Meta launches Muse, a consumer AI assistant, after losing billions on the metaverse. According to the presenter, success lies in small teams, polish and clear agent status.
Why this is useful to know
Mark Zuckerberg renamed his company around the metaverse in 2021. That led to heavy losses: Reality Labs lost 19.2 billion dollars in 2025 alone, according to the video. Now Meta launches Muse, an AI assistant that ordinary people use on their phone, without having to step into a new world. According to the presenter, ten million people started using it within six weeks. That figure comes from the video and has not been independently verified. Why does this matter to you? Because the lesson is not about money. It is about how you take a new bet after a failure: different people on board, small teams and attention to the details nobody asks for.
When you need this
You need this when an AI assistant or project disappoints after the first trial. The source advises against simply investing more in the same approach. First change something concrete: who works on it, which customer you have in mind or how you work. Keep the team small enough to decide quickly. Also check whether features hang together. Three separate features that together cost a lot of effort are one problem for the user. Always show what an agent is doing right now, for example with a clear status message. Polish the dull parts too, not just the opening demo. Pay particular attention to the moments where the user is bothered every day. After a failed project, ask: what would really make my next attempt better?
How it works under the hood
Technically this is mainly about orchestration and behaviour around the model, not about the model itself. According to the presenter, who refers to an interview with Meta, the team tested and improved more than 100 model behaviours before launch. That looks like an evaluation set per use case: when does the agent ask for confirmation, when does it stop, when is it done. For a setup with models, tools and memory, this means that the harness, the status of tasks and the tool output must be visible to the user. An agent that does not indicate it is waiting for input creates supervision overhead. Limitation: the video gives no technical details about architecture, retrieval or guardrails, so you have to test those assumptions yourself. The figure of fewer than 200 people on the Muse team also comes from the video and has not been independently verified.
This news touches chapter 7 and 16 and 23 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to a publication by Nate B Jones (YouTube): https://www.youtube.com/watch?v=JFoDdKuRmSk
