Quick answer
The Meta Muse AI assistant reportedly reached 10 million set-ups in six weeks, and Nate B Jones argues three founder habits made it happen: Mark Zuckerberg brought in new people to cover what Meta lacked, the team obsessed over small details nobody asked for, and he tried again after a very expensive miss. Meta’s Reality Labs division lost a reported $19.2 billion in 2025 alone on the metaverse push. Rather than playing it safe, Meta built a phone-based personal assistant that helps with everyday jobs. You don’t need Meta’s money to copy these habits. AI now makes a second attempt at a shelved idea far cheaper than it used to be.
Key takeaways
- 10 million people reportedly set up Muse in six weeks, a strong start, but small by Meta’s billion-user standards.
- Lesson 1, the team: Alexandr Wang and Nat Friedman reportedly led a group of fewer than 200 people, kept small on purpose so decisions stayed sharp.
- Lesson 2, polish: the team reportedly tuned more than 100 specific behaviours before launch, and made it obvious what the assistant is doing at any moment.
- Lesson 3, swing again: a second attempt only works if you change something real: the people, the approach or how you see the customer.
- Muse wasn’t a sure thing. Nate says it isn’t the smartest agent available and is expensive to run. It wins on being useful and easy to live with.
- For the rest of us: the money and skills barriers to trying again have dropped. The fear of embarrassment hasn’t, and that is often what really stops us.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Polish an AI-Built App Before Launch (the “Bathroom Test” Checklist) — a step-by-step checklist for finding and fixing the rough edges in anything you have built with AI
- How to Show Users What Your AI Agent Is Doing, Without Technical Jargon — simple ways to make it obvious when your agent or automation is working, waiting or finished
In this guide
- Watch: the source video
- What is the Meta Muse AI assistant, and why do 10 million sign-ups matter?
- How did a $19 billion loss lead to Muse?
- Lesson 1: Who did Zuckerberg put in charge, and why keep the team small?
- Lesson 2: Why does polish matter so much in an AI assistant?
- Lesson 3: What does “swing again” actually mean after a failure?
- Was Muse’s success inevitable?
- How can you use these lessons on your own projects?
- Frequently asked questions
- The bottom line
- Related posts
Most people remember Meta’s metaverse years for the jokes. Fewer people noticed what came next. In a new video, Nate B Jones looks past the headlines and asks a more useful question: what did Mark Zuckerberg do as a founder that the rest of us can copy?
This is our plain-English summary of that argument about the Meta Muse AI assistant, with the figures clearly marked as reported, and a practical take on how to apply it if you build products, content or automations.
Watch: the source video
What this video covers: Nate B Jones on why Meta’s Muse personal AI assistant reached 10 million set-ups in six weeks after Meta’s costly metaverse bet, and the three founder habits behind it: hiring for the missing skills, obsessing over polish, and trying again after a big miss.
Video: Nate B Jones (@NateBJones). This summary and the opinions in it are our own. User numbers, losses, team sizes and other figures are as stated in the video and have not been independently confirmed.
What is the Meta Muse AI assistant, and why do 10 million sign-ups matter?
Muse is Meta’s personal AI assistant: an AI character on your phone that you name, talk to and hand everyday jobs to. It is built to help with ordinary tasks, such as sorting out a bill, dealing with an account or answering a question about your day, using the phone you already own.
Nate covered Muse when it launched in September. He named his assistant Marvin, after the Douglas Adams character, and says it helped him cancel $1,285 of yearly subscriptions he had been putting off. That is one person’s experience, not a typical result, but it shows the kind of job Muse is aimed at.
According to the video, 10 million people set up Muse in six weeks. The interesting part is not the number on its own. It is that this time people actually turned up, which the metaverse never managed. We first covered Meta’s AI models in our look at Muse Spark Open.
How did a $19 billion loss lead to Muse?
Zuckerberg has wanted Meta to own a computing platform for years, and the metaverse was his first huge attempt. Meta reaches billions of people, but its apps mostly run on phones and operating systems controlled by Apple and Google. He wrote about that frustration in his 2024 letter on open source AI.
The ambition goes back at least to buying Oculus in 2014. By 2021, Facebook had renamed itself Meta around the metaverse. Most people never followed. Nate says Reality Labs, the division behind headsets, glasses and related research, lost $19.2 billion in 2025 alone, and that is just one year of the bill.
| Year | What happened (as described in the video) |
|---|---|
| 2014 | Facebook buys Oculus and talks about the platform after mobile |
| 2021 | Facebook renames itself Meta, betting its identity on the metaverse |
| 2024 | Zuckerberg’s open source letter spells out his wish to control the next platform |
| 2025 | Reality Labs reportedly loses $19.2 billion in a single year |
| 2026 | Muse reportedly reaches 10 million set-ups in six weeks |
The easy move would have been to retreat to Meta’s hugely profitable advertising business. Instead, Zuckerberg kept chasing a platform, but changed the shape of the bet. With Muse, you don’t have to buy into a grand vision. You just need a job done, and if the assistant helps, you come back.
💡 Tip: This is the first takeaway hiding inside the story. You might see the whole future of your product in your head, but your customer only sees the first thing you ask them to do. If that first step doesn’t make their life better, they never get to see the rest.
Lesson 1: Who did Zuckerberg put in charge, and why keep the team small?
Zuckerberg reportedly asked for outside help, then brought in Alexandr Wang from Scale AI to lead Meta’s AI effort, with Nat Friedman central to the product work. Nate cites an October interview in which Wang described Zuckerberg approaching him after being disappointed with Meta’s AI progress.
According to Wang, as quoted in the video, fewer than 200 people worked across both the model and the product. They had Meta’s resources behind them but were kept small on purpose, so that hard decisions were actually made instead of being smoothed over by committee.
You can feel the difference as a user. When too many teams own a product, one tab behaves differently from another, and you have to learn two sets of rules for similar tasks. Each choice made sense inside the company, but the extra effort lands on you.
Nate also gives credit where it is due: it wasn’t just Zuckerberg. Wang and a couple of hundred people in Menlo Park built it, and he “let them cook”. The lesson for any builder or leader is simple: find people who bring what you are missing, give them important work, and get out of their way.
Lesson 2: Why does polish matter so much in an AI assistant?
Polish matters because you live with an assistant every day, so a small rough edge in a demo becomes a daily irritation. Nate calls Muse the most polished consumer AI app available right now, and he means something specific: places where the team could have stopped at “it works” but kept going.
He compares it to judging a restaurant or hotel by its toilets. Everyone polishes the entrance. Checking the bathroom tells you whether they care about everything. In Muse, he says, the quieter tabs and screens that never featured in the launch video are just as consistent as the first conversation.
His favourite example is how the assistant shows it is working. Instead of a technical status message, his Marvin would pull out a tiny laptop and type, or scratch its head. You glance at it and know what is going on. That answers the question everyone has when they hand work to AI: “Is it actually doing the thing I asked?”
Nate also says Wang described a list of more than 100 specific model behaviours the team tested and improved before launch, such as knowing when to ask you something and when to just get on with it. Meta also reportedly built a model that generates facial expressions and movements synced to speech, ready for a live voice experience.
🔎 Unconfirmed / rumour: Nate mentions a possible small, Tamagotchi-like wearable you could hold up and talk to your Muse through. He says it isn’t available yet, and we haven’t seen Meta confirm details or a date. Treat it as a hint, not a product.
A charming character that fails at the job is still a failure, though. The character, the conversation and the actual work have to fit together so people feel comfortable asking for help. If you build anything with AI, the trap is clear: it is now so easy to make an impressive first version that it is tempting to keep starting new ideas instead of fixing the awkward parts of the one you have.
Built something with AI? Before you share it, run it through our bathroom test checklist for polishing an AI-built app. It takes an afternoon and catches the rough edges people notice first.
Lesson 3: What does “swing again” actually mean after a failure?
Swinging again means making a genuinely different second attempt, not repeating the first one with more enthusiasm. Zuckerberg admitted the first bet hadn’t worked, then changed who was leading, what was being built and how it reached people.
Nate turns that into a question for any disappointing project: what would actually make my next attempt better? The missing piece might be:
- A deeper understanding of the customer, from more time with the people you are building for.
- Skills you don’t have, such as design judgement or technical depth, from a collaborator.
- Focus, by stopping splitting your time across too many priorities.
- A different approach, like a smaller first step that delivers value straight away.
⚠️ Warning: Coming back with the same assumptions, the same way of working and more energy rarely fixes anything. If nothing meaningful changes, expect the same result.
Was Muse’s success inevitable?
No. Nate is clear that Muse was a bet, not destiny. Replay the facts: a leader fresh from a very public platform failure, a new AI chief who had never built a billion-user consumer app, a deliberately small team, and in-house models many people rated poorly. None of that adds up to a guaranteed win.
Meta does have advantages most of us never will: a huge existing business, distribution, computing power and access to talent. But Nate argues that using those as an excuse turns a useful story into a story about fate. Muse still has a long way to go.
| Claim in the video | What kind of evidence | How to read it |
|---|---|---|
| 10 million set-ups in six weeks | Figure stated in the video | Reported, not independently checked |
| Reality Labs lost $19.2bn in 2025 | Figure stated in the video | Reported; one year only, not the total cost |
| Fewer than 200 people built Muse | Wang’s interview, as quoted | Second-hand quote |
| 100+ behaviours tuned before launch | Wang’s interview, as quoted | Second-hand quote |
| Muse is the most polished consumer AI app | Nate’s opinion from his own use | Opinion, not a measurement |
| A wearable Muse device is coming | Hint in the video | Unconfirmed |
Nate also notes that Muse is expensive to run, and that 10 million users is small for Meta. To matter, it needs to get much cheaper to serve, grow towards 100 million and then a billion users, and connect people with businesses and other assistants.
How can you use these lessons on your own projects?
Start by revisiting the ideas you ruled out because they needed money, a team or skills you didn’t have. AI has cut the cost of a first attempt at many things that used to need specialist help. That changes what is sensible to try.
Nate gives an example. Say you have spent years in an industry where customers struggle to compare complicated quotes. You know which details matter. You always pictured hiring developers and funding months of work. Now you can often build a rough version yourself with AI, show it to real people and check it against what you know.
- Dig out your “can’t do” list. Nate calls it a haunted graveyard: ideas parked because of cost, past failures or embarrassment.
- Pick one idea and name what was missing last time. Money, skills, customer insight or focus?
- Change one real thing. Use AI to build a working version, find a collaborator, or talk to five potential customers first.
- Show, don’t pitch. Let people try something they can click, rather than explaining slides. Our guide to validating digital product ideas helps here.
- Polish the part people use most before adding more features, and make sure users can always see what is happening.
- Decide honestly. Nate says most ideas will need to be abandoned. That is normal. The point is that trying again is now cheaper.
💡 Tip: Nate suggests choosing projects that get more exciting when a new AI model comes out, not ones that make your stomach drop. If a better model would help your idea rather than replace it, you are building in the right direction.
One barrier hasn’t fallen: embarrassment. It still feels awful to show someone something rough, or to fail in public. Nate is honest that he has founded businesses that didn’t work. Sometimes “impractical” really means “I don’t want to look silly”. It is worth being honest with yourself about which one it is.
Building an AI agent or automation for other people? Read how to show users what your AI agent is doing. Clear status updates are one of the easiest bits of Muse-style polish to copy.
Frequently asked questions
What is Meta Muse?
Muse is Meta’s personal AI assistant: an AI character on your phone that you name, talk to and ask to help with everyday tasks such as sorting out bills, accounts and questions about your day.
How many people use Muse?
According to Nate B Jones, 10 million people set up Muse in its first six weeks. Meta has much bigger ambitions, so this is an early figure, not the finish line.
How much did Meta lose on the metaverse?
The video says Reality Labs, the division behind Meta’s headsets, glasses and related research, lost $19.2 billion in 2025 alone. That is one year, not the total spent.
Who built Muse?
According to the video, Alexandr Wang, who joined from Scale AI, led Meta’s AI effort with Nat Friedman central to the product work, and a team of fewer than 200 people built the model and the app.
Is Muse the smartest AI assistant?
Nate says no. He argues Muse succeeds on usefulness and polish rather than raw intelligence: it knows when to ask you something and makes it clear what it is doing.
What can small businesses learn from Muse?
Bring in the skills you lack, polish the everyday parts of your product rather than just the headline feature, and if an idea failed, change something real before trying again. AI makes that second attempt cheaper.
How do I polish something I built with AI?
Test the quiet screens as well as the main one, fix inconsistent wording and behaviour, and make every state clear to the user. Our bathroom test checklist walks you through it.
The bottom line
The Meta Muse AI assistant story isn’t really about a billionaire’s budget. It is about three habits anyone can copy: bring in the people who fill your gaps, care about the details nobody asked for, and make your next attempt genuinely different from the last. Your next step: pick one idea from your own “can’t do” list, write down what was missing last time, and spend an hour this week building the smallest version someone could actually try.
Related posts
- Muse Spark Open: Meta’s Open-Weight Model, Opus 6 Rumours and This Week’s AI News
- AI Agents and Productivity Theater: 4 Principles to Make Them Do Real Work
- How to Validate Digital Product Ideas Before You Spend Too Long Building Them

[…] Main article: This guide follows our article on the Meta Muse AI assistant and three lessons from Zuckerberg’s comeback, where polish is one of the main reasons Muse caught […]