Today AI Completes Over $200 Million First Round of Financing; Founding Team Responds to Ten Core Questions About Personal AI

By Deng Yongyi

Edited by Zhang Yuxin

Over the past month, Personal AI products including Muse and Instinct have caused a global sensation and become the most closely watched AI startup track. Among them, Muse surpassed 5 million downloads just 22 days after launch—a number that ChatGPT took 56 days to reach back then—giving the AI application track a rare boost of confidence.

An AI company called Today AI may have explored this earlier than Muse. Founder Qi Junyuan previously founded Teambition, and later oversaw various products at major companies such as Alibaba and ByteDance. After assembling the team earlier this year, Today AI launched its international and domestic versions successively starting September 9.

To sum it up in one sentence, what Today hopes to build is the user's "personal assistant." After the user authorizes Today to access their information and app data, Today will accumulate memories through ongoing conversations, with the goal of becoming a general-purpose assistant organized around the "person" rather than around tasks.

Compared with other Personal AI products, Today especially emphasizes that the AI assistant should be "like a person." This commitment is reflected in every aspect of the product: when users register, they must give the AI assistant a name; during conversations, Today has insisted from the very beginning on not dividing windows by task—all interactions are completed in a single window.

Interestingly, many years ago, when Qi Junyuan was still an entrepreneur, Teambition also launched a product called "Today" (Jintian). When that product launched, he once wrote: "I hope it can understand all the things you need to do in a day better than you do—from trivial to-dos to important work tasks, from casual personal dates to serious company meetings, and even every article you need to read—and then gently remind you what you should deal with right now... Our expectation for it is simple: it should be simple and easy enough to use."

Before Muse went viral, many investors still regarded this as merely a beautiful vision. But the Today AI team disagrees—by 2026, Claude Code and Cursor had already proven that Agents can autonomously complete complex work in professional domains; many teams are focusing on model memory and context engineering, and there have been quite a few breakthroughs. "An Agent that manages everything for you" is no longer just technical talk.

Today AI is also the team's first real answer to this vision.

△Today AI product interface

On October 21, Today AI also officially announced the completion of its first funding round of more than $200 million, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, Sequoia China, Transsion Holdings, Tongge Venture Capital, and Shanghai Jiao Tong University Hanyuan Assets.

Within the AI application circle, this is an investment of remarkable size. However, the Today AI team says that from day one, their vision has been not only to build AI applications but also to train models, so that they can become a company born entirely for "personalization"—by training their own interaction models, they can lower the unit cost of intelligence while making the product's tone, comprehension, and execution capabilities closer to "someone who truly understands you."

This is a common trend among recently prominent AI application companies. On September 28, Manus also launched its Personal Agent product "Cue," and soon afterward officially announced new funding of more than $500 million. We have learned that Manus is also training its own models.

In essence, AI application companies and model companies are moving toward each other from both ends—application companies are moving upstream to train models in order to lower the unit cost of intelligence and improve results and gross margins; model companies are also moving downstream to build applications and find revenue outlets. As the capabilities both sides need become increasingly similar, the boundary between AI applications and models is naturally becoming blurrier.

Another consideration behind startups needing to raise large sums is that building a Personal Agent requires not only training models but also connecting services and ecosystems—and that is by no means an easy path for anyone.

In China, there have already been cases of AI phone assistants being jointly blocked by service providers in ride-hailing, shopping, food delivery, and other sectors; overseas, although the platform ecosystem is more fragmented, Personal AI likewise faces access barriers—on September 20, Amazon restricted Muse's access and shopping on its platform on the grounds that "unauthorized AI Agents" violated its terms of use.

For many AI application companies, whether in terms of new startup directions or fundraising, 2026 has been an especially cold year. This has made many AI teams with Chinese backgrounds even more resolute in their choices—for example, focusing on overseas markets from day one, and then deciding whether to return to China depending on circumstances.

By contrast, Today AI's pace is bolder—it launched in the US region first in September, followed in the China region two weeks later, and quickly supported platforms such as Feishu, DingTalk, WeCom, and TickTick.

The Today AI team says this pace stems from the team's own judgment: launching as early as possible is itself part of competition, and voice, growth, and customer acquisition cannot wait until the day the product is "perfect enough."

"When it comes to Personal AI, startups have a battle to fight no matter what," they said. Although major companies have already entered the fray one after another, startups' advantages lie in greater concentration of talent density, faster decision-making, and the courage to place earlier and more aggressive bets on directions such as memory and proactivity.

Recently, the founding team of Intelligent Emergence and Today AI had an in-depth conversation. While learning about this product, we also tried to ask them to answer the many questions people currently have about Personal AI.

Below is the Today AI team's own account:

Is Personal AI a false need?

After Today launched, the thing we heard most was: Personal AI is a false need. But historically, every round of technological development begins with a great deal of skepticism, until it is gradually accepted—the previous wave of Agents was the same.

Based on our judgment, this round does indeed have incremental demand emerging. In the past, all Agents were handling office-related things, but needs in life will certainly be covered as well; everyone may need an assistant.

Of course, whether it has reached a very mature PMF stage, and whether users are already willing to pay for it—it is still too early to draw that conclusion, and we ourselves are still validating it. But we have no doubt about the demand itself.

The Personal AI boom is very closely related to Claude Code. Before Claude Code, models' function call and instruction following capabilities were insufficient, and what was built was almost an assistant with no generalization at all and all preset scenarios—it could not be personalized.

After Claude Code, Agent as a technology began to be proven capable of helping with many things, and the model's capabilities are also generalizing.

Muse's explosive popularity proved one thing: Personal AI can reach a large number of users in a short time. It surpassed 5 million downloads in 22 days after launch, far faster than ChatGPT did back then—this is a huge confidence boost for the entire track.

Muse also helped the market understand what a Personal Agent actually is. Before Muse, when we used language to describe that we wanted to build an OS Agent, it was actually quite hard to explain clearly. After Muse came out, everyone intuitively knew what a Personal Agent looks like and what it can do.

We judge that the demand truly exists—what the Personal Agent needs to handle is the large amount of trivial, high-frequency matters that everyone encounters every day. In the past, these were scattered across various apps and chat boxes, and no assistant could fully take them on.

Beyond the buzz, we think the most important thing is that downloads do not equal long-term value—whether a product is useful depends on whether users can keep using it and stay retained, and whether context can continuously accumulate. Whether users are still using it on the second day, the second week, and the second month—that is the useful definition.

Long-term retention, the balance between proactivity and interruption, the boundaries of privacy authorization, and access to platform ecosystems—these are all still common challenges facing the industry.

Without an independent app, it is hard to make users perceive the value of an Agent

If we summarize it in one sentence: Today is your personal assistant.

Today is different from those office Agents designed with tasks at the core. The latter start a session for every task initiated, and archive it when the task is completed. We want Today to be like a person—you keep chatting with it in one chat box, and you do not need to open a new window for every task.

A person's life is continuous, and you have many important goals that need long-term companionship—study, work growth, career planning. These things are very hard to achieve in task-centric products.

Back in March, we actually explored a direction similar to Instinct—it lives inside IM and has no independent app. But we found two relatively big problems: first, without a GUI (a graphics-based software interface), the guidance for users to explore the Agent's capabilities is very weak.

If you are an early adopter and are familiar with what this kind of product can offer, of course you can explore it yourself, but what we have always wanted to do is the mass market.

For ordinary people, at this stage the GUI is more like a map that can guide them to explore this maze of the Agent that hides treasure. Without an independent app and with only IM and a chat box, it is hard for users to feel how it differs from an ordinary chat assistant.

The second problem is that without a client, you cannot obtain some device permissions, and the Agent's own capabilities will also be limited.

After determining the form of an independent app, we also had to think clearly about what the Agent should connect to. We quickly realized that if we want to build an OS Agent, it should not serve only one operating system, but should connect to the various operating systems that already exist today.

More important behind this is to first figure out what problem the user wants to solve, and then choose the appropriate technical path. For example, operating mobile phone software for me is not my real need; solving travel problems is. Booking a plane ticket and operating Ctrip for me are two different needs; booking a ticket can be implemented in many ways, and it does not necessarily have to involve operating the Ctrip app on the phone.

We should first figure out what the user actually wants to do, and then choose the appropriate technical path, rather than redefining user needs for the sake of a certain technical path.

So, in terms of technical aspirations and principles, Today is an OS Agent, but for users, the value delivered is a Personal AI. From the very beginning, our device gateway was designed to be extensible to more devices, not limited to PCs, phones, or any single operating system.

To build a product for the broadest mass of users, and to deliver personalized services.

When ordinary people face an Agent, they are actually facing a black box—they don't know what problems this Agent can solve, and so they can't even think of which of their own problems they should hand over to AI to solve.

Intent should be generated by software and by the Agent, not provided by the person. So proactivity matters a great deal. The ideal state is that you very proactively, based on your understanding of the user, very precisely tell them what things you can help them do well, and you really can do them well.

Building products for ordinary people and for the broadest mass of users—this direction was aligned from the day Today was founded, and we have never wavered on this point.

So in product design, we want to make the Agent more like a person. For example, we have always insisted on a single session, meaning only one dialog box.

In the past, many Agent products divided dialog boxes by task, which was actually a technical compromise made when context capabilities were limited—once there is a lot of context, if the engineering is not mature, things will definitely become chaotic, and the model's attention will also become scattered.

But with breakthroughs in long-range memory and cross-session memory, we believe that in the future the Agent should take a single-session form, so we have also done a great deal of engineering optimization: for example, using the smallest possible model, increasing Context information density, reducing latency, and designing natural authorization flows and profile displays. Also, by replying in separate Bubbles to reduce the mechanical feel and enhance the human feel.

In many personal scenarios, users need to remember "which dialog box that thing from last time was in" before going to look for it, which is itself an extra action. The most natural way is to face the same dialog box and not need to manage a pile of sessions yourself.

In actual testing, for the same task, the token consumption of an Agent Team running through it is clearly higher than that of a single session, and on everyday trivial matters the efficiency is not necessarily better either; instead, it increases latency.

Agent Team does indeed have value in professional scenarios that require mutual discussion or review, such as code review—but for Personal AI, it does not.

What is a truly useful Personal AI?

In the past, users were unwilling to let AI handle many everyday small tasks, and the core reason was that the delegation cost of explaining the context was higher than doing it themselves. Schedules, appointments, recording the baby's vaccinations, recording the pet's medication—these very trivial things, I never used to think I would use AI for them, because every time I had to explain the background all over again.

Today's key design is to lower the delegation cost.

To achieve this goal, we believe the two most important things in building a Personal AI are: the memory mechanism and proactivity.

We accumulate memory through two paths: first, the user authorizes a Connector, letting Today enter your digital world—Feishu, health, Calendar—and it will go learn about you on its own; second, through everyday conversation, accumulating memories of your preferences and concerns during chats.

For example, for a period of time I often talked with it about papers related to memory. It found that I cared about this direction, so it proactively searched the internet for related papers and pushed them to me.

Besides everyday trivial matters, there is another category of long-term goals that task-centered office agents find very hard to achieve—learning, career planning, fitness, and so on. Today will help you formulate long-range plans in memory and proactively remind you of progress. For example, I delegated language learning to Today; it sends a lesson every noon, and improves the content during the process of my interaction with it.

The core difficulty of proactivity is that "being too proactive is harassment." We adjust the frequency of proactive pushes based on user feedback: first, whether the user actually replied to the proactive message; second, whether the user at least read it. If no positive signals are collected at all, we consider that we really are doing something harassing and will proactively reduce the frequency.

In the future, people's time will increasingly be liberated, and this is the value that Personal AI should bring—helping you reduce interruptions, load, and context burden. If it cannot bring that, then there is a problem with the product.

However, the operating cost of Personal AI is indeed much higher than that of the previous generation of AI products, and cost is indeed a problem now. Some market research speculates that instinct's daily operating cost is around 1-3 US dollars; calling the sandbox, search, the proactivity mechanism, and long-range task execution are all very large costs.

Does Personal AI have barriers?

To be very honest, Personal AI is currently in a stage of highly homogeneous competition—everyone has seen the role of memory and proactivity. It is hard to name an extreme differentiation, and even if there is one, differences in one or two feature points will soon be smoothed away by competition.

What we value more is not some single-point feature, but the overall product direction: interacting like a person, single session, making it truly usable for ordinary users, advancing long-term goals, and proactively discovering things users may need. Together these constitute Today's product philosophy, rather than some exclusive feature.

In particular, many veteran internet users already have a very, very large number of digital assets in their cyber world. If you give their Context to an Agent, can that Agent use this Context more efficiently? This is actually quite hard to truly create a gap.

Take email, for example—everyone will integrate it. Simply integrating and processing email is not difficult; what is truly difficult is finely categorizing emails, which requires a multi-layer Sub-Agent architecture that distills the know-how and insights of specific scenarios into it.

The barrier may not come from any single technology, but from the Context accumulation formed over long-term use. The longer users use it, the deeper the Agent's understanding of you becomes. Although technically you can export memories and switch to another product, the friction cost is very high. Although this kind of Context lock-in is not an absolute barrier, it does create switching costs.

Should Personal AI first be general or first be vertical?

Before building Today, some members of our founding team had also built vertical Agents. After finishing, we found that 80%-90% of the effort was all spent on the underlying engineering—how to make the Agent run more stably, how to build a memory system, how to make it remember the user's last progress. The tuning truly aimed at the scenario may have been only 10%.

If only 10% of our product effort is specific to a certain scenario, then when generalized to other scenarios, isn't it about the same? Health, learning, financial management... the core capabilities are all that same underlying thing: Context comes in—precipitates into Memory—then outputs results. And basically all scenarios can form such a loop, with the upper layer actually being a very thin layer of Harness.

Reasoning backward from this logic, the barriers in vertical scenarios are very thin—users actually have no need to download an Agent for every need.

In the future, everyone may need only one general-purpose personal assistant, which will become a unified entry point. Only needs that are extremely professional and cannot be handled by general solutions will be the space for vertical Agents, and they can also be connected by this general-purpose personal assistant.

Overseas, several different product paths have already emerged for Personal AI. For example, Town entered from the email scenario, first going deep into a single scenario, with a paid conversion rate of about 15%; I think the path of this type of tool product is viable, but it is different from Today.

There are also subscription-oriented products on the market, such as Grok Bot, with about 400,000 weekly active users. What Instinct explores is trusted networks: users first communicate with their own Agent, then the Agent negotiates arrangements with the other party's Agent, and finally feeds the result back to the user—the chain is relatively longer.

However, the boundary between general and vertical Agents is not so clear-cut. Now, in work scenarios, Today more often plays a collaborative role: scheduling time, recording important matters, doing research, and reminding about forgotten things. Work and life will definitely overlap on Personal AI; it is very hard to completely distinguish in a person's life which parts are work and which parts are life.

How to balance privacy and delivering task results?

This is a very real problem. If you don't take Context, there is not much you can do for users. Based on this premise, what we can do is make privacy controllable and transparent.

First, complete transparency. In our memory bank, everything the Agent records is completely visible to users, including basic information, preferences, and interests, and users can view it at any time.

Second, complete controllability. Users can delete memories at any time, and after deletion, the underlying database will not retain them at all.

We will also do progressive authorization. Only when Today truly needs certain information will it request permission, rather than asking for all authorization right from the start.

Authorizing access to relevant information does not mean authorizing Today to make all decisions for users. Schedule adjustments require user confirmation; when it comes to payments and important decisions, the final confirmation right remains in the user's hands. Proactivity means discovering and preparing in advance, not skipping the user's decisions.

Actually, we were also very worried about the cold start problem at the beginning. Will users really authorize? But internal testing found that the authorization rate is still quite high. Of course, it should be noted that these users still lean toward an Early Adopter profile and cannot directly represent the attitude of all mass users.

Our current user profile distribution is actually quite broad. Early Adopters like to explore various Agent capabilities; another type of user does not explore much but uses it steadily in life—recording, content understanding, personal management. Their usage trajectories are very different.

Users' Aha moment generally occurs at two time points: some users get it during the first day's Onboarding session, and they will feel, "This Agent actually understands so much about me," with a feeling of meeting an old friend.

Another type of user gets it when receiving a proactive push on the second day—it actually knows what I am interested in and directly pushes relevant information in front of me.

Whether doing Agents in China or overseas, the difficulty of commercialization is similar

We do not agree with the claim that Personal AI can only succeed overseas. Different markets simply have different scenarios; it is not that some market naturally has no opportunity.

Common overseas scenarios are email, calendars, and shopping price comparisons, and these indeed have mature Connector ecosystems. The domestic email scenario is not mainstream, but a large amount of Context is scattered across apps such as Feishu, Health, and Calendar.

Many people also question Personal AI on one point: that it cannot succeed without WeChat data, but we are very optimistic about this.

Most of the trivial matters users truly want an Agent to handle for them are outside acquaintance networks. What WeChat fundamentally accumulates is the acquaintance social network, which is precisely the part people least want an Agent to handle for them. As long as we can connect to Connectors such as Feishu, DingTalk, and WeCom, we can likewise collect enough Context.

It is true that data silos exist in China, but so far we have not encountered obvious obstacles when connecting to other Connectors, and the Connector market is gradually opening up.

As for PMF, to be honest, it is still being validated. Commercialization is also not our top priority at this stage; the top priority is delivering scenarios that are truly valuable to users.

Looking at our current users, the overall data of domestic users is not bad, but the Connector ecosystem is less mature than in the US, and privacy compliance issues also require extra consideration.

Two weeks after launching in the US, we also launched in China. China is a relatively homogeneous, unified, large single market. Sooner or later we have to do this market, so we might as well do it now.

The difficulty of commercialization in the US and China is, in a sense, the same. Muse is currently free, and other companies launching products in the US actually find it hard to charge—under such a free-competition landscape, commercialization for Personal AI is not easy even overseas.

We will explore subscriptions in the future, but for now subscriptions may not be the most suitable business model; shopping guides and advertising would harm the user experience. Right now it is a bit like the stage before Google invented the advertising business model—if Personal AI succeeds, its business model may be very different from what we imagine today.

AI is inherently driven by both engineering and models. Your goal can be to solve users' problems at the application layer, but to solve them well enough, you must have your own model capabilities and sufficiently good engineering and infra capabilities.

On the application side, you need a clear definition of user problems and solutions, and at the Harness and infra levels you need your own understanding and determination to build—always ready to adjust and modify the model.

Why was Anthropic able to build itself up in the face of such a strong competitor as OpenAI? It first built Claude Code, collected very good trajectory data, and used its own Harness for post-training, and the flywheel of model and application started spinning.

Cursor training its own models follows the same logic: lowering the unit cost of intelligence, raising gross margin, and improving intelligence.

But the model Today needs is not entirely the same as the direction of current mainstream models. We need more of an "interaction model"—one with a better sense of interaction, more human-like, and a more natural tone and style, while also being able to actually execute tasks.

Model companies are moving up into applications, and application companies are moving down into models, and the capabilities of both sides are getting closer. In the future, there may be no such thing as a pure application company or a pure model company.

"Startups have a battle to fight no matter what"

Basically all the major companies we know have entered the field. The speed of building products and forming consensus is now very fast, and competition is fierce.

To be very honest, startups will not have an absolute advantage. But no matter what, they have a battle to fight.

Startups' advantage lies in more concentrated decision-making and a greater willingness to take risks. For example, we may be one of the earliest teams to try building memory into an Agent.

Startups will have a short period of product leadership. Although it may not form a barrier-like advantage, this period is still important.

Our judgment is that the high-delegation-cost small matters Personal AI needs to handle do not all occur in acquaintance social networks. Work and life Context is also scattered across apps such as Feishu, DingTalk, and WeCom. As long as there are Connectors, there is an opportunity to access this information.

We also observe that the Connector ecosystem is gradually maturing and opening up. In the future, MCP and Agent-friendly service providers will certainly exist, and technologies such as Phone Use may all become ways to further open up the ecosystem.

However, whether every domestic company's Personal AI can ultimately build its own ecosystem remains a question.

Image source | "Interior in Strandgade, Sunlight on the Floor", Vilhelm Hammershøi

Welcome to exchange views

This article comes from the WeChat public account "Intelligent Emergence", author: Deng Yongyi, published by 36Kr with authorization.