meetings were held specifically to discuss breakthrough strategies beyond Meta's existing distribution channels.
Alexand Wang met and became friends with meme creators and people obsessed with internet culture, who he affectionately calls his "skitso friends." These individuals spend their days thinking about memes and internet dynamics. The X feed content is written entirely by Wang himself, often coming as moments of inspiration during meetings or even dinner with coworkers.
The goal was to make Meta cool again, as despite having culturally significant software like Instagram and Facebook, the company had lost its cool factor. The unhinged posting style emerged naturally over time, starting with retweeting interesting use cases and evolving into more spontaneous posts that performed well.
The internet rewards risk, surprise, and unexpected content. Some of the most successful tweets required minimal thinking time, while overthought content performed less well. Nat Friedman noted during the meme posting period that instead of learning about Muse users, they were learning about Alexand Wang's mind.
Getting users to an early "wow moment" was crucial for Muse's breakthrough. The product naturally lends itself to this because AI agents can perform surprising tasks, especially for users who have never used an AI agent before. The team focused on harnessing stories of these wow moments to help the product gain traction.
Meta ran some ads but initially relied heavily on retweeting user-generated content showing interesting use cases. Wang criticizes traditional AI company advertising (like Anthropic's outdoor campaigns and ChatGPT ads) for being company-focused rather than benefit-focused.
The core insight is that people don't care about companies or their research - they care about what the product can do for them. This principle comes from the early 1900s book "Scientific Advertising" by Claude Hopkins, which emphasizes that advertising should focus on benefits to the consumer rather than the company or product itself.
The Jolly mascot proved valuable for marketing because putting the character in different situations created surprising moments distinct from typical product marketing. Using racy memes to market a product was an unusual approach that generated intrigue.
Getting users to stay engaged requires moving them through a trust fall process with the AI agent. Users start with small tasks, see success, then gradually give the AI larger and more complex problems. This builds trust incrementally as users experience the AI successfully handling increasingly significant tasks.
College students particularly demonstrate full trust fall usage, recording voice notes and brain dumping to AI agents, trusting them to organize thoughts and provide clear plans.
Modern consumer product marketing is heavily screenshot-driven. Products need to work on screenshots that can spread through group chats, word of mouth, or online. Having the Jolly mascot in screenshots immediately identifies them as Muse-related content, providing powerful word-of-mouth marketing.
Wang primarily uses Muse by setting up workflows and treating it as a "second brain" during meetings. His vision for personal AI agents is that they serve as mechanisms to accomplish human wants, desires, and dreams. As humans accomplish their goals with AI assistance, they will dream bigger, creating an escalator of agency throughout their lives.
This contrasts with the current reality where people's ambition and agency get beaten out of them by corporate work and life commitments. The goal is to maintain the big dreams people have as children and help them accomplish increasingly ambitious goals throughout their lives.
The world has historically been shaped by a small number of fanatics who found ways to make their wants real. The promise of AI is enabling every person's agency to be fully switched on, creating a world where billions of people have the resources and intelligence to make incredible things happen. This could manifest as diverse, artistic, and interesting outcomes across different pockets of the world.
Alex Schultz, Meta's Chief Data Officer and longtime Meta executive, introduced Wang to Mark Zuckerberg around 2021 or 2022 to discuss AI. It took approximately one year from the introduction to scheduling the first meeting.
It took a full year from the initial introduction to scheduling the first meeting with Mark Zuckerberg. Mark has an extremely busy calendar and a large number of priorities to manage. He spends significant time on the most important matters, sometimes to the point of excess. This approach allows him to balance his professional responsibilities with being a good father and partner.
Mark writes very long, rigorous responses to WhatsApp messages with remarkable speed. This skill provides insight into his thinking process and demonstrates his ability to handle high-volume communication efficiently.
Meta was an important customer of Scale. Prior to the recent events, the interaction cadence with Mark was approximately once every six months.
The first conversation occurred after Llama 4, which was not on the trajectory Meta desired as a company. Internally, it was considered a disappointment. This occurred during a time when AI was becoming increasingly critical to Meta's future. Mark sought advice on what Meta should be doing and where to focus efforts.
Mark maintains a group of world-class entrepreneurs for advice, separate from Meta's formal board of directors. He regularly seeks input on his work and personal matters, demonstrating humility despite his accomplishments.
The entire process from initial outreach to deal announcement took five to six weeks. This rapid timeline is characteristic of world-class entrepreneurs, as evidenced by Jonathan Ross's experience with the $20 billion Nvidia deal, which took three weeks from first call to funds received.
One key recommendation was that Meta's heritage of focusing on what people want to do positions the company uniquely for the AI shift. Multiple strategic ideas were shared during the parallel conversations about potential collaboration.
The suggestion to "find a way to work together" was not initially recognized as indicating acquisition interest. The idea of a deal seemed improbable at first.
The logic for an acquisition seemed unclear initially. Scale's business model differs significantly from Meta's previous acquisitions like Instagram and WhatsApp, which were product-logic driven. Scale operates as a deep enterprise and government sales business, making the industrial logic less obvious.
Mark was predominantly interested in getting the Llama program and overall AI program on track. He wanted talent rather than Scale's existing business operations.
Clarity about the deal emerged when discussing specific numbers. The final structure involved Meta owning 49% of Scale while Scale continues operating independently. Alexandr Wang and a few others would join Meta.
The deal structure was considered unusual but represented a genuine win-win-win scenario. Scale shareholders benefited significantly, Scale continues operations, and the best days for Scale are believed to be ahead.
The predominant initial emotion was disbelief about the deal's reality. Multiple conversations occurred with Scale team members and investors. The emotional difficulty of potentially giving up a company worked on for nine years was significant, representing approximately one-third of Alexandr Wang's life.
Paul Graham's advice that viewing a company as one's life's work leads to different operational approaches was relevant. Scale was genuinely considered life's work throughout its development.
The potential of rebuilding Meta's AI efforts was compelling. Despite Meta appearing as "damaged goods" in AI, the opportunity to rebuild from scratch with proper principles and culture was attractive.
The situation created conditions to build something new with the right foundational elements. The conversation evolved toward recognizing that a complete rebuild was necessary rather than minor adjustments.
The strategy emphasizes small, incredibly talented, highly technical teams that operate in a flat structure. This approach enables rapid accomplishment, drawing inspiration from Elon Musk's demonstrated success with this model.
Mark operates with extensive deliberation before making decisions but commits fully once a decision is made. He has internalized that failure is less important than the magnitude of success when winning. This philosophy enables operation without fear and comfort with significant risk-taking.
Mark's view that "in a world that's changing all the time, the biggest risk is not taking any" reflects his deep internalization of this approach. This keeps Meta dynamic rather than becoming slow-moving or bureaucratic.
At Scale, the approach involved having strong opinions about every aspect of operations, making the founder the bottleneck but ensuring the company reflected their vision. This approach is recommended for first-time founders.
At Meta, the assignment required accomplishing significant work in limited time, making it impossible to personally handle every job. The focus shifted to creating an environment for hiring exceptional people at the top of their fields across all disciplines.
The opportunity to build a lab from scratch with a small team around the vision of personal super intelligence was compelling. The resources available at Meta eliminated typical startup bottlenecks around compute, infrastructure, and distribution.
Stock price appreciation at OpenAI and Anthropic meant compensation packages were often net neutral compared to staying at those companies. The decision was driven more by the opportunity to shape something new.
The role involves more traditional management focused on creating environments where exceptional people can achieve their greatest expression of talent. This contrasts with being the sole visionary whose point of view must be expressed throughout the organization.
The role functions more like a coach: setting the north star direction toward personal super intelligence and winning on consumer AI, then identifying special talent and providing environments and resources for maximum expression.
AI requires recognizing that humanity doesn't understand the limits of model capabilities. The process is fundamentally scientific, requiring trial and error, brilliant insights, and space for people to do their best work.
Two types of problems exist: diamond mining (power law payoffs where some outcomes are million times more valuable than others) versus skyscraper building (linear payoffs requiring process optimization). VC exemplifies diamond mining, while data annotation businesses represent skyscraper building.
At age 19 as an MIT freshman, the birthday party included students from Harvard and MIT who knew each other from math, science, and computer science competitions. Attendees included Jeffrey Yan (Hyperliquid), Scott Woo (Cognition), Jesse Zang (Decagon), and Vicky Yei (Anthropic researcher). HRT, where many worked as interns, generated $11 billion in revenue and $8 billion in profit in Q2 2020 with under 1,000 employees.
The current role focuses on diamond mining: nurturing promising ideas and identifying special concepts that justify the investment through exceptional outcomes.
Alexandr Wang explains how to create an environment where brilliant people can do their best work by maintaining extremely high talent density. The approach focuses on making it the best place for talented individuals to accomplish their life's work through a technically-focused organization where technical people have the opportunity to excel.
One of the core design principles is maintaining a very flat organizational structure. Researchers hired into the group called TBD report directly to Wang, serving as a statement of anti-bureaucracy. The philosophy emphasizes hiring brilliant people who have already demonstrated their capability through their previous careers, eliminating the need for micromanagement.
While the organization is anti-bureaucracy from a people management perspective, there is a technical leadership structure in place. The team uses a pod structure with various pods, each having technical leads who make tie-breaking decisions on technical matters. The focus is on setting clear technical directions for groups rather than traditional management hierarchies.
The organization makes many group decisions through debate and discussion, leveraging the talent-dense group of brilliant people. The approach involves hearing all opinions from team members who have each accomplished brilliant things, arriving at consensus on the best path forward. This unconventional team structure prioritizes creating an environment where capable individuals have the room to work rather than requiring traditional management.
Wang acknowledges he may not be the best manager to more than 200 people, but emphasizes that these individuals don't need managers—they need a great environment where they are brilliant and extremely capable.
Rather than serving as a traditional coach or manager, the role becomes that of a steward who creates and maintains an environment where great work can happen. The philosophy centers on providing the space and conditions for talented individuals to excel rather than directing their work through conventional management structures.
Keep David Senra in your library
Save the videos and channels worth coming back to, and find them again in one place.
