Meta's Muse: What a $19B Flop and a 10M-User Comeback Teach Founders
Meta's Muse AI agent hit 10 million users after the metaverse lost $19 billion in 2025. Here's what its launch teaches builders about agency.
What happened with Meta’s Muse, and why does it matter?
Meta’s AI agent, Muse, reportedly reached 10 million users within six weeks of launch. That number arrives after Reality Labs, the division behind Meta’s VR and AR bets, lost $19.2 billion in 2025 alone. Mark Zuckerberg spent years and tens of billions of dollars chasing the metaverse, watched the world treat it as a punchline, and responded not by retreating to safe, profitable advertising but by funding another expensive, ambitious bet. The story matters to builders because it is less about Meta’s balance sheet and more about what happens after a public, expensive failure: who you hire next, how much polish you insist on, and whether you’re willing to swing again instead of just quietly shelving the idea.
TL;DR
- Meta’s Reality Labs division lost $19.2 billion in 2025, a cost tied to years of pursuing the metaverse vision Zuckerberg staked the company’s name on in 2021.
- Despite that loss, Zuckerberg didn’t retreat to Meta’s core ad business. He backed a new AI agent, Muse, which reportedly reached 10 million users in about six weeks.
- Alexander Wang, who joined Meta from Scale AI, and Nat Friedman led the effort with a team Wang has described as fewer than 200 people, deliberately kept small to avoid consensus-driven compromises.
- The product’s polish shows up in unglamorous details: more than a hundred specific model behaviors tuned before launch, and an avatar whose expressions and gestures signal status so users don’t have to guess what the agent is doing.
- Muse isn’t positioned as the smartest model on the market. Its edge is in removing friction, like design choices that make a task feel finished and trustworthy rather than merely functional.
- The underlying lesson for founders is a repeatable pattern: fail publicly, change something real about the next attempt (team, approach, or customer understanding), and ship again instead of defending the original plan or quietly abandoning it.
| Vendor | Tier | Risk | ApprovalNEW |
|---|---|---|---|
| Northwind Supply | Tier 1 | Low | Approved |
| Acme Logistics | Tier 2 | Medium | Approve |
| Helios Data | Tier 1 | High | Pending |
| Granite Freight | Tier 3 | Low | Approved |
| Pinecrest Labs | Tier 2 | Low | Approve |
Why did Zuckerberg keep betting big after the metaverse flop?
Zuckerberg’s frustration predates the metaverse rename. Meta reaches billions of people, but its apps run on devices and operating systems controlled by Apple and Google. In his 2024 letter on open source, he wrote directly about wanting more control over the next computing platform, a goal traceable back to the 2014 Oculus acquisition and the 2021 decision to rename the entire company around the metaverse.
That rename was a public, structural bet: company identity, capital allocation, and leadership time all pointed at a future where people would spend significant time in virtual worlds. It didn’t arrive at the scale Zuckerberg described. Consumers didn’t reorganize their evenings around headsets. The result was years of losses and a durable meme of digital Zuckerberg as the symbol of an overreach.
The common move after a failure like that is to get conservative: shrink the ambition, protect the core business, let the bigger vision quietly die. Zuckerberg didn’t do that. He kept investing in a platform bet, this time channeling it through AI and an agent people could use immediately on the phones they already own, rather than a device they’d need to buy and strap to their face.
How did a small team build Muse inside a company Meta’s size?
According to Alexander Wang’s account from an October interview, Zuckerberg approached him directly after being unhappy with Meta’s AI progress and asked what the company should do differently. Wang moved over from Scale AI to lead the effort, with Nat Friedman central to the product side. Wang has said fewer than 200 people worked across the model and product combined, a deliberately small number for a company with Meta’s headcount and resources.
The smallness was a choice, not a constraint. Large teams tend to resolve hard product disagreements through consensus, which often produces compromises that users feel as inconsistency: a feature that behaves one way in one tab and differently in another, small frictions that accumulate into real effort for the person using the product. Keeping the team small forced sharper, more opinionated decisions about what the product should actually do, with someone accountable for the whole experience rather than a patchwork of locally reasonable choices.
The other half of this is that Zuckerberg reportedly stepped back once he found the right people. He recruited outside talent, gave them real scope, funded the work, and let them build without forcing it through the company’s usual scale of process. For founders, the transferable piece isn’t “hire Alexander Wang.” It’s recognizing when an existing team and process can’t produce the thing you actually need, and being willing to bring in different judgment rather than asking the same group to try harder with the same assumptions.
What does “polish” actually mean in an AI agent?
Polish here isn’t a marketing word. It shows up in specific, testable behaviors. Muse’s team reportedly tested and refined more than a hundred distinct model behaviors before launch. That is a meaningfully different claim than “the model is smart.” A capable model can still make a frustrating assistant if it doesn’t know when to check in with the user, when to act on its own, or how to signal progress.
One concrete example: the agent’s avatar visibly “does” the work it’s doing, through small animated cues, rather than leaving the user to interpret a technical status message. That answers a question every person delegating a task to AI has: did it understand what I asked, and is it actually doing it? Removing that uncertainty is design work, not model work, and it has to happen everywhere in the product, not just in the demo flow.
Meta also reportedly built underlying technology for generating facial expressions and movement synchronized to speech, anticipating a live voice experience and a future physical wearable, ahead of shipping the phone app. That’s investment in an experience layer most companies would defer until after proving the core idea works.
Is Muse actually the smartest AI agent available?
No, and the product’s traction doesn’t depend on that being true. Muse is not positioned as having the most capable underlying model on the market. Its apparent strength is in reducing the gap between what a model can technically do and what a normal person experiences when they try to use it for something real, like canceling a subscription or tracking a bill.
That distinction matters for builders because AI tooling now makes it easy to produce an impressive first pass of almost anything: an interface, a prototype, a pitch demo. The temptation is to keep generating new first passes instead of doing the slower, less visible work of fixing the parts of an existing product that are merely functional. A charming, fast demo that confuses users in daily use isn’t a finished product. The unglamorous second and third passes, the equivalent of cleaning a restaurant’s bathroom instead of just dressing up the entrance, are where products that get used repeatedly are actually made.
What’s next for Muse, and does the lesson scale down to smaller builders?
Ten million users is a strong early result, but it’s small relative to Meta’s usual ambitions, which operate at the scale of billions. Reportedly, Meta still has to bring down the cost of serving Muse at scale, extend it into a broader network across assistants and businesses, and grow toward hundreds of millions and eventually a billion users. Meta has advantages most builders will never have: an existing massive ad business funding the bet, global distribution, and enormous compute capacity.
What doesn’t require that scale of resource is the underlying pattern. Take a real swing, accept that it might not land as planned, and when it doesn’t, change something substantive rather than repeating the same approach with more enthusiasm. Keep the team deciding the hardest product questions small enough to stay sharp. Treat the boring parts of the experience, the error states, the status indicators, the second and third tab, as seriously as the opening demo. None of that requires $19 billion to lose first.
Frequently Asked Questions
What is Meta’s Muse?
Muse is Meta’s consumer AI agent, designed to handle everyday tasks like managing subscriptions and answering questions, accessible through an app with a character-driven interface rather than a plain chat box.
How many users does Muse have?
Muse reportedly reached 10 million users within about six weeks of its launch, according to accounts discussed in coverage of the product’s rollout.
How much did Meta lose on the metaverse?
Meta’s Reality Labs division, responsible for VR and AR hardware and research, lost $19.2 billion in 2025 alone, part of a multi-year pattern of heavy investment in the metaverse vision Zuckerberg announced in 2021.
Who led development of Muse?
Alexander Wang, who joined Meta after leading Scale AI, headed the effort alongside Nat Friedman, with a team Wang has described as fewer than 200 people across model and product work.
Is Muse the most advanced AI model available?
No. Muse is not marketed or positioned as having the most capable underlying model. Its distinguishing feature, based on team accounts, is extensive refinement of everyday usability details rather than raw model capability.
