Beneath the new entrances and the old city walls, how far away is the "Chinese version of Muse"?

By | Li Zhaofeng

Edited by | Zhang Yuxin

10.4 million cumulative downloads — that is the figure of Muse, Meta's AI personal assistant, in less than two months after its launch. This has rapidly sparked a boom in Personal Agent products across China, and building a "Chinese version of Muse" has become a topic drawing attention from the entire industry.

A group of startup teams are accelerating their entry into this track. Qi Junyuan, former head of Doubao PC product who left ByteDance last year, started a new business this year and launched Today AI, a product focused on continuous memory and proactive services. Manus, which rose to fame for its general task execution capability, recently rolled out Cue, a personal assistant equipped with tools including email and computers for Agents. Rabbit, which started with the AI hardware r1, now shifts its focus to the new generation OS3: users can access this assistant either through web pages or the r1 device, and the team is also developing supporting new-generation AI hardware...

These teams are testing different product paths: one is to provide independent Personal Agent software on mobile phones or computers, and the other is to deeply integrate with hardware and take hardware as the entry point.

On the software side, some products emphasize continuously understanding users and offering help proactively; some equip AI with independent computers and accounts to execute tasks; others try to connect to users' existing computers and receive instructions through multiple entrances such as software and hardware. These approaches are not mutually exclusive, but each team chooses a different starting point.

However, in China, the first challenge that Personal Agent startups have to face is the free products from large tech giants. Doubao and Qwen have already gathered a large number of users, and ByteDance and Alibaba are continuously adding task execution capabilities to them. Giants can push new functions to their existing users, while independent Personal Agents urgently need to answer the question: how many users will keep using the convenience they provide, and even be willing to abandon the free products from giants to pay for their services?

On the other hand, gaining users does not equal being able to complete tasks for them. Domestic services such as shopping, travel and payment are concentrated on various platforms, and Personal Agents need to obtain permission from service providers to perform tasks across apps. After the second-generation Doubao mobile assistant recently launched the SAEP protocol, apps including WeChat, Taobao, Alipay and Meituan appeared on the list that refuses AI operation on behalf of users. At the current stage, assistants are far from being able to move unimpeded across different apps.

Recently, Intelligence Emergence talked to multiple Personal Agent startup teams, and we found that although a general assistant like Muse is the industry's vision, for startup teams, they have to make trade-offs in their current product logic.

Personal life assistant LiDan starts with memorizing things and scheduling, and then gradually takes over more life scenarios; WishSeen, which also develops personal Agents, plans to focus on polishing one life function first; UniCone, a personal assistant focusing on proactive services, gives priority to groups including creators and personal IP owners. Some of them start with functions, while others start with customer groups, but their paths share a common point: they first strive for a part of users who keep using the product and are willing to pay, and then consider expanding the service scope of the product.

Cao Fangyongzheng, co-founder of WishSeen, said: "Personal Agent is a good narrative, but it will definitely not work if it cannot keep users using it continuously and keep them willing to pay for it."

When the boom of "AI taking over ordinary people's lives" sweeps across China, the differences in user habits and service ecosystems also emerge. Can the "Chinese version of Muse" take off along the same path? Even including Muse itself, can these products really become a viable business? Investors, startups and large tech giants are all involved in this gamble whose answer is yet to be revealed.

Do You Need to Hand Over Your Life to an Agent?

To enter the lives of ordinary people, Personal Agent has to prove first that such demand truly exists.

Since the beginning of this year, various Agent products have achieved explosive growth, covering scenarios from coding to work, and now to personal assistants oriented to life scenarios. As application scenarios become more and more segmented, Agent manufacturers must give users a reason to use their products.

The team of LinkAI founder Chai Yujie is based in Shenzhen. Starting from the chatbot project chatgpt-on-wechat, they gradually developed it into CowAgent, an open-source assistant that can call tools and operate computers. The team also provides the enterprise platform LinkAI to help customers build agents, knowledge bases and workflows.

"Actually, Personal Agent is not a very new category or technology, it is more of a usage scenario," Chai Yujie told Intelligence Emergence. His concern is that ordinary users already have Doubao and Qwen, and professional users also have office Agents. "If we develop another similar product, it is actually a problem how to get individual users to pay for it."

In the consumer AI report released by US venture capital institution Menlo Ventures this year, it is found that among US AI users, the proportion of those who use both general assistants and AI tools dedicated to specific types of tasks has risen from 60% last year to 72%. The institution believes that even if users are already using ChatGPT, they may still choose another more user-friendly tool for other specific demands.

This report indicates a possibility: Personal Agent and general AI tools may not conflict, on the premise that they can provide real convenience in daily scenarios.

LiDan is a personal assistant that helps people manage their lives. When users state what they need to do, the product extracts to-do items from chat records and puts them on the main interface. Founder A Tang calls this design "event-centric": at present, it first helps users memorize things and arrange schedules, and in the future it hopes to further complete tasks such as booking hotels and flights for users.

The social media homepage of LiDan writes "5 people in Chengdu are building the Chinese version of Muse". But A Tang admitted that they are already planning to go global: "Our future focus will be on overseas markets, but we want to polish the product through the domestic market first."

If a Personal Agent wants to continuously serve users, it has to know when to offer help.

The UniCone team from Shenzhen is developing a proactive personal assistant, trying to sort out to-do items and break down tasks from information such as user-authorized screen content and chat records. "It has to be proactive, which is different from the previous situation where it only helps you complete a certain task," said the CEO of UniCone. This requires continuously perceiving the user's situation, computing in the background, and judging the timing to provide help.

Such judgment is not easy, especially since users' personal information is constantly changing. Lance Ulanoff, editor of tech media TechRadar, once shared a scenario he encountered when testing Muse: Muse found that the Portugal travel plan in old materials was inconsistent with the Italy itinerary in recent emails, and then Muse confirmed with Ulanoff and corrected the destination information.

Muse's proactive service capability benefits from its real-time access to a large amount of user information in emails, but this path is obviously hugely different from China's user ecosystem — in China, a large number of information and services exist in multiple super apps such as WeChat, Alipay and Dianping. If Personal Agents cannot obtain these information, can they provide qualified proactive services?

Suki, product manager of Today AI, is relatively optimistic. She believes that there are still available information sources outside WeChat. "WeChat cannot monopolize all the information of a person's work and life. At least a lot of work information is not deposited on WeChat," Suki said. Many users of Today AI have authorized the assistant to access data from Feishu, health apps, calendars and other applications, which can also provide clues for the assistant to understand users.

But not all Personal Agents want to take charge of users' lives from the very beginning, especially startups that have not yet raised large amounts of capital.

WishSeen, which is developing a mobile personal assistant, plans to retain memory and proactivity, but only focus on polishing one life function. Co-founder Cao Fangyongzheng explained that pursuing an all-in-one product will make it difficult for the team to afford the cost and stand out from peers. "I would rather make a certain function good enough that users are willing to use it every day, and even pay for it." For startup teams, temporarily putting aside the idea of a universal assistant and making one specific life function solid seems easier to find an upward breakthrough under the resource squeeze of large tech giants.

New Entry Point and Old Barriers

At present, this batch of Agent startup teams have basically verified the feasibility of their product logic in a small scope, but this is only the first step of building a product — from 0 to 1, startups can verify their ideas among hundreds of seed users; but moving forward, they will face resource competition with large tech giants.

In January this year, Alibaba demonstrated using Qwen to order milk tea at its press conference, which was supported by the access to services such as Taobao and Alipay. Wu Jia, then president of Qwen C-end Business Group, said at that time that he hoped in the future "most life and service scenarios can be completed with only the Qwen App". By September, Alibaba launched Qwen Intelligence for mobile phone manufacturers; in October, Qwen Go opened experience applications, supporting functions including claim settlement, price protection and appointment.

Since the end of September, Doubao under ByteDance has also been successively connected with travel and utility bill payment capabilities. Large tech giants can add life service capabilities to their existing products, and have sufficient resources to connect part of the external ecosystem.

But for independent Agent teams, customer acquisition has to start from scratch.

"Large tech giants do have abundant capital, can provide free services, and can spend a lot of money on advertising, which we cannot compare with," A Tang said. He sets the breakthrough point of LiDan on interaction: "If our product is exactly the same as theirs, we definitely cannot beat the giants."

"Reminders, to-do items and other such functions should not be sunk in the conversation stream, they should be extracted and placed on the main page," A Tang said. "After the truly valuable content is brought out, chat records are not that important anymore."

Talking about whether startups can widen the gap with large tech giants through product interaction, Chai Yujie said: "There are differences, but the duration of such differences will definitely get shorter and shorter." The opening of frameworks and the progress of development tools make it easier to build an Agent, and also make peers catch up faster.

Agents can be built quickly, while model manufacturers are evolving even faster, which makes many investors relatively cautious when facing AI application projects. An investor from an industrial fund told Intelligence Emergence: "Now models are iterating so fast that the software application you invest in today may be eliminated by a new model tomorrow."

However, from the perspective of startup teams, it is still worth striving for a short-term lead.

"We may be one of the earliest teams trying to implement memory in Agents, and also one of the more radical teams in developing proactive strategies," Suki from Today AI said. In her opinion, large tech giants have advantages in distribution and resources, but they may not keep investing continuously: "If the data is not good after the product goes online for a month, or it cannot generate hundreds of millions of revenue, the boss may make a snap decision to cut the project." The product lead of startups may only last for a period of time, "although it may not form a barrier advantage, this period of time is still very important".

However, even if they gain users, the threshold of "handling tasks" still depends on the coverage of services provided by Agents. Alibaba and ByteDance can first connect their own services, but users' lives will not only unfold in the ecosystem of one company.

If AI wants to handle things for people, it will have to step into other platforms' territories sooner or later.

"If everyone shares permissions with Doubao, each App will only have a background left, which is unacceptable for all major manufacturers," when talking about the resistance to cross-App task execution, Chen Mo (pseudonym), an Agent product manager from a large tech giant, told Intelligence Emergence.

Taking ByteDance's recent actions as an example, before the launch of the second-generation Doubao mobile phone, the Doubao team launched the "Screen Automated Operation Statement Protocol" (SAEP), allowing App developers to declare whether they accept AI operation on behalf of users, and they can also specify which pages and which operations the assistant is allowed to access. When the public notice ended on October 14, the official website announced that the rejection list included as many as 155 apps, including WeChat, Taobao, Alipay and Meituan.

At least for now, the ecological barriers between large tech giants are still solid. It is not easy for startup teams to let assistants cross these barriers, and neither are peers from large tech giants.

Even Muse has encountered this threshold. In September, Amazon blocked it from making purchases on behalf of users. Foreign media reports show that Amazon believes that third-party applications that buy goods from other merchants on behalf of customers "should operate in an open and transparent manner and respect the service provider's decision on whether to participate".

So when will the walls of China's App ecosystem open to external AI assistants?

"In the long run I think it will open up, but it may take several years, and it definitely won't happen in the short term," Chai Yujie admitted. Consumer platforms have their own data and business barriers, and it is difficult for small and medium-sized teams to connect with them one by one.

"If a certain large tech giant does not open up, when other platforms open up, traffic will flow to those open platforms," the CEO of UniCone believes that when users look for services through Agents, platforms will eventually choose to open interfaces.

Chen Mo, who has experienced multiple Agent projects in large tech giants, is more cautious about the speed of platform opening. He believes that large tech giants usually push changes very slowly, and cross-platform connectivity "cannot be solved in at least five years".

At least at the moment, this contradiction seems unsolvable: Personal Agents want to manage users' daily lives and save them the trouble of opening apps one by one; but platforms are unwilling to give up the opportunity to interact directly with users — where users stay, business happens. But if Personal Agents can never take over real life affairs, who will pay for them?

Who Will Pay for Personal Agent?

The difficulty in the commercialization of Personal Agent lies in that the services it provides need to have a good enough experience to make users stick to the product, and even be willing to pay. But this means that in the cold start process, the team has to continuously bear the computing power costs brought by background continuous operation and proactive services — only after burning money to get through this step can they discuss how to get users to pay for subscriptions.

"What we want from starting a business is not just user volume, but revenue. Each additional user brings additional costs, and we cannot afford to support so many users for free," when talking about the commercialization of the product, A Tang said that LiDan may stop its free quota later, to see how many people are willing to pay, and then decide whether to invest in advertising and look for KOLs. "We will not spend a lot of money to expand the user scale at the beginning."

Different from traditional Internet products, a large part of the cost of Agent products comes from the computing power consumed by model operation.

A Tang and his team tried multiple models when developing LiDan, "either too expensive or too slow". When they switched to cheaper and faster models, "sometimes the model just cannot understand what people say", and tool calls and long-term tasks are also prone to problems. However, he is still optimistic about the long-term trend: "I am optimistic about the improvement of model capabilities and the decline of prices. In the future, easy-to-use models that can help ordinary people complete tasks will definitely become affordable for everyone."

Not only LiDan is constrained by costs. The WishSeen team once calculated that to make the assistant both general and proactive, and have strong execution capabilities to meet their expected effect, the cost per user per month is at least 1000 yuan. "No user will pay 1000 yuan a month for your product," Cao Fangyongzheng said.

This makes the competition of Personal Agent carry the meaning of resource consumption. Large tech giants can subsidize traffic and computing power, while small startup teams must calculate the account of a single user as early as possible. "The biggest problem is not technology, but cost," Cao Fangyongzheng said.

Many entrepreneurs believe that the existing models have not yet found the balance required by Personal Agent: strong enough capability, fast enough response, and low enough cost. "The current models and Agents are not designed for the Personal AI scenario," said the CEO of UniCone, who researches AI at Tsinghua University. UniCone tries to train smaller models, and also tries to post-train existing models, hoping to reduce the cost of continuous services.

Finding payment demands from specific groups is also a kind of operation idea. Anish Acharya, partner of A16z, wrote in an article in July last year that high-priced AI subscriptions usually serve specific demands such as programming and video creation. He advocates that startups can deeply meet the needs of such a small group of users and charge accordingly, and they do not necessarily have to pursue the mass market first.

As the industry-leading product Muse, in addition to providing free quota, it directly launched paid plans of 20 US dollars and 100 US dollars per month, but it has not