The metaverse has become Muse's AI widget.

"Sometimes, the whole year has been tough," Mark Zuckerberg said about Meta's setback in the large model competition, "but we persisted, because that's who we are, and Muse is the core of everything we've built."

At the press conference this Wednesday, Zuckerberg centered a series of product launches and keynote speeches around Muse.

With this personal AI assistant launched in early September gaining massive popularity, Meta's stock has risen by 30% so far this month, on track to post its best monthly performance since July 2013.

This year's conference is also drastically different from the themes of the events held several years ago, when Zuckerberg declared that the metaverse was the future and renamed the company to Meta accordingly. At this year's annual Meta Connect conference, the company's previous key focus areas were barely mentioned. "We've decided to temporarily reduce our focus on the metaverse," Zuckerberg stated.

Muse will be available on Meta's Meta VR glasses. In addition, the company's upcoming Muse Charm is about to launch, a dedicated handheld device that allows people to talk to Muse without needing a smartphone or glasses.

Zuckerberg said engineers are putting the final touches on the product, which will hit the market before the holiday season.

At the conference, Zuckerberg also unveiled a new monetization plan for Muse AI.

"We believe Muse can help you make money," Zuckerberg told the room full of developers. "We make this commitment because Muse is free to access for a large number of token users, and we expect to generate revenue by charging a small transaction fee over time."

The newly popular Muse has become the only AI hope in Meta's global ecosystem.

Keychains, Smart Glasses: Muse's Ambition for All-day AI Access

The detour on the metaverse track was not a total waste, as the metaverse has stepped down from the main stage this year and become a supporting peripheral for Muse.

Wearing a T-shirt printed with the words "Building is my love language", Zuckerberg demonstrated the new features of Muse glasses on site. He asked his agent through the glasses to arrange a meeting time with the Chief Technology Officer for him.

Meta's goal is to allow users to use the Muse assistant completely without downloading the app that currently ranks first on Apple's App Store. Integrating Muse into Meta's smart glasses is expected to further boost the popularity of this AI assistant.

Apart from the glasses, Zuckerberg also demonstrated a new handheld device called "Muse Charm", which can be hung on a keychain. He said users can use it to contact their Muse agent. This feature will allow wearers to activate the smart assistant via a "wake word" to achieve completely hands-free operation.

The launch of these products was confirmed long ago.

In an interview on the Sources podcast on September 8, Zuckerberg said "Our idea is to give everyone in the world a very powerful personal agent that can understand your goals and work for you around the clock."

This is the first time that the vision of "personal superintelligence" proposed by Zuckerberg a year ago has appeared in product form.

The host of Sources tried Muse for several days and described it as "OpenClaw for ordinary people".

But Muse's ambition is not to be a smarter assistant. What Meta wants to build is a new personal computing system: AI is responsible for understanding your goals, calling a cloud computer that belongs to you, and handling your life, work and transactions for you.

Traditional chatbots handle one-off problems, while Muse wants to handle long-term projects. Zuckerberg said users can give Muse "projects" or "goals", and it will "work around the clock until it helps you achieve the goals".

The Meta team calls this feature "night learning". Muse will organize the day's reflections into its memory, and also propose new projects on its own initiative.

Zuckerberg gave an example. He played the game *Civilization* with his daughter. Muse first helped his daughter make a game strategy guide, then took the initiative to expand the strategy guide into a set of history courses about different civilizations, and even created a new page in the app.

Zuckerberg's three-year-old daughter loves baking but "doesn't know how to bake". So he asked Muse to arrange a weekly project suitable for three-year-olds and baking beginners, and prepare the ingredients through Instacart. When they tried cake pops for the first time, Muse found that the task was more difficult than expected.

He also asked Muse to apply for a mountain climbing permit for his daughter. When the permit application opened, Muse automatically grabbed a slot for a certain day, and Zuckerberg had to make a temporary decision: "Then I have to take a day off to go mountain climbing with my daughter."

These cases reveal Muse's goal: AI is no longer just for answering questions, but becomes a long-existing "digital staff" that keeps learning about the user.

Inviting Industry Leaders to Customize a Privacy Lock for Muse

The more useful Muse is, the more information it needs to know.

To understand users' goals, it must have access to emails, messages, files, schedules, payment information and even health data. To act on behalf of users, it must be able to log in to websites, purchase goods, send messages and complete payments.

Privacy has shifted from a "compliance requirement" to the core of the product.

Zuckerberg said Meta has specially built Confidential VM, or confidential virtual machine, for Muse. He and Meta executive Nat Friedman personally invited Moxie Marlinspike, the founder of Signal, to participate in this work. Marlinspike once participated in the construction of WhatsApp's end-to-end encryption system.

Zuckerberg made a promise: Even if Muse runs in the cloud, Meta cannot see the content inside.

"We can promise that even Meta cannot see the content inside."

This is an important difference. In the past, if users wanted their agent to have a relatively private computer, they often had to buy a Mac Studio by themselves, and then complete the installation, configuration and maintenance. Zuckerberg believes that this method can only serve millions of people at most, and it is impossible to cover billions of people.

"I don't think billions of people will buy a Mac Studio, configure it and run it at home."

What Muse wants to build is an alternative solution: running in the cloud, but with privacy and confidentiality close to that of local devices.

This is also the biggest contradiction facing Meta. It owns the world's largest social network and advertising system, but as a result, it carries a heavier privacy-related historical burden than many AI startups. The more users trust Muse, the more the product needs to prove that Meta will not misuse these data.

In addition to the confidential virtual machine, Muse is also designed with several layers of restrictions.

The first is "least privilege". After users connect their mailboxes, the default permission is read-only. If they want Muse to send emails, users need to grant separate authorization.

The second is the security credential system. Users do not need to directly hand over their credit card numbers and passwords to the agent, and the system can call one-time credentials when needed.

The third is the Sentinel agent. They are responsible for monitoring the input and output of Muse, identifying potential risks such as prompt injection and abnormal data sharing. If Muse wants to log in to a website, make a payment or transmit sensitive information, users need to perform manual confirmation.

Zuckerberg said Muse cannot judge on its own whether these sensitive actions should be executed, and this restriction is written into the system architecture. But these are only designs and promises. Whether Muse can build trust ultimately depends on whether the technology is verifiable.

Creating the Profitability Effect of the Agent Network

Muse's business model is also different from traditional AI products.

Zuckerberg revealed that in the initial stage, Muse provides a free quota of up to 100 million tokens per week, including one virtual computer. Users can choose to purchase a subscription, but the vast majority of people can use it for free.

Meta's judgment is that personal agents will eventually be able to help users make money and save money, so there is no need to charge high subscription fees from users first.

"We think this thing will actually make people money and help people save money," Zuckerberg said, "That's how it will eventually pay for itself."

If users use Muse to run small businesses, the product can connect to Meta's advertising system, helping users create products, place ads and manage operations. Zuckerberg envisions that Meta may not mainly charge users in the future, but draw a small percentage of fees from transactions, and even the fees will be borne by the merchants cooperating with users.

This is a typical Meta-style calculation: distribute the product to as many people as possible first, and then generate revenue from transactions, advertisements and commercial activities generated by user behavior.

Muse is not only connected to users, but also potentially connected to Instagram, Facebook, advertising systems, payment tools and merchants. As long as users are willing to authorize, Meta can put the agent into the existing commercial network.

Meta is also testing another network effect: allowing the Muse agents of different users to learn anonymized common experiences.

At present, most AI agents are still "single-player games": one user owns one agent, and the agent only serves that person.

Zuckerberg believes that interactions will occur between agents in the future. A problem solved by one agent may help another agent find a solution faster. Meta has internally tested scenarios where multiple agents collaborate with each other.

This feature is not fully included in the first version of Muse, but it may become the key for Meta to compete with other AI companies.

What Meta is good at is not just building a stronger model, but pushing products to hundreds of millions or billions of users. When the number of users is large enough, the more anonymized experiences generated by agents, the faster the system can find new tasks and usage patterns.

Zuckerberg said Meta's advantage lies in its "social DNA". Muse can not only help users write code and run businesses, but also help users maintain relationships, spend time with family and arrange daily life.

This is connected to the product logic Meta has built over the past decade: the core value of Facebook, Instagram and WhatsApp is not just information distribution, but connecting people with each other.

Based on the network effect of its products, Mark Zuckerberg, Meta's CEO, elaborated on Muse's monetization plan beyond subscription payment at the Meta Connect conference on Wednesday.

"We believe Muse can help you make money," Zuckerberg told the room full of developers. "We make this commitment because Muse is free to access for a large number of token users, and we expect to generate revenue by charging a small transaction fee over time."

Meta's Chief AI Officer said that support for PayPal has been added, and integrations with retailers including Walmart, Best Buy and Sephora have been announced.

In terms of travel, Meta is adding connectivity to Expedia. The company will also provide grocery shopping services for Muse through Instacart.

However, in terms of foundational models, Meta still needs to continue to catch up.

Zuckerberg admitted that he once mistakenly thought that the machine learning experience Meta accumulated in the fields of information flow, advertising and content moderation could be naturally migrated to large language models.

Llama 3 performed well, but after the release of Llama 4, Meta found that the model did not develop along the expected trajectory.

But all these ideas have already been put into practice in China. For Zuckerberg, who has just squeezed onto the AI game table, all of this is still a fresh major AI breakthrough.

This article is from the WeChat official account "Intelligence Surge", written by Ai Ti, published with authorization from 36Kr.