Muse, Meta’s inspiration, is only a little better than “AI ordering coffee”

Text | Xiaojing

Editor｜Xu Qingyang

Muse became the "goddess of inspiration" of the giant Meta, and single-handedly raised the stock price of the giant Meta by 15%.

Muse, Meta's smart phone, was released on September 8. As of September 18, it topped the free list of the US Apple App Store, occupying the top spot for three consecutive days. By September 21, Meta's stock price reached a climax, rising 11.4% in a single day, increasing its market value by approximately US$192.3 billion.

Its popularity has even spread to the semiconductor industry, and even CPUs have become the target of market pursuit, driving AMD to rise by 9.9%, and its market value has exceeded US$1 trillion for the first time.

For Chinese users, this scene seems a bit familiar. From "ordering takeout in one sentence" to cross-application price comparison and order placement, personal agents (AI agents that can perform tasks autonomously) have demonstrated many exciting scenarios in the past year. Even the personal terminal "Doubao mobile phone", which is said to be a born AI personal assistant, has also appeared, trying to let AI directly operate the screen and complete tasks between different apps for users.

Many people have drank the coffee served as an “AI treat” at the press conference.

But the real experience is still nakedly cruel. The smooth and continuous operation in the press conference demonstration may be stuck in login, payment or a sudden pop-up window in daily use; the limitations of banking and other applications make it impossible for AI assistants to get around it.

Now Muse debuted with similar product promises, and was even quickly blocked by Amazon. Why are these familiar abilities and obstacles given so much room for imagination in the U.S. market?

Picture: Meta Chief Artificial Intelligence Officer Alexandr Wang (Wang Tao) and Muse, the picture is generated by AI

01The reason for people to download is "help me get my money back"

Muse was released on September 8th and reached the top ten days later. Recurring protagonists in communications are chargebacks, inactive subscriptions and insurance bills. Meta chief AI officer Alexandr Wang promotes the "MuseMoneyChallenge" (Muse money-saving challenge) to encourage sharing of money-saving results. Foreign media reported that one entrepreneur said that Muse helped him find a car insurance plan that was $3,500 cheaper per year; however, the report also reminded that many of the sharings here actually came from Meta employees.

But this type of case gives users a reason to try: to check whether there is any money paid in vain. For people who don’t care about AI, the amount of savings is easier to understand than the benchmark scores of a bunch of models, and is more suitable for forwarding.

Muse’s product dissemination is also solving another obstacle: users don’t know what tasks to assign to the AI. The Muse design team disclosed that in early testing, there were too many functions that made it difficult to start, so they added a design that actively generates suggestions based on goals and conversations. Technology reporter Alex Heath also praised Ideas and Feed pages in public posts because they will propose tasks that they have never thought of.

Wharton Business School professor Ethan Mollick commented: Other AI laboratory systems have been able to accomplish similar things. Muse’s dedicated design makes it easier for ordinary people to get started; Meta has also invested a lot of free computing power to support the experience.

Meta is obviously better at C-side gameplay, providing users with a straightforward appeal of "saving money". The functional design makes it easy for users to get started. The entrance is the most familiar WhatsApp, and more importantly, there are free tokens that can be squandered. Users can try Muse for free first, and can also let it handle tasks in WhatsApp just like sending messages to contacts, which lowers the threshold for trying and using it.

Money-saving cases, task suggestions and familiar portals constitute several noteworthy links in this round of communication. Moreover, the money-saving story provides a communication trigger, which is transmitted to the performance of the list and becomes a signal for investors to re-evaluate Meta.

1. Where does the novelty of technology lie?

So, does Muse have any technical innovations?

The model is responsible for understanding the requirements, arranging the steps, and then calling tools such as browsers to open web pages and fill out forms. This computer runs on Meta's servers, so users can close the mobile app and the task can still continue. Meta uses virtual machine technology to allocate an independent running space for each user to save personal data and task progress. Computer Use (computer operation ability) allows AI to use software on this computer.

There is already a precedent for Muse’s “cloud computer”. According to the official OpenAI documentation, ChatGPT Work runs in a virtual machine-based sandbox, and the execution state is associated with the user. The environment can be reused across tasks, and can also be replaced after retaining part of the state. Its cloud browser continues to work after users turn off their computers. GPT-6 can also write code to operate the browser and desktop, and cooperate with plug-ins to complete tasks.

But the engineering feature of Muse is to draw permission boundaries around personal data. According to Meta disclosure, each user has a Muse Secure VM (Muse Secure Virtual Machine) as the main storage location for personal data. Execution programs and tools are put into restricted containers inside the virtual machine, and account credentials, security checks, and persistent states are managed separately; model inference calls external infrastructure through interfaces, and not the entire large model is installed in the personal virtual machine.

We can think of it as a partitioned workspace. Agents can read files and run programs in the operation area. Account credentials are kept in another place. Connector actions and network access also need to go through Sentinel (authority review agent). This design is to make it difficult for programs that process unfamiliar web pages to directly obtain full permissions even if they are misled. However, the actual safety effect still needs to be tested and verified.

The engineering value of this design lies in organizing execution capabilities and authorization boundaries: calling the connector when the service has an interface, and trying browser operations when there is no suitable interface. The interface requires cooperative access, and interface operations must deal with page changes and verification.

This set of designs serves continuous delegation, retains progress, continues processing transactions based on time or related events, and requests authorization before critical actions. Independent virtual machines, containers and permission audits all have existing technical foundations; it is currently impossible to prove that Muse's complete solution is ahead of all competitors.

Figure: Schematic diagram of Muse security architecture. Each user has an independent cloud virtual machine; task execution programs, account credentials, and persistent data are managed separately, and Sentinel is responsible for auditing external access and connector operations.

Domestic products have been tried for a long time with similar forms. AutoGLM 2.0 in 2025 provides both cloud phones and cloud computers, covering mobile applications, web pages and office tasks; Doubao phones are directly purchased as "AI native phones" that can operate apps on users' phones; Qianwen app takes advantage of Alibaba's own ecological advantages to access ordering, payment and other services through cooperative interfaces, reducing screen-by-screen clicks.

Interface operations and interface calls can be combined, and tasks can be executed locally or in the cloud. AutoGLM's "cloud computer" refers to the environment for AI operations, while Muse's "independent Linux virtual machine" mainly describes how it divides the running space for users. The two are at different levels of comparison.

Muse explained that some AI mobile phone experiences can be delivered through existing mobile phones and cloud services; to call applications, cameras and system permissions on the mobile phone, device support and user authorization are still required.

It seems that Muse is essentially the AI ​​assistant that domestic users are familiar with and is good at ordering coffee. However, this AI assistant has no way to ignite the excitement of users and capital in China for the time being.

1. Change your thinking and don’t just order coffee.

Muse is for individuals and currently focuses on life management, but it can also generate documents and web pages. "Personal user" and "life tasks" are not the same concept. One person's assistant can handle home and work at the same time.

China has not given up on life agents either. Alibaba disclosed that Qianwen facilitated nearly 200 million orders during the Spring Festival; the event was accompanied by a 3 billion yuan incentive, and the order volume cannot alone prove natural demand and profitability. At the same time, Doubao launched a professional version for complex offices in June, with a monthly subscription starting at 68 yuan, and Tencent launched WorkBuddy Enterprise Edition. Domestic life and office scenes are being explored in parallel, but the office scene is more popular.

It is easier to "charge" for office work, sort out a batch of forms, complete surveys, and the labor saved can be compared with the delivery time saved. However, ordering a cup of milk tea originally only requires a few clicks. The incremental convenience brought by AI is limited, and taking over after making a mistake will only increase the trouble. For service platforms, such tasks may still generate value through order conversion.

Muse is smarter in selecting scenarios, such as canceling forgotten subscriptions, comparing renewal quotes, tracking refunds, and coordinating family schedules. The frequency of these tasks may not be high, but they can save money, time and worry. The potential lies in handling long-delayed tasks for users and avoiding missing deadlines.

The Muse designer cited a personal case: when her assistant was sorting out her children's school opening emails and the school website, she discovered that registration for a sports selection was about to close, and promptly reminded her.

In contrast, domestic life services already have many convenient processes, which may reduce the incremental value of regular orders. However, the values ​​of insurance renewal, app subscription, after-sales appeals, and care coordination are still worth verifying.

Figure: Official demonstration video of Muse running

1. Is this business really established?

Stratechery founder Ben Thompson said on September 21 that Muse is the easiest personal agent he has ever tried. The Muse Spark 1.3 it uses has not yet reached the level of cutting-edge models, but it can already support a good product experience. In his view, once users hand over their account information to an assistant and let it participate in daily tasks, the cost of replacing products will increase.

Meta also develops basic models. This round of popularity more directly reflects the market's revaluation of the prospects of model + application consumption. When the model capability reaches a usable level, Meta can rely on product experience and distribution channels to allow users to continue to delegate tasks and earn income from it. Analysts at Truist predict that Muse could bring $28.5 billion in incremental revenue to Meta by 2030.

Oppenheimer analyst Jason Helfstein maintains a neutral judgment: paying people may already have a ChatGPT or Gemini subscription, and trust is also a threshold. It requires about 115 million paying users to significantly affect Meta's profitability; the model uses assumptions of US$20 per month and 80% incremental operating profit margin. The more free computing power there is, the harder it is to achieve profit assumptions.

But this also explains from the side why the market can overflow: the market is revaluing the scale of consumer agents. CPU and memory are also affected by this expectation. In addition to model inference, the browser, code, and scheduling require a CPU (central processing unit); maintaining running programs requires DRAM (server running memory). The more concurrent tasks there are, the greater the two types of load may be.

When the user has no tasks, Meta can pause the cloud virtual machine, save the task progress to the hard disk, and resume it after receiving new tasks. Therefore, "one virtual machine per person" does not mean that each person uses a fixed amount of CPU and memory throughout the day. The popularity of Muse has led investors to expect more people to use such services, but the actual number of chips needed to be purchased depends on how many tasks are running simultaneously and how long each task takes to run.

05Muse still hasn’t solved the fundamental problem

Muse puts the user's data and execution environment into a dedicated cloud virtual machine, but do users dare to let it have long-term access to emails, orders, and payment information? According to foreign media reports, Muse during internal testing once stopped refreshing after monitoring ticketing for about 15 minutes, and private data exposure issues also occurred. Meta confirmed that it had postponed the original April release to improve security.

Even if the user is willing to authorize, the Agent may not be able to access all services. On September 21, Amazon blocked Muse from browsing and purchasing items on behalf of users, arguing that such access and transactions were without permission. Doubao Mobile was also forced to temporarily suspend the operation of financial apps due to login and risk control issues. The product paths of the two are different, but they still encounter the same problems. User authorization cannot replace the platform's permission. The platform still has the ability to decide whether the Agent can access its own services.

Picture: On September 21, user Jonathan Wegener shared the prompt Muse encountered when accessing Amazon: Amazon said that the unauthorized AI Agent continued to access the website in violation of its terms of use.

Every powerful platform has a strong incentive to guard this door. If the agent selects products and completes transactions for users, the platform may lose advertising revenue and recommendation opportunities brought by users browsing pages. Shopify chose to connect to Muse, and Amazon chose to block it without hesitation. The platform’s stance is a difficult nut to crack.

In addition, Eric Sheridan, head of research at Goldman Sachs, predicts that consumer agents may rely on advertising and subscriptions to monetize in the future; but if the assistant charges promotion fees, users also need to know whether the recommendation results have been affected. It is difficult for consumer agents to have a clear motivation for paid subscriptions like agents in office scenarios. If you still follow the old path of advertising and recommendations, you will inevitably get someone else’s cake. From whom to collect money, how to collect money, and how to share the money after collecting it is still a difficult problem.

On September 22, Meta's stock price closed down 0.63%. As of the time of publication, it rose by 0.6% after the market opened. The market started to gradually enter a cooling-off period from excitement. After giving Muse a try, the next step may be to look at user retention.

However, the Meta Connect conference is about to be held. According to foreign media speculation, this conference, which used to be dominated by hardware, may update more information about Muse. We can wait and see whether Muse has any real new ideas.