A Personal Agent Founder’s Reflections | A Conversation with 子皓, Founder of Ambi

Summary

• 子皓 believes personal agents have the ingredients to become consumer products, but the distribution loop is still broken. He doesn’t think the opportunity faded after OpenClaw’s moment in the spotlight. OpenClaw’s problems were that its use cases were hard to explain, setup was costly, and users still had to connect their own connectors: “This isn’t a technology problem; it’s a product and distribution problem.” In his view, Muse already has many of the ingredients of a product for ordinary consumers. Business Insider reported that Mark Zuckerberg and GitHub CEO Fred Freeman were surprised that the big tech companies hadn’t shipped something like it first.
• The underlying variable behind demand is the cost of accessing intelligence; interaction is his biggest non-consensus opportunity. “If you can just say what you need and it does it for you, maybe everyone will use it a lot.” He compares it to an extended-range electric vehicle: you might not need the range extender for 100 days, but still worry, “What if I need it once and it really hurts?” When the cost of acting approaches zero, demand in everyday life expands. That’s why, he says, “it’s inherently wrong for interaction to happen on a screen.”
• He sees ChatGPT and Claude among the eventual competitors, while distribution is a major constraint for independent products. He believes the opportunity for these products in China is limited in part because reliance on big tech distribution leaves startups exposed to upstream cost and channel constraints. Offline hardware in overseas markets faces fewer of the same closed social-distribution constraints. Ambi is betting on offline context and a new interaction point, with AI “present alongside you” rather than keeping people in the loop. Proactivity should be triggered by changes in the environment, not by repeatedly scanning connectors and guessing what users want.
• Models are still improving, but training is mainly about Pareto trade-offs; on-device LLMs don’t fit Ambi’s form factor. He has revised last year’s claim that “intelligence was already good enough”: after model upgrades, adjusting the harness and system prompt still produces clear gains on the same tasks. Model training is mostly about finding a better balance among cost, speed, and intelligence. Ambi isn’t building an on-device LLM, mainly because of battery, weight, and wearability constraints. Computer Use extends existing model capabilities, and he expects it to improve further over the next 3 to 6 months.
• He believes personal AI can develop network effects between agents, and that the timing is right for consumer applications. He cites a colleague @-ing his Hermes bot in a Lark group to ask it to review a document: products don’t need A2A to come first. Users build up personal content such as memory and a file system first, creating the basis for agents to collaborate. Once digital-task capabilities and infrastructure are mature, products can be built for a much broader audience.
• His business playbook has stayed consistent: timing and sector matter, and business ultimately comes down to efficiency, ROI, and demand. From brokering ice-skating tickets in high school, to analyzing user journeys at Qingsongchou, to using red flags as sales incentives for a Hainan telemarketing team, he has focused on identifying demand and finding ways to meet it. “Recognize demand externally, hire the right people internally”: the gap between the two is the residual profit. His enduring advice is to “go where the fish are” (去有鱼的地方抓鱼).
• He is pessimistic about China’s product environment, though the reasons remain an open question. He says few products from Chinese tech giants or startups have stood out this year: “I think Granola is pretty average, but there aren’t any products made by Chinese people that are better than Granola.” On the US market, 曲凯 brought up Harvey’s valuation of more than $10B, and 子皓 agreed that Tom and Instinct are good products. He says he still doesn’t know why things have turned out this way in China.

Deep dive

1. Ambi started with a personal AI architecture: a virtual machine for every user

• 子皓 confirms that the technical architecture from the company’s earliest days was “a virtual machine environment for every person.” It was only after investors raised the idea of personal AI 2 or 3 months ago that the team thought they might belong to the category, even though “we never defined ourselves that way.”
• The starting point was to change how people interact with AI: “Intelligence is already so powerful, but each of us might use it for only 2 or 3 hours a day. That’s the biggest irrationality in the world.” Intelligence had become broadly accessible, but people still weren’t using it like water, electricity, and gas.
• The ambition is not to build a vertical product. The goal is to create something many people can use, and that changes life for most of them.

2. Personal AI is more than a passing trend: OpenClaw failed on distribution, not technology

• 子皓 doesn’t think the opportunity disappeared after OpenClaw’s moment in the spotlight. The team deployed OpenClaw, found it wasn’t good to use, and switched to Hermes. They have since used Hermes for work, life, and health, as well as scheduled tasks and proactive services. He says that pattern has stayed the same “from the beginning of the year until now.” 曲凯 adds that when 子皓 says it “isn’t hot,” he means from a consumer perspective.
• 子皓 concedes that there may not yet have been a product with consumer PMF. In his view, OpenClaw was difficult to explain in terms of clear use cases and value; it was expensive for third parties to install; and, after installation, users still had to connect connectors and add context. “That’s not a technology problem. It’s a product and distribution problem.”
• Business Insider reported that Zuckerberg and Fred Freeman, whom 子皓 identifies as GitHub’s CEO, were both surprised that the other big tech companies hadn’t launched first, while Muse had. 子皓 thinks Muse and Grok have many of the ingredients for products aimed at ordinary consumers: no technical background is needed, and users don’t even need to know what a virtual machine is. “It really feels like an internet product from the previous era.”

3. Taking back “intelligence is already good enough”: model upgrades create new aha moments

• 子皓 walks back his earlier claim: “When I said last year that intelligence was already good enough, I may have been wrong.” At the time, the tasks they designed weren’t complex enough, there wasn’t enough context, and the sources weren’t varied enough. They also hadn’t constrained the agent’s scope to everything that could be done in the virtual world.
• His process is to adjust the harness after a model ships, then tune the system prompt to make it work better with the new model. The company’s Slack, GitHub, Figma, and its chat, files, and meeting content are all available. Even on an ambiguous task like identifying how a feature has deviated from its original design, Model 6 showed a clear improvement.
• Given the same task, Model 6 can say more with fewer words and greater precision, and sometimes surfaces something people wouldn’t have summarized themselves. 子皓’s conclusion is qualified: “The upper bound on people’s need for intelligence may really be very high. I can’t say it’s infinite, but I haven’t reached that point.”

4. Computer Use extends model capabilities; cost isn’t a major concern

• 子皓 divides product bets into 2 types: experiences that become possible just as model performance reaches the required level, and capabilities built by hand that work today but that models may be able to deliver natively later. He puts Computer Use in the second category: it may still be slow and require repeated attempts and exception handling, but he expects it to improve over the next 3 to 6 months.
• Ambi isn’t building Computer Use for now. The plan is to treat it as a foundational capability and use it once models have the necessary permissions and are mature enough.
• On cost, 子皓 says one feature of personal agents is that users don’t care much which model is doing the work. Tasks like booking a restaurant or buying a ticket require reliability, speed, and quality, but users can tolerate a little more latency for most tasks. Ambi doesn’t want to compromise the experience to cut costs. He believes the current trade-off between model prices and the value delivered can work, at least as a business model.

5. The cost of accessing intelligence determines how much people need it

• In response to 曲凯’s question about how many everyday needs people really have, 子皓 brings the discussion down to the cost of taking action. If “you can just say what you need and it does it for you,” people might use it a lot. But today, they have to set things up and spend time explaining, while everyday tasks are often trivial, so they don’t bother.
• He compares it to range anxiety and extended-range electric vehicles: you may not use the range extender once in 100 days, but still wonder, “What if I need it once and it really hurts?” If you could access intelligence anywhere, anytime, at very low cost within 2 to 4 hours, why wouldn’t you?
• He gives the example of a colleague driving through Xinjiang who suddenly needs to change hotels, find the nearest gas station and repair shop, and estimate the cost of repairs. That’s when intelligence becomes essential. Looking only at shallow work and life use cases oversimplifies people’s needs, he says: the cost of accessing intelligence can change the demand itself.

6. ChatGPT and Claude are among the competitors for personal agents

• Asked by 曲凯 whether Muse, GrapeBot, and similar products compete with ChatGPT and Claude, 子皓 says, “I think they do.” He expects life use cases to ultimately revolve around clear needs; once a need is clear, all kinds of products will move to serve it.
• 子皓 says that in Silicon Valley, you can see OpenAI is about to launch a bot with something like a virtual machine, and Toutiao is also going to launch a similar product.

7. There’s no single right organizational model; coworker products are slow because they require organizational change

• On the debate over whether multiple agents should serve one person, each person should have one agent, or each team should have one, 子皓 says, “All of these approaches could work.” Different organizational structures and practices can succeed: some organizations have almost no privacy, some are competitive, and others preserve some privacy while sharing context.
• He thinks coworker products haven’t developed as quickly as expected because deploying them within an organization involves organizational change. Progress will ultimately be constrained by the narrowest point in the “big pipe.”
• Many large companies, he says, have failed to transform through “AI native” efforts alone. A new company that sticks to old ways of working could take more than 5 times as long. It’s like the arrival of electricity: building factories designed around electric power may have taken 30 years. With tools like coding agents, new companies may be more efficient, have denser talent, and accumulate knowledge that compounds, eventually displacing old companies—but the transition won’t happen overnight.

8. The consumer route still has room; winner-take-all dynamics are more about distribution

• 子皓 says some Silicon Valley companies went B2B in the past because they couldn’t build consumer products. Back then, consumer products had high penetration and network effects, and anything that gained traction was quickly copied. He questions why companies should still follow the old playbook in the new era. 曲凯 adds that B2B markets are more fragmented, offer more exit routes, and have a higher floor. 子皓 frames the difference as whether a founder wants to change the world or make some money.
• On winner-take-all dynamics in consumer markets, he says powerful network effects in upstream channels make it difficult for independent giants to emerge. The problem is more about channels and distribution than the product itself. He also says that after The Innovator’s Dilemma was published, new disruptive innovation may become hard to come by.
• On overseas markets, 子皓 notes that TikTok once spent heavily on Facebook ads. He believes offline hardware doesn’t have the same network effects or face such a closed social environment, so there is still room to compete. He later adds that the opportunity for this kind of startup in China is indeed constrained: when you rely on distribution from big tech companies, their tight control over upstream costs and restricted access to channels can limit creativity.

9. A2A network effects may emerge at the agent layer, like iPhone and the App Store

• 子皓 says whether personal AI has network effects depends on the case, but agents themselves could generate them. The key is whether connecting agents creates value for both sides. The causal direction is: “It’s not that you have a product because A-to-A is possible; A-to-A becomes possible because people use your product.” The memory, file system, and other content accumulated through product use are what give agents a basis for sharing and collaborating.
• He compares it to the iPhone and App Store: “The iPhone didn’t exist because there was an App Store; the App Store was built because there was an iPhone.”
• They already have real examples. A Lark group includes colleagues and 子皓’s Hermes bot; colleagues @ the bot to ask it to review documents or explain what 子皓 would make of a particular direction. 子皓 is in the group so everyone can see the bot use his memory and file system according to the right etiquette and instructions. He doesn’t need to say anything himself; he just handles permissions.

10. A product-by-product breakdown: Tom, Google Bot, Muse, and Instinct

• Tom: 子皓 says Tom has the imprint of Y Combinator and recalls that its CEO was formerly Plaid’s CTO. The product takes permissions, references, and connectors seriously, making it suitable for many B2B use cases. It handles a lot of email, automatically tagging and categorizing messages and sometimes drafting replies. Some parents use it to manage emails from schools about summer camps, classes, and homework. But 子皓 thinks there are few everyday tasks that can be fully automated with replies, A2A is complicated, and the overall product architecture is not especially promising. In the long run, he says, intelligent email replies aren’t an extension of intelligence.
• Google Bot: 子皓 thinks it’s a good product. It handles authorization and pass-through well; when MCP or an API connection fails, it opens a browser and lets the user enter an email and password in a pop-up window. He thinks the backing of a big tech company makes ordinary users feel safer, and that the product does a good job with communication and context sharing between bots. It comes from the “Kezi” team, and the team’s growth lead was involved too. 子皓 thinks they got the details and distribution right. He adds that it isn’t aimed at ordinary consumers so much as people who aren’t programmers.
• Muse: 子皓 sees the unmistakable imprint of the mobile internet era: polished execution and heavy investment, with 100 million tokens for every user from the start. On his first day using it, he thought its spending, storage sources, and token consumption would make it hard to replicate at scale. Muse promotes itself at the top of Instagram and is easy to get started with; it also has a goal feature for long-horizon tasks spanning programmers and ordinary users, as well as the physical and digital worlds. 子皓 thinks it is more consumer-oriented than Google Bot.
• Instinct: 子皓 says they don’t have the right environment to use it. “Maybe it’s hard for us, living in a country with Meituan, to understand what the need is.”

11. Growth narratives are returning to the internet era; 子皓 says big tech has lost its ambition

• 子皓 thinks marketing around booking flights and restaurants solves the problem of OpenClaw being hard to describe. Users can easily understand: “I download an app and have to pay for it; now it’s free. After I download it, it might save me RMB499.” Lovable’s website-building pitch is similar: use a tool to build a site that could make you money or bring in customers. Growth narratives are back to “helping you save money and make money,” which he says is close to the fundamental growth instinct in China.
• 曲凯 says that in an environment like the past, Meituan, Tencent, and ByteDance might have fought for users with subsidies. 子皓 thinks big tech companies have lost the ambition they had in the previous era to become number one in the world and lead globally. Many products are “China’s version of this” or copies of “that foreign product”; truly original products that people love are rare.

12. The latest boom was predictable; 子皓 says founders need their own taste and conviction

• 曲凯 reflects that it usually takes 6 to 12 months after a model or a product like OpenClaw becomes a hot topic for it to turn into a real product. He gives Manus as an example, saying it started about 6 months after o1 launched. He also notes that more than 100 teams were working on OpenClaw-like products, but almost none ultimately raised money, after which startup activity in Chinese applications cooled.
• 子皓 thinks the domestic environment and the enthusiasm burned up in that first wave of the year both played a part. Since the arrival of LLMs, people have repeatedly tried to build “China’s version of” something. Short-term, opportunistic projects are unlikely to succeed, he says; startups need their own taste and convictions. He cites Jensberg, Melius, and Paulin as examples, saying none is “China’s version of” anything.
• On timing, he says he once wrote that “there is no such thing as timing”: before agents emerged, no one believed you when you said you wanted to build agent hardware; before OpenClaw, no one believed you when you said you wanted to build something similar. Different timing simply brings different foundational assumptions.

13. Ambi’s differentiation: offline context and a new interaction point keep AI present with the user

• 子皓 says Ambi combines software and hardware, adding offline context—and especially an offline interaction point—to the personal agent. One gap in existing personal agents is that every interaction takes 30 seconds to 3 minutes, or even 5 minutes. Muse’s Charm hardware, a prototype worn on the wrist and more than 50 millimeters thick, also shows that the other side sees this direction.
• Ambi is betting on AI being “present alongside you,” rather than asynchronously requiring the user to participate in many steps. He agrees with some of Jeff’s views: software should run continuously without requiring people to manage it, and AI should also run continuously, in real time, alongside the user.
• Ambi could build Computer Use, food ordering, and discount features, but won’t lead with them. 子皓 says the differences from other products lie in interaction, context handling, and technology, and that the company is investing heavily in real-time interaction.

14. The time cost of “say it and it happens”: users shouldn’t have to repeat offline context

• 子皓 says AI glasses require users to wake the device, say “Hey Meta,” and repeat what the other person just said before making a request. He finds that repetition strange because the device is on your face and can hear what’s happening. “It’s because it isn’t always running—not like software, and not like air that’s just there around me by default.”
• He invokes Her and Iron Man: the characters don’t repeat the conversation they just had, so why should people get used to doing that in real life? He also borrows Steve Jobs’s example about computer startup time: if everyone uses a computer every day, reducing a wait from 30 seconds to 3 seconds saves time for the whole world.
• 子皓 says people may talk to their personal agent only 3 times a day now, but it’s easy to imagine that rising to an average of 30 statements a day within 6 months. If it takes 5 seconds of engineering to make “say it and it happens” work, but the user has to spend 30 seconds repeating something, that wastes a lot of time. When 曲凯 mentions manually entering offline information after a meeting, 子皓 says, “That’s stupid.” AI should run continuously like software, hear conversations, and understand why they’re happening.

15. Privacy: 子皓 says Ambi handles data much like Apple

• 子皓 says Apple does all its AI processing locally and doesn’t retain recordings; Ambi is short for “ambient.” He says none of the audio files ever touch Ambi’s servers, and after they’re processed on a third-party platform, “we may not even have them anymore.”
• Asked whether images are processed locally or uploaded, he mentions that some desktop products use Apple’s Accessibility API to retrieve content instead of taking local screenshots. He says the approach is similar to Apple’s, except Apple does it on its own chips, while Ambi may need the backing of Google GCP or something else.
• 子皓 compares concerns about the device to early concerns about rear-facing phone cameras and says trust is the key to protecting privacy. The best approach is for no app provider to be able to access this data at all.

16. Training models is a Pareto trade-off, not a push for greater intelligence; on-device models are constrained by battery

• 子皓 cites a Sequoia partner’s view that model APIs offer a set of points on a Pareto frontier. Application companies train models mainly to find the right balance among cost, speed, and intelligence, not to pursue greater intelligence—much like optimizing code and improving ROI. If it has the capability, Ambi will also do more to improve speed and cost over time.
• But he rejects the idea that application companies can train models to be more intelligent, calling it absurd. “Who am I to prove what I know that Claude doesn’t?”
• On-device, 子皓 says hardware has to respect the laws of physics; the constraint is more often battery than compute. Given Ambi’s form factor, the company has opted not to run a large model on-device: the bigger the battery, the heavier and thicker the device, which affects wearability. A device that sits on a desk or goes in a suitcase could make a different choice.

17. Lessons from several careers: timing, business models, and efficiency

• 子皓 spent 6 years in social-impact entrepreneurship, raised 3 rounds, and eventually sold the company, returning the money as well. His biggest takeaway was: “Timing and the sector matter more than ability and everything else.” He thinks choice matters more than effort, and that fate or a larger wave may matter more than choice.
• He then spent 4 years on insurance, innovative payments, SaaS, and other businesses. His view is that business models are simple at their core: he has worked across 6 or 7 models, including Douyin ads, CPA marketing, state-owned enterprises, central government-owned enterprises, foreign companies, foreign pharmaceutical companies, and innovative payments; in each, he could reach at least the top 3 nationally, often the top 2. Business ultimately comes down to efficiency, ROI, and demand.
• Hardware experience gave him a grounding in R&D, warehousing, and design. He thinks building hardware is a bit like B2B software: eliminate downstream risks as early as possible to make delivery smoother and cheaper, rather than constantly iterating by trial and error. The LLM era also reminds him of the early mobile internet: when foundational technology changes rapidly, short-term winners can quickly lose their edge, while the pace of change and core methods become competitive advantages.
• 子皓 describes himself as a “product and business” person, though he says his stronger suit is business. 曲凯 calls him steady and someone who delivers results. 子皓 jokes, “No matter how bad the business is, I can still make money from it.”

18. How to make money: ice-skating tickets, Qingsongchou, and Hainan telemarketing

• After graduating from high school, 子皓 wanted to see how much he could earn without studying, just by doing physical work. On his first day selling ice-skating tickets on Beijing Road, he sold only 5. He spent 5 hours outside an EF Education office and got just 1 person to come upstairs. Then he noticed young people liked to go ice-skating and recruited student agents through a part-time jobs website: the tickets cost RMB25, he paid agents RMB27, and they sold them to classmates for RMB30. He eventually became the rink’s exclusive agent, and by his junior year of college, people were still calling his sister to get tickets.
• After Qingsongchou acquired the company he had started, he needed to prove himself by delivering results. He spent a weekend going through the insurance sales process, traffic sources, and user journey page by page, using URLs and basic data to find where users got stuck. His analysis helped the business grow by 30% to 40%.
• When he first ran a telemarketing center in Hainan, he asked for a budget to give everyone 100 red flags. When a deal closed, they would bang a gong, put up a flag, and tell everyone how much each flag was worth. That month, sales doubled and profit reached more than 200. 子皓 thinks the key is finding gaps in the business and understanding what people need.
• Later, when he worked in innovative payments, price wasn’t the only reason he won partnerships. In a market dominated by a single competitor, he emphasized that working with him was “clean,” meeting the partner’s needs for safety and for maintaining its position.
• His summary after all these years is that business means identifying demand and finding a way to deliver a solution. The constant lesson is to “go where the fish are” (去有鱼的地方抓鱼). Asked whether he prefers large or small markets, he says that when he was young he would go where there were fewer fish and fewer people; now he goes where there are more fish and more people. “At a certain age, you start asking: why can’t you make money in a big market?”

19. There’s no shortcut to hiring: be genuine, invest time, and notice signals of discomfort

• 子皓 sums up business as “recognizing demand externally and hiring the right people internally.” The gap between meeting demand at a reasonable cost and bringing in the right people is the residual profit. When hiring, he thinks experienced people can tell truth from fiction. There’s no need to lie to someone more capable than you: share what you really think, why you’re doing this, what you can offer, and the “process-level rewards” that may not be financial.
• He says many employees were well-paid freelancers before joining. To attract them, the key is whether you can give them a reason to do something greater. When he was recruiting the head of the Hainan telemarketing team, he waited for 5 or 6 hours at a small shop downstairs and ate several bowls of qingbuliang before the person agreed to meet. If you say hiring matters, he says, you need to put in the time and sincerity.
• He also has a hiring rule: if something feels a little uncomfortable in a conversation, don’t move forward, because that discomfort may eventually grow. Move on quickly. There are no shortcuts to hiring or building a company.

20. Do 3 years’ work in 1; speed is the biggest concern

• 子皓 has been thinking about how work that once took 30 people 3 years to code might now be done in a year or a few months. He is also rethinking the traditional hardware go-to-market route: test on Kickstarter, find media coverage, launch a standalone site, go on Amazon, then go global. Why does that have to take 3 years? He wants to use coding agents to rethink the business, organizational structure, and ways of working.
• His conclusion is that many ways of working can be much more aggressive: “I hope to do in 1 year what others do in 3.” But he opposes reducing a company to an OPC. Ambi’s dozen-plus employees all work hard, and that’s what made the current product possible; looking only at coding can’t explain where the company is today.
• He sums up the experience he has built over the past decade-plus as trying not to make mistakes, especially one-way-door mistakes. His biggest concern now is speed, because everyone else is moving fast too. He points out that an Apple designer went to Meta and built Charm in just 10 months.
• When 曲凯 wonders why the world has to get more and more competitive, 子皓 replies that many people have finally reached an era worth competing in. After 曲凯 talks about the changes AI is bringing, 子皓 adds: “Most people only get one life.”

21. Hardware’s strengths and weaknesses; there’s a gulf between a Shenzhen demo and reliable mass delivery

• 子皓 says hardware acquisition channels don’t have particularly strong network effects on their own, but once established they can become an advantage. He cites Anker, DJI, and Insta360 as companies that can use existing channels to launch new products. Douyin is different: without building an account and buying ads, even Dongfang Zhenxuan might not be able to sell.
• Hardware has exclusivity: at any one time, you carry only 1 piece of hardware, while there may be 5 software products running in the top-right corner of your computer. Once someone owns a device, the value of add-on purchases for that device may increase in the AI era.
• If software works well for 98 out of 100 minutes and poorly for 2, users can turn to another product to fill the gap. Hardware, by contrast, can have 2 particularly bad minutes every day. So even if a use case takes up little time, if it happens every day, it has to work well.
• The downside is that hardware is slow, requires iteration, and requires taste; it has drawbacks as well as advantages. 子皓 thinks the AI era leaves more room for hardware to survive, and notes that Apple isn’t a monopoly and GoPro hasn’t gone out of business.
• He warns against treating Shenzhen’s ability to build demos as proof that a product can be delivered reliably at scale. For software, going from a demo to scale may just mean hiring a few more people; for hardware, there may be a 100x gap between reliable delivery and high integration. Shenzhen is a good place to try building something, but mass manufacturing requires completely different capital and capabilities.

22. Interaction is the biggest non-consensus opportunity; he says one kind of proactive product is the wrong direction

• 子皓 thinks interaction is personal AI’s biggest opportunity. “Say it and it happens” has 2 parts: the user says what they want, and AI gets it done. That depends on a lot of foundational technology. In the model era, many technical concepts quickly become taken for granted: skills, function calling, multimodality, and other capabilities that once seemed impressive soon stop being the focus. He expects that in 6 months, anything that can be digitized may be assumed to be solvable with intelligence. What matters is how people use AI.
• He thinks interaction can create a degree of defensibility because the value it builds up may produce network effects and switching costs. Specific technical capabilities will keep improving, so the technical problem itself doesn’t deserve too much discussion.
• On proactivity, he points out that scanning for connector changes or running scheduled scans both cost money; the smarter the system, the higher the cost. But when it pushes something in front of the user, it is guessing what the user wants, which may not match the situation at hand. The odds of content feeling like an “aha” without being intrusive affect the product’s value.
• 曲凯 illustrates the problem with a secretary knocking on the door 20 times a day, and getting it right once. 子皓 thinks a product that guesses what users want and pushes things at them frequently is the wrong direction: “We’re not building that.” He adds that Ambi’s proactivity has a clear trigger: a change in the environment or context, not simply more proactivity.
• He thinks work and life shouldn’t be separated, because context from both helps AI understand the user better. That may not work for everyone, but it makes sense for them. They also don’t plan to define a fixed persona: AI is part of the user, an extension of them, not some unrelated object. He says there’s a difference between someone looking for “曲老师’s assistant” and “an extension of 曲老师”; the latter is more likely to generate network effects.

23. The end-state vision is service you can carry with you; 子皓 says consumer timing is right, but China has few standout products

• 子皓 hopes everyone will eventually think of AI as “a service you carry with you.” He recalls wanting a drone as a child that would follow him around and carry things: people want access to intelligent services anytime and anywhere. 曲凯 says that as a child he wanted a Doraemon.
• In non-entertainment settings, people spend only a small fraction of their time on phones and computers. If AI is to create more value, 子皓 says, it shouldn’t just compete for screen time; it should help people handle things outside the screen so they can do more in parallel.
• 曲凯 estimates that Manus, Pluto, and similar tool products have hundreds of thousands of paying users, and compares them with Muse, Grokbot, WorkBuddy, Doubao, and other products aimed at hundreds of millions of users. 子皓 chooses the latter route and thinks the two kinds of products are equally hard to build. The reason mass-market consumer products hadn’t emerged before, he says, was that the timing wasn’t right.
• 子皓 thinks the timing is right now. On the one hand, users don’t care which model they’re using; on the other, models have progressed from instruction-following and multimodality to long-horizon tasks. He predicts that in 6 months, all tasks that can be digitized will be assumed to be doable, infrastructure will be as mature as water, electricity, and gas, and the next step will be scaling up. He means everyday consumer use cases, not continued work on problems like cancer and other diseases.
• On China’s application environment, 子皓 says that in all his years in the industry, this is the first time he has seen so few Chinese products that make people sit up and take notice: by late September or October this year, there were almost none that everyone knew about. “I think Granola is pretty average, but there aren’t any products made by Chinese people that are better than Granola.” He says US companies have built good products in data, B2B, and vertical consumer markets, and he agrees that Tom and Instinct are good products. As for why China is in this position, he says he doesn’t know. 曲凯 ends with: “Still, I hope our country does well.”