Building Muse, Working with Mark Zuckerberg & The Future of AI | Alexandr Wang, Meta

Summary

Alexandr Wang describes Muse as Meta’s effort to build a personal AI agent, guided by the broader vision of “Personal Super Intelligence”: AI that helps people pursue what they want to do, rather than only automate work tasks.

Muse’s development and technology

• Wang says early experiments with OpenClaw convinced him and Nat Friedman that personal agents could be powerful. Their experiences ranged from useful to unsettling: Wang describes an AI session that used information from his email and photos to conduct intensive psychological questioning.
• The team quickly built a prototype, then spent roughly seven months turning it into a reliable consumer product. Wang contrasts Muse with early agent tools that could produce impressive “magic moments” but were often unreliable.
• The development process was highly methodical. The team identified more than 100 specific behaviors and capabilities an ideal personal agent needed, built evaluations for them, and repeatedly trained and tested the model against those requirements.
• Model quality was central to the experience: the agent needed to carry out tasks reliably, communicate clearly, and ask users for guidance when necessary. The team combined numerical launch thresholds with hands-on testing. Wang says the first model checkpoint to meet all the team’s criteria was a specially trained version of Musepark 1.3.
• The product was tested with new users through ongoing model comparisons and A/B tests. Wang emphasizes that introducing people to an agent—not just a chatbot—requires careful onboarding.
• The initial prototype included elements such as the “Jollybot” mascot and an early version of Muse Charm. Wang says the finished experience depended on extensive refinement, not just the initial concept.

Product strategy and user experience

• Meta chose to focus on personal and consumer AI, rather than pursue the coding-agent market, where Wang says competitors such as Anthropic were ahead. He argues that Meta’s social apps, consumer-product experience, and reach to billions of people make it particularly suited to personal AI.
• Wang describes adoption as a gradual “trust fall”: users start with small requests, see that the agent succeeds, and then entrust it with larger tasks. Successful users may come to use Muse as a second brain or set up workflows.
• The interview gives an example of an agent handling a multi-step personal logistics task: a user asked Muse to recover an ID left on a private jet and coordinate its delivery, with little further involvement from the user. The example illustrates how end-to-end assistance can make an agent feel different from a chatbot.
• Wang’s longer-term view is that personal agents could help people overcome everyday obstacles and pursue ambitions they might otherwise abandon. He frames this as an increase in people’s agency, rather than AI simply replacing human goals.

Launch, marketing, and product design

• Wang says Muse had millions of users but does not provide a precise figure. He credits user stories, screenshots, and social sharing with helping the product stand out amid a crowded stream of AI launches.
• He also used memes and the product’s mascot in social posts to attract attention. His marketing principle is to show what a product can do for people, rather than focus on the company or its research.
• Wang argues that a coherent product needs a clear point of view. He says a small, focused team helped Muse avoid becoming a collection of disconnected features and priorities.

Meta’s AI organization

• Wang says Mark Zuckerberg recruited him after Meta’s Llama 4 effort fell short of the company’s expectations. Their discussions led to a deal in which Meta took a 49% stake in Scale AI, Scale continued operating, and Wang and some colleagues joined Meta.
• Wang describes his role at Meta as different from founding Scale. Rather than personally directing every detail, he focuses on setting the direction, hiring exceptional people, and creating conditions for them to do their best work.
• He describes the lab as deliberately small and flat, with high technical talent density and limited bureaucracy. He characterizes AI research as “diamond mining”: many experiments may fail, but a few breakthroughs can justify the investment.

Reviews, guides, and tutorials

No formal product review or step-by-step tutorial is presented. The interview discusses Muse’s user-testing process, model evaluations, launch criteria, and examples of how people use the agent.

Main speakers and sources

• Alexandr Wang — Meta AI leader and Scale AI co-founder; describes Muse, its development, and Meta’s AI strategy.
• David Senra — Host and interviewer.
• Other people and projects discussed: Mark Zuckerberg, Nat Friedman, OpenClaw and its creator Peter Steinberger, and competing AI products and labs.