Ant Lingguang CTO Mo Linjian on Lingguang 2.0: How Personal Agents Shift from Building Tools to Getting Things Done
In an interview, Ant Lingguang CTO Mo Linjian explained that Lingguang 2.0's product focus has shifted from generating flash apps from a single sentence to helping users accomplish goals that span time periods, with the ability to remember requirements, track them continuously, and proactively remind users. He noted that 1.0's flash apps were not a high-frequency need, with three-day retention in the low 20s percent, so 2.0 places more emphasis on whether it truly solves problems. Lingguang 2.0 has integrated services including Alipay bills, email, DingTalk, and Feishu, connected through the MCP and A2A protocols; risky actions such as payments are still confirmed by users at this stage.
Why read this
AI score 62
Ant Lingguang's CTO reviews the pivot from 1.0 to 2.0 and explains why the super-app ecosystem leads domestic personal agents down a different path.
Topics
Body · AI translation
By Silicon Valley 101
In the second half of 2026, global tech giants are launching a new battle for the AI entry point.
From Silicon Valley's Grok Bot, Muse, OpenAI Dots, and Anthropic Duet to the startup Instinct, AI products in the form of Personal Assistants (PA) are rolling out one after another.
After AI exploded on the enterprise side, the personal consumer side is a bigger market, and this is the next battlefield all the giants are joining. In China, the battle among PA-type products has also begun, but China's market and ecosystem will be very different from Silicon Valley's.
China has the giants' super app ecosystem, which means agents within the ecosystem can more easily connect payments, office work, travel, and other links. But in the mainland market, where AI product competition is fiercer, how such products can be commercialized is also an important challenge.
We spoke with Mo Linjian (nickname: Linyi), CTO of Ant Lingguang, which just released Lingguang 2.0, about his thinking on the evolution of Agent products.
A year ago, Lingguang 1.0's main feature was generating flash apps from a single sentence, letting ordinary people without programming skills create their own tools—a direction full of imagination. But after launch, the team gradually discovered that being able to create an app does not mean the app actually solves a problem, and creation itself may not be a high-frequency need.
So in 2.0, Lingguang tries to shift the product's focus from helping users create tools to helping users truly accomplish their goals. It is not just about answering a question once, but about remembering your needs, continuously tracking tasks, proactively reminding you at the right time, and even acting on your behalf when authorized.
At the same time, this battle over personal AI assistants also raises many questions worth discussing. Are users really willing to pay for an AI personal assistant? If AI can complete shopping for us, what happens to platforms that originally relied on traffic and advertising? As AI gains more and more personal information, how do we build trust and ensure it does not do the wrong thing?
Chen Qian: Linyi, welcome to Silicon Valley 101. I know you just released Lingguang 2.0, a year after Lingguang 1.0 was released. Why did you choose this moment to release 2.0?
Linyi: We had been brewing 2.0 for a long time. What left the deepest impression about 1.0 was probably creating flash apps, but we kept thinking about how to help users get things done better. Whether Q&A or apps, they often solve only one aspect of a problem, while tasks like travel span time periods.
We probably started brewing the related work around June or July this year, and releasing at this point is a natural result for us. It just so happens that many similar products have also appeared in the industry.
Chen Qian: So Lingguang 2.0's audience is still more to C, hoping to help the general public get things done and solve problems. Is this somewhat benchmarked against Muse?
I can talk about my own experience. I started using tools like Codex early on to manage life and handle office affairs. For example, to correct my child's wrong answers, I would photograph the child's mistakes, have it analyze what went wrong, and then create a new targeted question. If the child gets it wrong again, the process continues—a bit like manual RL (reinforcement learning). But this process is still complicated for ordinary people. My wife cannot use it either, and it is even harder to imagine my mother or other elderly people using it.
Linyi: Right, but I think similar capabilities can be generalized. Take my mother as an example. She has high blood pressure and needs to manage it. Right now she records her blood pressure in a small notebook every day.
But recording is only the first step. More critical is analyzing trends, and perhaps also giving advice: what should be done when blood pressure is high. These things are suitable for Codex or a similar agent to manage, but she simply cannot use it.
We see such powerful capabilities, yet my mother cannot use them. Does she need this capability? Definitely. We want to turn powerful capabilities into products that ordinary people can also use.
Chen Qian: Was this also inspired by OpenClaw, the "crayfish" from early this year, and the later rise of Hermes? At the time, ordinary people likewise could not use them. Many people were anxious, and even after installing "crayfish," they did not know how to use it.
Linyi: Yes, it was more of an inspiration. This interaction style is very meaningful as a reference for us, but our main line has never changed: we hope to help users better complete a task.
We hope to move from means such as looking up information and making apps to interacting directly around the user's goals. Users explain their goals and constraints clearly, and we proactively help them complete things along the way. This is the result of the product's natural upgrade, and in the process we also absorbed inspiration from these products.
Chen Qian: To build an agent product that everyone can use, what technological advances over the past year have made today's agent capabilities possible?
Linyi: Technically, the first is certainly model capability. Last year everyone may still have focused on programming and mathematical reasoning; now long-horizon task execution capability has made great progress. This is one of the conditions.
The second is the harness layer. It may not contain particularly profound technology, but it is very critical for agents to be deployed.
Models have their own limitations. Even as their capabilities grow stronger, some problems still cannot be overcome. For example, they are still probabilistic models and cannot guarantee 100% correctness; many things, including new developments, must be backstopped by engineering means. Because model training takes at least a month-long cycle, new developments also need to be supplemented through peripheral mechanisms such as memory.
In addition, the development of the ecosystem, MCP (Model Context Protocol), and so on are also pushing this forward. Cost is equally critical; otherwise, it cannot be sustained to accomplish so many things.
Chen Qian: These are all very important; we will expand on them later. First, let's review the birth of Lingguang 1.0: from developing 1.0 to 2.0, what lessons did you learn from market feedback?
Lin Yi: Our product is not about adding features to an existing business; rather, we hope to work close to the technological frontier. For example, we judge where technology may develop in six months or a year, and then build products around those capabilities.
When we started last year, a major trend was progress in coding capability. Could it help us achieve the goal of "helping people complete things"? A very natural direction is helping people write programs and build apps. Generating an app with one sentence has a low enough barrier that ordinary people can use it.
Chen Qian: Let me ask a challenging question. We saw data showing that after Lingguang 1.0 launched, its three-day retention rate was just over 20%, lower than that of chatbots. What thoughts did this bring you, and how did it affect the product design of 2.0?
Lin Yi: On the one hand, you have to look at these numbers; on the other hand, you still have to look at the core: are the apps that are truly useful and repeatedly used growing?
Frankly speaking, making flash apps is not a high-frequency need; many people prefer to use things directly. This is also why we later created Lingguang Circle: not everyone has that many ideas and wants to build apps themselves. A fairly big reflection is that we should focus on how much value the product ultimately brings to users. Making an app does not mean value has been created; only when it is truly used and solves users' problems does it have value.
By 2.0, we will continue this way of judging value. We will not look only at numbers or only at how many people created something, because there are too many other interfering factors in between. The core is still whether it helps users truly complete things and provides incremental value.
Chen Qian: I see that Muse's current retention rate may also be just over 20%. Is this number higher or lower than your expectations?
Lin Yi: If it were me, I would want to see whether those who have already created goals and use the product deeply are continuing to grow and retain.
Overall data will be affected by other factors; this is a bit like our flash apps at the time. If a product becomes explosively popular, it will inevitably bring in a large number of users. If everyone who comes stays, that would of course be ideal, but in reality there will be many disturbances.
So I would rather look at whether this group of users with three to five goals they truly care about is continuing to grow and retain. After we launch, we may also look at our own situation from this angle.
Chen Qian: I see its users should have exceeded 10 million recently. It is growing very fast, but the speed of new user growth is declining. Is this inevitable?
Lin Yi: We may encounter this situation, and it may also be related to the evolution of capabilities. For us, one thing is common: we should still return to the most core user value.
Chen Qian: Lingguang 1.0 launched a 100 million yuan flash app incentive program at the time. Do you think the money was well spent? If so, will 2.0 launch a corresponding program? If not, what lessons were learned?
Lin Yi: For 1.0, incentives definitely had value in cultivating habits. Letting more people understand it and make their first app of their own had market education significance; this point is fairly clear.
Second, what we incentivized was not just creation itself, but also making the product good. So the plan at the time was strongly tied to Lingguang Circle and closely integrated with it. Making something is the first step; after that, we hoped everyone would make good things, share them, gain recognition, and let more people continue modifying based on the results.
But truly making something excellent still has a relatively high barrier, and there is still room for improvement in comprehensive capabilities. Being able to make simple apps that solve nearby problems at hand is already good.
Chen Qian: Do you think general agents still need market education now? For example, Meta is also using incentives, giving both inviter and invitee free token quotas through invitation codes, using subsidies to achieve social fission. But ChatGPT did not adopt this kind of promotion method at the time. Why did chatbots not need it then, while agents need it now? Is it insufficient market education, or are capabilities still not enough for many people to use them spontaneously?
Lin Yi: I think there may be multiple factors. Meta certainly has considerations about competing for the market, and the timing is different from ChatGPT back then, when ChatGPT stood alone.
Chen Qian: There was no competition at that time.
Lin Yi: Right, the situation is different. Meta faces fierce competition, and using fission-based marketing is understandable; this is related to the market environment.
On the other hand, users have already developed the habit of using chatbots. Using a personal agent requires telling it your goals, not just asking a question. It is not asking "What is the weather like in southern Xinjiang?" but saying "I want to go to southern Xinjiang; help me arrange it." This habit needs to be cultivated so that users know they can use it this way. Although a shift is needed, I think the latter is more in line with human nature: users only need to state their purpose.
Chen Qian: Changing user habits is not easy. In the internet and mobile internet era, everyone is used to opening an app and completing a task through clicks. Now it has to change to telling the PA (personal agent) in natural language what you want to do, and then having it complete it automatically. Will this kind of interaction paradigm change bring a tenfold improvement in efficiency or convenience?
Lin Yi: I don't think it will be a replacement; the two should integrate better instead.
Chen Qian: You don't think the original paradigm will disappear?
Lin Yi: No, and both will get better. For example, GUI (graphical user interface), this way of operating, has great natural advantages. On the one hand, many people have already been educated on it. On the other hand, it encapsulates the operating steps. For example, for a complex task, the first step is to check the destination, and the second step is to select other conditions. Users understand it at a glance; the interface encapsulates the optimal process inside.
If it is pure conversation, it is of course great if users can say all the information completely in one sentence. But if it is too free, users may not know what to express next. So I think GUI is a very good form of encapsulation, because for repetitive processes, clicking may be more convenient than speaking; one click and it is done.
So I think it will not disappear, and the two will definitely integrate better. We have also done a lot of exploration in this area. When we were building 1.0, we adopted a combination of GUI and conversation, hoping to interact with users in the most appropriate way and present results in the way with the highest information efficiency. Sometimes a picture is worth a thousand words, and sometimes an interactive animation is more useful than a lot of text. The same is true for 2.0.
Chen Qian: Natural language is not necessarily the most efficient way to communicate, and many things cannot be expressed in language.
Lin Yi: Right, for example, if I ask you to describe what this space is like, how do you say it in one sentence?
Chen Qian: You might have to write an 800-word article.
Lin Yi: Taking a photo is actually best. So there will definitely be multiple input methods, including multimodal and unstructured input, while the output should be structured.
Chen Qian: What brings the most value from general agents? Taking Lingguang 2.0 as an example, what scenarios have you seen that users find most valuable?
Lin Yi: We have integrated some capabilities from the Alipay ecosystem. A friend asked it to analyze bills to see where money was wasted and where money could be saved. As a result, it found that the amount he refunded was five times higher than the amount he spent. For example, he bought six pieces of clothing, returned five, and kept one. It analyzed that he was an impulsive consumer, and he said it hit the mark.
I have also analyzed my own bills. Its suggestion was that my telecom plan was too extravagant. I took a look and indeed I did not need such a high allowance, so I downgraded the plan by one tier, which also counted as saving money. Although the example is small, I think it is valuable.
Also, during the summer vacation, I took my family to Xinjiang for a trip and did trip planning, and my personal experience was also very good. I told it the goals, itinerary, and constraints, such as accommodation requirements, and it kept an eye on things for us throughout the trip. The key is that it did not just give a travel plan before departure; the experience during the trip was even better. It would tell me how to adjust the next step according to changes in the situation. For example, the scenery in southern Xinjiang is greatly affected by the weather. When the weather is bad, the scenery cannot be seen, and it would help you adjust the itinerary accordingly.
Chen Qian: Give new suggestions based on real-time changes each day.
Lin Yi: Right. For example, for a period of time I had an upset stomach and felt unwell, and it suggested readjusting the itinerary. During the trip, it would update the overall arrangement according to the conditions we cared about and external changes. It would not simply push what was not completed yesterday to tomorrow, but replan according to the goals of the whole trip. This process is very worry-free.
Lin Yi: I have seen many that are shopping-related.
Chen Qian: Right. I will also share two examples: one is helping everyone sell things on Facebook's second-hand marketplace, Marketplace. The other is asking Muse to search the inbox for small refunds or checks that need to be claimed. There are too many emails, and users easily overlook them, but it can help find them, making everyone feel that they have indeed saved back a lot of money.
Lin Yi: If people are willing to spend the time, they can actually do these things, such as checking bills one by one. But often they either forget or find it troublesome, and the PA (personal agent) can do it for you.
Chen Qian: What do you think agents can actually do, and where are their high-value scenarios? Is it more valuable for agents to help you save money and time, or to help you spend money and time?
Lin Yi: I think they are not contradictory. Both things can be done now, and they are done quite well.
For example, in productivity and office software, the Office trio most commonly used by white-collar workers, making PPTs, making Excel sheets, and writing Word documents, were among the first batch of scenarios to take off. Products such as Claude Code and WorkBuddy are also oriented toward office work, and only later expanded to personal life scenarios.
These two scenarios are not contradictory and can both be done well, but cost and cost-effectiveness will be involved. Inference costs are still high now, and office scenarios are relatively easier to recover costs from.
Chen Qian: For enterprise users, the commercial value may be higher.
Lin Yi: But the market for pure office scenarios is certainly not as large as that for life scenarios. Basic technology is advancing and costs are falling, which is also one of the prerequisites for the product to be viable. Saving time has very clear value. Isn't it good to use the saved time better for life?
Chen Qian: Which scenarios where agents handle tasks will be relatively high-frequency? Only truly high-frequency and reproducible ones will have higher value and user stickiness.
Lin Yi: Office work is certainly high-frequency, and there are also many in life. Every day we may have things on our minds, such as child-related matters. When they are young, they need vaccinations; when they are older, their studies need to be managed.
Chen Qian: Check the wrong answers.
Lin Yi: Right. There's also managing finances, handling money, how to save money, how to make money, and going out to have fun—there are many scenarios.
And many things can't be solved in a single round of conversation; there are implicit goals behind them. Having me help my child manage homework is to help him learn better. To achieve that goal, there are many things that can be done in between.
Chen Qian: Will it become everyone's personal secretary, making work simpler?
Lin Yi: It feels that way. Through getting used to each other, the more it understands me, the more it can handle those tedious matters for me. But things that truly require my decision should still be told to me. When it doesn't need to disturb me, it should proactively complete them—that sense of boundaries must be there.
Chen Qian: Will users pay for such a personal secretary?
Lin Yi: It depends on what incremental value it creates. For example, if it can really raise my child's score by 10 points, why wouldn't I pay? I've already spent so much money on education. So it still depends on what it creates; I think the demand to pay definitely exists.
Chen Qian: There's a business judgment about whether a product has found users' "pain point" or "itch point." An itch point is optional—nice to have, but users will compare prices and free competitors. A pain point is something people are genuinely willing to pay for. Now that agents help everyone improve efficiency and handle life affairs, is that a pain point or an itch point?
Lin Yi: That depends on capability progress and the specific scenario. Education, just mentioned, is definitely a pain point, but it still depends on how much incremental value we can provide. It's not that having it solves all problems—that's certainly impossible. But it can indeed greatly improve efficiency in some steps, much better than ineffective repetitive drilling.
Some scenarios can be called itch points. For example, we often use it to summarize DingTalk messages. Originally you had to go through group after group; now it automatically summarizes for you every day, greatly improving efficiency. You can't say it's absolutely indispensable, but it really gives me time to do other things.
Chen Qian: On commercialization, discussions in Silicon Valley have also been heated recently. Instinct has already said it won't charge users and is currently burning venture capital money. Whether it will use advertising in the future or take a certain percentage commission from the money it saves users is still uncertain.
Big companies may have another logic. For example, Muse may not need to directly make money for Meta; it may ultimately improve its advertising system and generate revenue through improved advertising efficiency. Grok Bot may target enterprise users, charging by token or in other ways.
Do you think domestic commercial and product competition and users' payment habits will differ from international markets?
Lin Yi: Frankly, it's hard to judge, but many underlying logics should be similar—it's just a matter of charging individual consumers or charging enterprises.
Right now everyone hasn't reached the stage where PA (personal agent) commercialization is very successful. The enterprise market already has many successful cases, but the personal market may not yet.
As for charging models, I personally lean toward membership fees and the commission mentioned earlier. Whether the advertising model is compatible with PA—I think it may be difficult. PA is quite different from chatbots or other tools. Tools still lean toward platform and traffic logic, while PA is like a very intimate partner. If commercial interests are mixed into interacting with you, I think it's very hard.
Chen Qian: If it plays you two ads first thing in the morning, the experience might be completely ruined.
Lin Yi: PA (personal agent) must earn users' trust; mixing in commercial interests will make this difficult. There may be innovations, but I think it's quite hard.
Chen Qian: Then how does it commercialize? Does that mean it can't commercialize?
Lin Yi: That's not what I mean; I'm talking about the advertising model. Membership fees and commissions can both be tried, and we may try them at an appropriate time, but not necessarily right now.
The core is still whether it brings incremental value. First make the cake bigger; how everyone divides it is the next step. If it truly solves a pain point, makes users' lives better, and creates extra benefits, it will definitely get the share it deserves.
Chen Qian: I ask this because I recently saw some investors believe that PA in China is not a game for small players, but a game for big players.
Domestic product competition is very intense. Big companies have resources and may also do better in interactions across various aspects, so small players don't have many opportunities, and it's not easy to achieve commercialization from the start.
Lin Yi: Yes. Now may not be the best stage to immediately implement commercialization. We should consider it, but not necessarily push it aggressively.
Entrepreneurs have another difficulty, related to the boundaries of PA's capabilities. Models now have strong action capabilities and long-horizon planning capabilities; the key is what environment you give it. In a virtual world, or a controlled environment, it can only do those things. If you only give it a virtual machine, what it can do is limited to this virtual world. But if it connects to real services and the real world, it can do more. It's not that the model's capabilities are lacking; the environment limits it. The ecosystem capabilities that big platforms have accumulated in the past will be of great help in this regard.
Chen Qian: This is also the focus of my next question: the capability boundaries and access permissions of agents. When designing Lingguang 2.0, will you refer to the overall ecosystem of Ant or Alibaba and then decide the product form or flagship use cases?
Lin Yi: We will definitely consider it. The first bill example earlier indeed can only be done by relying on an ecosystem like Alipay.
Products from different companies may be similar in broad concepts, but the core value they provide to customers based on their own ecosystems may differ. For example, around bills and payments, many money-saving-related services can be derived.
Chen Qian: What advantages does the Ant and Alibaba ecosystem give Lingguang? Delivery is a very high-frequency occurrence, and so is travel.
Lin Yi: Right. The second question is what attitude and position service providers take in this change. If they have their own platform and invoke their own capabilities, the same problem does not exist.
Chen Qian: For example, we saw Amazon shut Muse out. What do you think the motive and the conflict are? As the platform, Amazon, along with users, and Meta as the agent provider—what conflict exists among the three?
Lin Yi: The biggest reason is probably still commercial interest. Amazon's business model relies heavily on traffic and advertising.
Under the logic of a PA (personal agent), users may be able to complete transactions directly without visiting the platform, so that part of the value will certainly be hit hard, and there will be a clash of interests.
Chen Qian: Between Lingguang and Alibaba, for example Taobao, is there no such conflict?
Lin Yi: There is. It will not directly replace the platform; for a long time, the two user habits will certainly coexist. I don't think it will ever get to the point where the agent is the only entry point and people no longer visit platforms at all. But within a large ecosystem, collaboration is indeed much easier than with the outside, especially inside Ant.
Chen Qian: Is there a way to align the interests of all three?
Lin Yi: I think the game will continue for a long time, but the core is whether this approach creates incremental value. Without increment, it is a zero-sum game; with increment, there is new room for interest coordination.
For example, buying coffee is essential for some people, but not for me. When I am driving, I originally could not go buy it, but if I say a word when I want to drink it, it can buy it for me—such scenarios exist. This is not just a capability issue; it also involves a series of things such as payment and trust, and the industry should explore it together.
Chen Qian: How is the interface for Lingguang 2.0 designed? What can be connected now?
Lin Yi: Huabei bills, and many externally open services, such as email, DingTalk, and Feishu. Interfaces come in various forms. If a merchant opens MCP (Model Context Protocol), we connect according to that protocol; some use A2A (Agent-to-Agent protocol).
Chen Qian: What stage have these two protocols reached now?
Lin Yi: The protocols themselves may not be the bottleneck; they have already become de facto standards in some cases. For stateless queries or relatively simple queries, an interface is enough. But some queries carry business logic behind them, such as connecting to internal credit business. If it is just asking "how much is my Huabei bill," one sentence and one interface are enough. Sometimes users ask "why is my credit limit so low" or "how can I raise my limit," and there is business logic behind that. It is best answered by the credit business's agent, because the outside may not accurately grasp it; in this case, A2A (Agent-to-Agent protocol) may be more appropriate.
Chen Qian: So neither MCP nor A2A is absolutely better; it depends on the situation. If the other party does not grant permission, is Browser Use the only solution?
Lin Yi: Browser Use may not necessarily solve the problem either; it depends on the reason for not granting permission. For example, if Amazon explicitly does not allow you in and you force your way in through Browser Use, it becomes confrontation, and there will not be a good result.
There is a case where the other party is willing to open up but is limited by conditions. For example, small merchants may lack technical capability or the extra budget to develop interfaces; in this case, Browser Use may be a good approach, because they are willing to open up in the first place. But if the other party simply does not let you use it, then it is not just a technical issue. Browser Use certainly has technical difficulty, but the bigger problem may lie outside technology.
Chen Qian: If the two sides of this conflict cannot coordinate and reach agreement, will it become a bottleneck for agent development?
Lin Yi: Yes.
Chen Qian: So the bottleneck is not just technology itself; all parties also need to coordinate interests first.
Lin Yi: Right. Many complex problems are not purely technical problems.
Chen Qian: What industry trends do you see? Is everyone converging and starting to cooperate, or is it still in a confrontation stage?
Lin Yi: It is very complex; everyone is generally experimenting, and it is hard to say what the future will definitely be like. Take Amazon again: attitudes differ among parties, and each has its own interests.
But some things I think will definitely happen: users will not go in one step to "only interacting with agents from now on," but the agent will definitely become a default option.
Chen Qian: Once people try it and find it useful, it is hard to go back. A colleague of ours was selling furniture two months ago. Originally he had to take photos, upload them, and write descriptions himself on Facebook Marketplace; now Muse can sell it for him directly, and the time and effort involved are much less. After getting used to this way, he will not want to do it himself anymore.
Lin Yi: Yes, the trend of this habit change is relatively clear. But under this trend, different merchants and brands will have different attitudes.
High-margin merchants may be more willing to embrace it. For example, some brands have thick profits and may not be afraid of price comparison; instead, they worry that the agent will say the wrong thing on their behalf, because the way it answers is uncontrollable and may affect brand image. Rather than letting others speak for me, I would rather connect proactively myself.
On the other hand, agents naturally compare prices, and pricing systems may be very complex, with different prices across channels. Some merchants will be very worried, especially low-margin ones.
Chen Qian: It will strip away part of the middleman arbitrage space.
Lin Yi: Right, so small merchants may have a harder time. But they can't avoid embracing it either, otherwise after traffic shifts, users may no longer be able to find them. These problems are quite difficult to solve.
Chen Qian: You just mentioned the conflicts of interest at large platforms, such as Amazon refusing Muse access, which is also because of the existing business model and advertising revenue. Advertisers pay to gain ranking; once agents enter, could this part be disrupted?
Lin Yi: Yes, right.
Chen Qian: Then the only way is that it may need to change its business model.
Lin Yi: I think it will definitely try to do similar things itself. The U.S. ecosystem has many standardized parts and is relatively open overall.
Domestically, there are basically some super ecosystems, forming closed loops from payment to shopping, and there are several such closed loops. Abroad, Amazon may be a fairly typical representative, while other parts are relatively fragmented and more open overall.
Chen Qian: Is it possible that domestic agents will land earlier because these ecosystems already exist within super apps?
Lin Yi: There are enough scenarios and space, so I think it may be possible.
Chen Qian: Let's talk about security. AI will always hallucinate, and agents may make mistakes and do unexpected things. How have you designed security aspects?
Lin Yi: I think this is a matter of life and death; it may even affect the development of the entire industry. As mentioned earlier, PA (personal agent) will become one of users' default habits, but some conditions may change this trend. For example, if a large-scale malicious incident really occurs and users lose trust, this will definitely move backward. So it is crucial. We have done a series of work, and we must continue doing it.
On the one hand, Ant has deep accumulation in security. For example, we work with the internal security team's APASS platform to provide some security protection; users' accounts and passwords are isolated from the agent.
On the other hand, for things that may be done wrong, instructions and execution actions need to be graded, with different levels triggering different permissions. For example, for payments, the simplest way at this stage is for the user to confirm everything. We are also exploring other ways. Ant Group as a whole is working on "AI Pay": when certain conditions such as a certain amount and scenario are met, it can help users pay automatically, which requires systematic coordination.
At present, we still lean more toward information queries and have not yet executed risky actions such as payments on a large scale. Before the infrastructure is complete, we really need to be especially careful in this area.
Chen Qian: Right now the final decision-making power is still in human hands. For example, whether to pay still needs to be decided by a person. Under what conditions can machines start making payments? For some small payments, can machines pay automatically?
Chen Qian: Alipay should have a lot of experience.
Lin Yi: Quite a lot of accumulation. But doing it at scale still requires special caution. At the same time, we also have some confidence, because Ant Group has accumulation in this area.
Chen Qian: You mentioned earlier that technical optimization over the past year has reduced token consumption costs, which is also an important progress. Specifically, what technologies have improved token usage efficiency?
Lin Yi: The biggest driver is certainly progress in model technology. On some key leaderboards, a 100B model now may score higher than a 1T model did in the past. This is huge progress, and it is still continuing. Models can work and do things while consumption drops significantly. I think this is the biggest driver.
There are also changes in product form. For example, in PAs (personal agents) such as Muse and Lingguang, many tasks are executed in the background. Background tasks can be scheduled, which means they do not have to be crowded into peak periods; they can shave peaks and fill valleys, execute at more suitable times, and use hardware inference resources more efficiently. These changes may all reduce costs.
Chen Qian: Are the largest models called behind the scenes for everything? Or can small models also complete some tasks?
Lin Yi: Quite a few things can be completed by small models; not everything requires the largest model.
Chen Qian: So underneath there are large models and small models, and different models are scheduled according to the task?
Lin Yi: Right, theoretically this can be done. But we still have a lot of room for optimization in this area; it is not that we are already doing it very well.
Chen Qian: Is the main room for optimization next mainly in scheduling?
Lin Yi: I think it is mainly the models; scheduling can be optimized in parallel. Models will continue to improve, and we have already seen great progress on Bailing. For example, on public leaderboards such as AA, models of the same size can improve very quickly in just a few months compared with previous scores.
Chen Qian: Models are indeed progressing very fast now, and with RSI (recursive self-improvement) next, it may be even faster. What scenarios that AI cannot do now do you expect will become feasible in the next six months or a year as models improve?
Lin Yi: There is one thing I very much look forward to, but it is still quite difficult now: making PA (personal agent) get better the more it is used.
Lin Yi: "Understanding you" is more reflected in memory, that is, knowing you better, which can be solved through context.
But "good" is not only reflected in memory; it may also be reflected in execution paths. For example, when helping you buy a plane ticket, knowing your preferences is understanding you better. But search paths and execution paths also have room for optimization. Suppose it checks ten platforms or ten airlines; can it save the best experience among them and complete the task more efficiently next time? This concerns both cost and efficiency, as well as effectiveness.
Complex long-horizon tasks cannot be guaranteed to succeed every time; sometimes search fails, and sometimes it times out. After accumulating experience, the success rate may also improve. Similar concepts have existed for a long time, and Hermes also has a practice of automatically summarizing skills. But it cannot be implemented arbitrarily, because it may produce side effects.
Lin Yi: For example, if I say "I'm not eating spicy food today," it summarizes it as "this person never eats spicy food." This boundary is very hard to grasp.
Chen Qian: It categorized you incorrectly and failed to identify your intent correctly.
Lin Yi: Right. Just because one execution path succeeded this time, does that mean it will definitely be the best going forward? Not necessarily.
Chen Qian: And I might change too—today I don't want spicy food, tomorrow I might want it again.
Lin Yi: Right, preferences are one thing, and execution capability is another.
It's not that it's completely impossible right now. On the matter of "understanding you," memory can play a big role. But to fully automatically optimize the user's execution trajectory and distill that optimization into system-level capability, so that it gets better the more it's used—there are still many conditions that need to be handled.
This closed loop involves not only memory but also skills and models. Which experiences should the model absorb, and which should it not? Many things can work in concept, but maintaining continuous improvement in real operation is still very difficult, and the noise is very large.
Chen Qian: Do you think it will gradually be realized in the coming months?
Lin Yi: We hope so, but it may not be solved immediately in the coming months. I think it will keep improving, so that users feel it gets better the more they use it.
Chen Qian: The issues of long-term memory and long context mentioned earlier—have they been solved now?
Lin Yi: We definitely can't say they've been solved; that example just now wasn't easy either. I think memory can be broadly divided into three categories.
First is long-term preferences, which are relatively stable. For example, liking a certain hotel brand, or long-term habits.
Second is facts—things that actually happened. For example, during a trip, today I went to three places and didn't go to the fourth—that's a fact.
Third is temporary matters, such as "I'm not eating spicy food today."
These three need to be distinguished. When should something be classified into which category? When can a temporary matter rise to a long-term preference? All of these need to be handled separately.
Compared with a chatbot, a PA (personal agent) has certain advantages in this regard; some information can be better used for memory processing. Take Lingguang, for example: it continuously helps you manage a certain matter, and there are many interactions and processes in between, so it can better accumulate fact-level memories—that is, what actually happened.
In a chatbot, the fact that I looked up a certain brand of car doesn't mean I'm definitely going to buy it; I just looked it up. But if I set a goal: "I want to save money to buy a certain car," that is a fact explicitly set by the user, and it can also help understand related queries afterward.
Putting temporary, one-off conversations into a continuous thread is helpful for memory processing. Long-term preferences can be extracted from things that actually happened, or answers given to the user can be adjusted accordingly—compared with one-off conversations, there is more basis for it.
Chen Qian: This involves another dimension: the proactiveness and passiveness of AI. A chatbot responds passively, while a PA (personal agent) can be proactive and sometimes even nudge you.
Lin Yi: Right, that degree needs to be grasped well.
Chen Qian: It will remind you what needs to be completed today and tomorrow in order to achieve long-term goals. It can be more proactive, but how should that scale be designed?
Lin Yi: That's a very good question, and we're exploring it too. I think information can be divided into three layers.
First, matters directly related to action goals and on the main path. For example, during a trip, "if you don't get up now, you'll miss the sunrise." At that point a reminder is appropriate; otherwise, missing it would be its failure. This type of matter must be handled well.
Second, information that has some probability, or a relatively high probability, of changing the user's actions—I think it should also be appropriately delivered to the user. For example, it knows you love coffee, there's a highly rated coffee shop near the destination, and you've previously shown willingness to change your itinerary for a cup of coffee. It's not on the main line, but it's very likely to change your actions, so this kind of information should also be told to you.
Third, things like introductions to the destination. This kind of information is more like a secretary placing a document beside you—you can look if you want, and you don't have to.
Therefore, information needs to be layered, and the interaction methods may also differ. The first type should be strongly reminded; the second type might be reminded in conversation or presented in a dedicated spot. The third type can be placed in an additional location for the user to view, but at the very least it should not be actively pushed as a reminder. Lingguang is still rather conservative now and doesn't proactively disturb users that much.
Chen Qian: In the future, could different modes be offered? For example, I want a proactive assistant that reminds me not to be too lazy and to do more things every day.
Lin Yi: Right, there's personalization in this too.
Chen Qian: It could be PAs (personal agents) with different personalities and characters.
Lin Yi: Right, and a person's personality will affect information layering. The same good coffee shop—if you love coffee, it might belong to the second type for you, but for me it might not. It depends on how much impact it will have on you.
Chen Qian: This is also the incremental value you mentioned at the beginning. If I didn't know this shop existed, I wouldn't go spend money there. Knowing it, and wanting to try it, brings increment.
There are also things of the "I don't know what I don't know" kind. If a PA (personal agent) tells me that this thing exists and asks whether I want to consider it, it provides increment in cognition, information, and choices, and in the end it may also bring consumption increment.
Lin Yi: Right, possibly. We're mainly doing the first type now, and trying to explore the second type.
Chen Qian: I think it's very valuable. I used an app made by a friend's startup; its agent connects newly met people with existing contacts.
For example, if I want to meet someone, it will tell me the other person's background, how to introduce myself, and it roughly knows what we've talked about and where I've been recently. Once when I arrived in Beijing, it proactively told me that a certain person might also be in Beijing recently and suggested we meet. Or it knows I'm about to go to the airport and have a two-hour gap in between, so it suggests using that time to meet someone. This is very helpful for my network and social connections.
Lin Yi: Yes, it can bring a great sense of surprise. The second type mentioned earlier is something we definitely should do.
Chen Qian: So the truly valuable opportunities in the future may lie in the second type?
Lin Yi: I think both the second and third types are valuable, but you have to judge: after the user knows this information, how likely is it that their actions will change?
If the probability of change is low, it may belong to the third type—nice to know, but no big deal if you don't. The second type, on the other hand, has a greater chance of influencing action, which is the increment and surprise you mentioned.
Chen Qian: What prerequisites are needed to do the second type well?
Lin Yi: First, you definitely have to understand the user, because the same piece of information may change one person's actions but not another's. Whether you call it memory or "understanding this person," that's the prerequisite.
Second, you need enough information. I first have to know there's a coffee shop there, and that it's well reviewed. Once you know that, there are many technical means you can use.
Chen Qian: Are most agents still stuck at the first type right now?
Lin Yi: I don't think you can put it that way. The first type is definitely something most agents must do; without it, it's hard to call it an agent. But different products have different strategies—some are very proactive, to the point that you find them verbose.
Chen Qian: Nudging too much, even ordering me to go meet people.
Lin Yi: Right, it may be treating some third-type information as first- or second-type. Different products make different choices; overall we may be too conservative right now and should more proactively explore incremental value.
Chen Qian: Right now Lingguang is mainly doing the first type, and will develop toward the second type in the future?
Lin Yi: Yes, right now it's the first type plus a small amount of the second type, and we hope to explore more.
Chen Qian: Are there any other interesting real cases from the internal beta of Lingguang 2.0?
Lin Yi: I'm deeply dependent on it now—I've really entrusted some things in my life to it.
Chen Qian: Deeply dependent means you open it many times a day?
Lin Yi: Yes, and usage time is relatively high. Take the Xinjiang trip as an example—I might use it for one or two hours.
Before, if I wasn't deliberately testing, I might interact only a few times or a dozen times a day, not for that long. But when you treat it as a PA (personal agent) rather than just an information query tool, I have a lot I want to tell it. For example, if I feel unwell or my itinerary changes, it will help me plan. More importantly, the more it understands me, the better help it can provide.
So at the end of each day, my child is also very willing to tell it what they liked and disliked today, and proactively write a travel diary entry. This information can help with the next trip planning and let it understand us better. Only when you see the value are you willing to spend time. If the interaction is more interesting, the child is more willing to do it, and usage time naturally increases.
Also exercise—since I like running, for example if I want to run a half marathon, on one hand it can make a training plan, arranging how many kilometers to run today and strength training tomorrow. On the other hand is connectors. Now it can connect to a Garmin watch, a body fat scale, and strength training apps. This data is sent back regularly every day; I don't need to tell it how long I ran today—it can summarize it and then arrange the next day's plan, which is very convenient.
Sometimes when I open it, I'll report that I'm too tired today and can't run. Other times I open it to check the day's schedule; before, I had to actively query "what should I do today," but now it pushes it to me, and I just run according to the schedule.
Entrusting the goals you truly care about to it and using it as a PA (personal agent)—that's what deep use means. It also makes mistakes; tell it where it went wrong and it will correct itself. Used this way, I think stickiness is very high.
Chen Qian: The more you use it, the better it understands you.
Lin Yi: Yes, and the value gets higher too. But we also have to solve how to let users use it better. It's not as direct as a conversation where you ask a question and get an answer; it requires a process.
Chen Qian: Will the final form of a PA (personal agent) have an always-on form? It feels like that's the only way to understand you better, but there are also many technical challenges.
Lin Yi: It doesn't necessarily have to be truly always-on. For example, if we really had an assistant, we wouldn't need him to follow you and watch you 24 hours a day.
Chen Qian: But a good assistant would be great if available 24 hours a day.
Lin Yi: I think being there when you need it most is enough. Take fitness—the key moments in a day are just a few time points: telling you the training plan, analyzing results afterward, and giving suggestions for the next day. When there's a problem, it can also explain clearly. Do these, and it's qualified for this matter.
Going further means doing more second-type things, such as providing better training suggestions. Users feeling it's always online doesn't mean the backend must run constantly. It can wake up at the right time and get things done well.
Chen Qian: If you want to achieve always-on, what technical milestones are needed behind it?
Lin Yi: Still the same view—always-on as users perceive it and always-on as technically implemented are two different things. Keeping a PA (personal agent) running 24 hours a day is mainly not about technology, but about cost, and what incremental value it actually creates.
Chen Qian: Still a balance between cost and incremental value.
Lin Yi: Right. A lot of the time it's just running idle with no increment, so the value isn't that great.
Chen Qian: Are you optimistic about agent phones? If it's just an app, you always have to unlock the phone and open the app to use it. On-device hardware can better accompany users, receive information, and form an always-on PA (personal agent). Will that be the final form?
Lin Yi: I think it will develop in that direction. For example, when driving, there are times when you want to ask a few questions or chat a bit. But unlocking the phone and then opening an app isn't the best choice.
If there are hardware entry points like car systems and glasses, the PA can be everywhere. I think this is a trend, because the need exists. When I'm driving myself, sometimes I want to pull out my phone and ask a few questions, but it's a bit dangerous.
Chen Qian: Right, I do that too. Because I drive a Tesla, when using FSD, sometimes I want to pick up my phone and do things. But if my eyes aren't looking ahead, the in-car camera will flag it as a violation.
Lin Yi: So the need exists, it just hasn't been well met yet. I think it will definitely be met through earbuds, car systems, and other means.
Chen Qian: Let's imagine—what will Lingguang 3.0 look like?
Lin Yi: I wouldn't dare say what 3.0 will look like, but there are a few directions that are fairly clear and will keep advancing.
One is the ubiquitous PA (personal agent), which can be invoked at any time in the most convenient way and interacted with in the most natural way. Right now a lot of it is still voice and text, and it also supports images; real-time multimodal interaction and so on can all be explored, but the technical challenges are enormous.
The second is connecting to a broader physical world. Right now it's still the virtual world plus some services, and this scope will definitely keep expanding.
And there's the layer in between: the PA has to be able to complete tasks very reliably. When evaluating a PA, the core question is whether it can truly complete the tasks users hand to it—reliability is the most fundamental requirement.
Chen Qian: Connecting to the physical world—what specifically does that mean?
Lin Yi: Through interfaces, for example connecting to smart home devices and robot vacuums. Many home devices already provide these kinds of interfaces, so connecting is very natural—the conditions are already in place.
Chen Qian: I think that's very much needed. For example, a Tesla can have its interior temperature set in advance. In the US, sometimes it's over 30 degrees outside and maybe over 40 in the car, but I often forget to turn on the AC in advance. If the agent knows I'm about to get back in the car and will arrive in about three minutes, and connects to the Tesla app in advance, it could automatically lower the temperature for me. That's a feature I'd be willing to pay for.
Lin Yi: The benefit is that it can access a broader range of information. The Tesla app can already let you turn on the AC in advance.
Chen Qian: But it isn't proactive, because it doesn't know when I'll be back.
Lin Yi: Right, that's the key—the PA has access to a broader range of information. For example, accompanying a child to a one-and-a-half-hour tutoring class—I don't know if you do that now; it's fairly common in China.
Chen Qian: My child is a year and a half old now and hasn't started yet, but maybe in the future.
Lin Yi: It knows the tutoring class lasts an hour and a half, and it might also know when I enter the parking lot. At least in China, smart seamless parking information gets pushed to Alipay.
If it knows the parking time, the class duration, and also knows the outside temperature is over 30 degrees, these things might not need you to say anything—it can coordinate and turn on the AC in advance at the right time. If it's cool that day, there's no need to turn it on. How far in advance specifically also depends on the arrival time. All of this can be imagined in theory, and I think it's not far off.
Chen Qian: Two final, more personal questions. I've talked with many agent entrepreneurs, and they're all very excited, feeling they're creating products that make human life more convenient and efficient. Do you have similar feelings? What moments make you feel you're building an exciting product?
Lin Yi: Yes, and the feeling is strong. Some of the earlier examples are like that—for example, using it to work through wrong answers, my child's math grades really did improve.
Lin Yi: Right, it's really quite good, and the key is it saves time. You might have achieved the same by grinding through practice problems, but it helps you in a targeted way. That was the first thing that made me intuitively feel it had value.
Also when traveling, my child proactively said, "Tonight I haven't told it my travel diary yet," and I felt that was valuable too.
And there was another small moment: my child wanted to buy a gift for my wife and asked it what to buy. It knew we were going to Japan for National Day, first gave some conventional suggestions, and then said, "You're going to Japan for National Day—why not pick something out for Mom on site."
At that moment, I felt this product was different from before. It wasn't applying a fixed formula to answer directly; it really knew your itinerary and suggested you pick out a bracelet for Mom on site in Japan.
Chen Qian: Very personalized.
Lin Yi: These moments make me feel it's useful, and I even hope my mom can use it too. If one day she tells me this thing is very useful to her, I'll feel even more accomplished.
Chen Qian: I think elderly people will need it very much. I chatted with another interviewee before, and he was very worried about global aging. The children are all working and under a lot of life pressure, and many elderly people need companionship—in seeing doctors, coping with loneliness, and daily socializing, they all need more help, and the current help isn't enough.
If there's a product that can really help elderly people order takeout, track health indicators, or offer small suggestions for daily life, I think that would be very meaningful.
Lin Yi: Yes. I hope to help ordinary people live better lives, hand the trivial things over to Lingguang, and let everyone have more time to live.
Chen Qian: And I hope our agent products keep getting better and help more people. Thank you very much for your time.
Lin Yi: Thank you.
That's our conversation with Lin Yi. After finishing this episode, we'll find that a very obvious trend right now is that AI is becoming increasingly "proactive" and "capable of getting things done." But for personal AI assistants to truly enter our lives, there are still several thresholds that need to be resolved:
Behind these key questions is actually a logic that connects technology and commercialization: the ultimate competition among personal AI assistants may not be about who has the strongest model, but about who can earn users' trust, connect to enough real services, and continuously create measurable value.
AI Trends