How Did Meta Create the Hit Muse? Developed by Fewer Than 200 People, Launched Only After Passing Over 100 Tests

Beating AI News Flash: Meta Chief AI Officer Alexandr Wang discussed the development process of Muse in a recent interview. This personal AI Agent, capable of handling tasks for users, took seven months from prototype to launch, with a core team of fewer than 200 people responsible for the model and product.

At the beginning of this year, Anthropic was far ahead in the coding Agent space, and many in the industry believed that coding was the most important application direction for advanced AI. But Wang and former GitHub CEO, now Meta AI product lead Nat Friedman, were more optimistic about personal Agents. Wang believes that Meta's apps are used by about 3.5 billion people every day, and building AI products for ordinary people is the company's strength.

Meta proposed the direction of "personal superintelligence" as early as 2025, and OpenClaw gave this vision a more concrete product form. In February of this year, after deeply experiencing OpenClaw, Wang and Friedman each submitted memos to the board. Friedman wrote two, one of which was titled "Trust Is a Must," emphasizing that personal Agents must earn users' trust. The team then produced a Muse prototype in just one to two weeks, and even had the beginnings of the mascot Jolly and the dedicated hardware Muse Charm. However, the product was very unstable at the time, occasionally impressive, but most of the time unable to even complete basic tasks well.

Over the next few months, the team repeatedly adjusted the model. They listed more than 100 required capabilities, designed tests for each one, and set launch thresholds. Some model versions had already passed about 80 items, while about 20 remained unqualified, so training had to continue. It was not until the specially trained Muse Spark 1.3 version met all standards that it passed this evaluation system.

Meta's senior leadership regarded Muse as one of the most important products of the year, yet never significantly expanded the R&D team. Wang explained that products at big companies can easily become a patchwork of many people's ideas, with every product manager wanting to add something, ultimately turning it into a mishmash. From the beginning, Muse aimed to maintain a unified product experience, with Friedman leading the oversight.

The AI research team managed by Wang also adopted a similar streamlined approach. He revealed that he currently has more than 200 direct reports. Researchers on the TBD team basically report directly to him, with no layered people-management roles in between. Each group still has a technical lead responsible for deciding technical direction. Wang admitted that he is not the best manager for these 200-plus people, but these researchers have already proven their abilities and do not need him to oversee everything. Giving them enough resources and freedom is enough.