The company calls it a system-level agent, but this is not China’s first attempt at such a creature. Xiaomi, Honor, and Oppo’s Breeno all once flirted with device-anchored agents before. Doubao Mobile Assistant is different. It arrives not as an extension of a handset brand but as a product of China’s most formidable attention machine.

The last time ByteDance ventured into phones, it did so through Smartisan, the quirky Chinese handset maker whose patents and staff ByteDance acquired in 2019. That effort produced a single device, the Smartisan Jianguo Pro 3, briefly marketed as a “Douyin phone” for its deep integration with ByteDance’s video apps. The project never expanded into a sustained hardware strategy, and much of the original team was eventually absorbed into other ByteDance units, including education.

This time, ByteDance did not attempt to build a phone of its own. Instead, it created an experimental operating environment and deployed it inside someone else’s hardware.

Through a partnership with ZTE, Doubao appeared inside a modest Nubia smartphone priced at 3,499 yuan (US$482). It sold out on day one, not because of the phone itself but because buyers wanted to glimpse what happens when an AI is allowed to live inside a handset rather than merely accessorize it.

In fact, the Doubao team has already been moving toward this moment for some time. On PC and Mac, their AI assistant already has the power to read screens and perform automated actions.

Luxuries made possible by the relative openness of desktop systems. Smartphones, however, operate with tighter boundaries. Without system-level injection, any agent, no matter how intelligent, is confined to its own sandbox.

And this is where the double edge shows. On mobile, the question is almost never about what the agent can do. It is about who watches the agent once it is permitted to move on its own.

I luckily snatched a unit of the Nubia M153 before it was gone. ByteDance insists it is only an engineering prototype, a vessel for the assistant rather than the beginning of a Doubao Phone line.

It comes with a Snapdragon 8 Elite Gen 5 processor, a 6.78-inch LTPO display, 16GB of RAM, 512GB of storage, a 6,000mAh battery, NFC, infrared, and a silver-white chassis punctuated by a physical button designed for one purpose: summoning the Doubao AI. Also, the phone runs Obric UI, ByteDance’s customized environment for this experiment.

In demonstration videos circulating online, Doubao behaves like an invisible “roommate” with inexhaustible patience.

Ask it to compare the price of a chicken sandwich across Meituan, Taobao Flash Sale, and JD.com, and it slips silently into each app. It searches, taps, logs, and returns with a tidy summary. Instruct it to place the order and forward a screenshot to a friend on WeChat, and it does so with clerical devotion. The only step still reserved for you is the moment of payment, a boundary the assistant has not yet been invited to cross.

In my own tests, the device lived up to its viral reputation. It could swipe through TikTok videos on its own, fill in online quizzes, and perform actions that felt not dangerous exactly, but unsettling in their plausibility. A simulated bank transfer, for instance. It never entered the password, since Doubao does not yet impersonate the user’s hand to that degree, but it maneuvered seamlessly to the threshold of transfer, as if waiting for a quiet nod of approval.

Still, none of this qualifies as magic. Doubao works because it sees the screen, identifies UI elements, and takes over the gestures that would otherwise require a human finger. It is not communicating with apps through underwater interfaces. It is just impersonating the user, frame by frame, tap by tap.

And that impersonation quickly became a problem. As the Nubia M153 reached more users, major Chinese apps began to reject it outright.

WeChat temporarily blocked logins from the device. Alipay and several banking apps followed, either through active bans or through automated risk-control systems that detected unfamiliar patterns and locked users out.

For a moment, it felt as though the country’s most widely used apps were pushing away a foreign body, not because the device was faulty but because it behaved too much like a person. Or perhaps too little.

What Doubao attempts to embody is what the industry has begun calling agentic AI, a category of intelligence that no longer waits for the user to tap their way through an app but instead performs the task itself.

In the Chinese imagination, this promise has been circulating in the last 3 years. Every major smartphone keynote has gestured toward a future where the phone becomes an active collaborator rather than a passive tool, a kind of digital aide-de-camp that can navigate the fragmented landscape of apps on behalf of its owner.

@user90tymlsymv赵明荣耀发布会现场用手机AI给观众点奶茶，网友： 3万块上个热搜，这广告费划段。

Tiktok failed to load.

Enable 3rd party cookies or use another browser

Doubao does not reinvent what an AI assistant is supposed to be. It simply carries out a familiar vision with a level of completeness that previous attempts never reached. What makes it feel new is not the idea itself, but the distance it is willing to travel to make that idea actually work on a modern smartphone.

That distance becomes clearer when placed next to Apple and Google. Their approaches differ not in technical capability but in philosophical posture.

Apple Intelligence, introduced with no shortage of ambition, still behaves like a cautious house guest. It suggests, summarizes, polishes, and retrieves, yet avoids touching anything that might belong to someone else. Google’s Gemini, even when deeply embedded into Android, operates through carefully negotiated APIs. It speaks to apps through doors left politely half open. The boundary between assistant and application remains intact, preserved by a kind of tacit social contract.

Doubao operates as if no such contract exists. Instead of waiting for app makers to build integrations, it studies the screen directly and decides what to do next. The difference is subtle but consequential. Apple and Google designed assistants that ask apps for permission to help the user. Doubao assumes that the user’s request is permission enough.

For anyone watching the evolution of AI on mobile devices, the arrival of Doubao marks a moment when the future forks. One path favors assistants that remain deferential, bound by formal integrations and explicit trust structures. The other imagines assistants that act first and justify their actions through results.

Doubao chooses the second path, not because the vision is especially radical, but because it takes the fastest route available. In a landscape where every major app operates as a walled garden, the only way for an agent to function today is to slip between those walls rather than wait for doors that may never open. The approach works, but in the same way an iOS jailbreak works: as a hack, driven by impatience as much as ambition. It produces striking results, yet it also raises the question of whether such a method can ever be viable, or safe, for the mainstream.

The ability to reach that far comes from a single Android permission called INJECT_EVENTS. It looks minor on paper, but it is one of the most powerful privileges in the system. Google classifies it as a signature-level permission, reserved only for components signed with the same cryptographic key as the firmware. In practical terms, the assistant is not just an app installed on the phone. It is treated as part of the operating system itself.

Most permissions open a door to a specific function. INJECT_EVENTS gives access to the entire body. With it, software can write directly into the phone’s input stream, generating touch signals that are indistinguishable from human interaction. The system cannot tell the difference, and neither can banking apps or fraud detectors.

Technically, this means Doubao bypasses the UI hierarchy and acts at the hardware level. It does not ask apps for cooperation. It simply moves through their interfaces as if it were the one holding the device.

The existence of such a permission raises a question that is less technical than structural. Why is a system-level capability like this permissible here but unimaginable in the West. The difference has little to do with engineering talent and more to do with how ecosystems define authority. Apple will never delegate this level of control to an external agent because its entire platform depends on a strict separation between the system and the apps running on top of it. Google restricts INJECT_EVENTS to manufacturers because a global marketplace requires predictable lines of trust. In both cases, the assistant is allowed to assist, but never to operate as a peer of the operating system.

This is also why Doubao ends up arriving inside a ZTE device rather than as a download available to anyone. Without firmware-level signing, the system would reject the assistant’s capabilities outright. Embedding the agent at the operating-system level is not a display of ambition so much as the practical requirement for how this particular model of assistance is built. It is less about breaking rules and more about bypassing the structures that were never designed with agents in mind.

The contrast becomes even sharper when placed next to the path Apple and Google have been taking. Both companies have spent the past few years building what they call Intents frameworks — Apple’s App Intents and Google’s expanding library of system-level intents. These frameworks move slowly and often feel frustratingly limited, but their premise is straightforward. They create permissioned, auditable pathways through which assistants can request specific actions without overstepping application boundaries. The assistant does not impersonate the user. It negotiates through a shared vocabulary that the ecosystem has explicitly agreed to honor.

From afar, this looks conservative. In practice, it reflects a belief that cross-app automation only becomes sustainable when the rules are explicit and the lines of authority are preserved. App Intents represent a long-term path. It is slower, but it does not require the assistant to jailbreak the operating system to get work done.

By comparison, Doubao’s method feels like a short circuit: effective, even elegant in its way, but undeniably a hack. It reaches for outcomes the ecosystem has no native protocol for, filling the gaps with system-level privileges that were never intended for agents. This makes the device impressive in the short term and difficult to justify as a mainstream model in the long term.

The more time I spend with the device, the clearer it becomes that Doubao is not simply illustrating what an AI assistant can do. It is illustrating what happens when an assistant begins operating outside the boundaries that apps, platforms, and users once assumed were fixed. Its power is not a side effect. It is a structural consequence of an ecosystem built from many self-contained worlds that do not willingly speak to one another.

In that environment, any agent that hopes to function must cross walls that were never designed to be crossed. Doubao does so with an ease that feels impressive at first and unsettling immediately after. The discomfort is not because the assistant is misunderstood. It is because its method—reading every screen, acting as the user at every layer—is fundamentally indistinguishable from overreach.

If anything, the ecosystem’s resistance is not misdirected. It is a reminder that capability and legitimacy are not the same thing. A system-level agent that navigates financial apps, social platforms, and private content on behalf of the user without the consent of those apps is not merely innovative. It is introducing a new political reality into mobile computing, one in which the operating system can quietly overrule the autonomy of every application in its path.

Doubao is, without question, the most competent mobile agent I have ever used. But its competence exposes something larger. Perfect automation in a fragmented ecosystem requires a form of intrusion, and that intrusion forces us to reconsider what a phone is allowed to do on our behalf. The question is not whether Doubao succeeds. It is whether the model it represents should exist at all.

The future of AI on mobile devices will not be shaped by who builds the fastest model or the most convenient assistant. It will be shaped by whether we decide that an agent acting “as us” is an acceptable default in a world where neither apps nor platforms were built to accommodate such authorit.

I barely gave it a second thought until she followed up with another message: “How about that? Written by Yuanbao, reviewed by me.”

For those unfamiliar, Yuanbao isn’t a teacher from my mom’s elderly class. It’s an AI assistant made by Tencent, just like ChatGPT or Gemini.

Slowly, I started noticing other changes too. She was cooking new dishes, singing with better technique, even handling my son with a level of precision that made my wife say, “Mama’s parenting is getting surprisingly scientific.”

Is it because of Yuanbao? Maybe not entirely. But I couldn’t help noticing the timing.

Which made it all the more striking. My mother, a retired drafter from our local metallurgical factory, once approached even routine phone software updates with the kind of trepidation usually reserved for major life decisions. Now, she was casually crediting artificial intelligence for her writing, and she did so without consulting me, the supposed tech-savvy member of the family.

Before I could even begin to publicly apologize for not introducing AI to my parents sooner, it became clear that her embrace of this technology is hardly an anomaly.

In fact, her curiosity didn’t stop with Yuanbao. She soon discovered Doubao, ByteDance’s AI assistant, primarily because Doubao’s responses sometimes include videos.

This proved to be a significant draw. When my son would bombard her with endless children’s questions, she found herself turning to Doubao for motion explainers, illustrating concepts visually for him.

And remarkably, she gradually began to compare these AI tools with the discerning eye of a seasoned reviewer. She would engage in discussions, debating which AI “feels more natural” and which is better suited for “formal occasions.”

I was genuinely impressed. The acceptance of AI within my family, much like my own journey with these tools, seems to have spread quite organically.

However, a lingering question remains: how did she stumble upon all of this without my intervention?

For years, I've been the self-appointed tech evangelist in our family. I was the teenager who insisted we acquire our first home computer back in 1999. I was the one who pushed for getting it connected to the internet in 2003. I even gifted my parents their first iPhones and walked them through the intricacies of setting up Apple IDs and navigating the App Store…

Sure, this time I didn’t offer them ChatGPT because its geopolitical complexities made that a non-starter. But the fact that she managed to independently find and adopt these alternatives is a source of genuine curiosity for me.

So I asked my mom directly who guided her to these tools. Her answer was disarmingly simple: "They're already there." She meant they were inside her daily necessity apps: WeChat and Douyin (TikTok’s China version).

And just like that, the puzzle pieces clicked into place. This is precisely why she never had something like DeepSeek on her phone. AI was already woven into the fabric of the platforms she uses every single day.

Yuanbao, for instance, has its own WeChat bot and was once prominently promoted in WeChat’s Life Services section. Douyin offers several access points for its Doubao assistant, one embedded directly within the short video interface, alongside likes, comments, and shares. Another is tucked inside the message list, blending AI with social functions.

This subtle integration, it seems, is the true engine behind Chinese AI's quiet, organic spread into the hands of millions, even those, like my mom, who might otherwise shy away from new technologies.

It turns out, she’s not alone.

Across China, AI adoption is quietly spreading from cities to villages, driven not just by curiosity but by an aggressive push from tech giants.

Yuanbao, the same assistant my mom now uses in the city for writing, cooking tips, and even parenting advice, is showing up deep in the countryside. Tencent has painted promotional slogans on village walls, a format as familiar as it is effective in the countryside.

Douyin and Kuaishou push AI-generated selfie filters to villagers. These playful features help demystify the technology. Also from there, awareness of tools like Doubao, Yuanbao, and even DeepSeek builds gradually, often through word of mouth.

Beyond marketing, AI is also taking root in agriculture. Systems powered by machine learning are monitoring pig farms, optimizing crop yields, and automating greenhouse operations. Tencent’s AI team even trained models to grow cucumbers more efficiently, a quirky but telling example of how AI pilots start in rural China.

Healthcare is another frontier. Baidu’s AI eye-screening machines are providing diagnostics in areas where qualified doctors are scarce. Other initiatives use smart speakers and AI assistants to deliver medical advice or basic education where resources are limited.

Yet beneath this seemingly cheerful embrace of new technology lies a more sobering reality. China’s AI race is indeed reaching the everyday user, but the competitive landscape is brutally unforgiving.

While AI embedded in super-apps like WeChat and Douyin accelerates adoption, it also masks an increasingly cutthroat market. Companies are slashing prices, offering services for free, and burning through resources at alarming rates, with no guaranteed path to sustainable monetization.

Ironically, it’s not only global chip shortages or geopolitics that threaten China’s AI players. It’s relentless domestic competition, say, “内卷 Nei Juan”.

At first glance, it feels like a winner-takes-all game dominated by giants with seemingly bottomless pockets. But the reality is more layered. Survival depends not on who has the largest model or flashiest demos, but on who achieves early product-market fit or has robust existing revenue streams to subsidize the long game.

Take DeepSeek, for example. Their decision to open-source models isn’t philanthropy. It’s strategic confidence, backed by Quantitative hedge fund High-Flyer’s lucrative trading empire.

By contrast, players like 01.AI, once one of China’s celebrated Six AI Dragons, are already retreating. Not for lack of ambition, but because pure model development without real products, burns capital and patience at an unsustainable pace.

Another challenge is that for most people, AI tools still feel remote. Unless a model rides a viral wave like GPT-4 or DeepSeek R1, building meaningful adoption is slow, tedious work.

That explains why companies are embedding AI into the platforms people already use. Baichuan targets industry verticals. Tencent weaves Yuanbao into the routines of WeChat users, from city dwellers to villagers.

My mother didn’t start using Yuanbao because of a tech demo. She used it because it quietly appeared inside her most-used app, offered help, and asked for nothing more than a few seconds of trust.

Millions of others are doing the same. Most won’t and can’t pay for ChatGPT, Gemini, or Grok. But they will keep paying for Tencent Video, Tencent Meeting, or WeChat services, especially if AI subtly enhances those experiences.

Watching my mom toggle between Yuanbao and Doubao, comparing their diagloue styles with the same discernment she once brought to microwave ovens, I’m reminded the real AI race won’t be won by parameter sheets or technical benchmarks.

It will be won, quietly, in moments like these. When AI stops feeling like the future and simply becomes useful.

More specifically, to visit Pangdonglai (胖东来, literally “Chubby Donglai,” with Donglai being the founder’s given name), the viral supermarket that had overtaken Douyin and Xiaohongshu with something close to messianic fervor.

I had gone shopping across provincial or national borders once before, back in 2012. It was somewhere in the Netherlands, near the western edge of the German border, though I could not identify the precise location now. What I do remember vividly is seeing a €500 banknote for the first time, handed across the counter of a Gucci or Chanel store.

And this time, a decade later, we drove six hours to a local supermarket in Xuchang, a lower-tier city in central Henan and better known as the capital of Cao Cao’s Wei empire during the Three Kingdoms era.

Unlike that first cross-border trip, I do not think I will forget a single detail ten years from now.

The first surprise came the moment I stepped inside: the entrance lobby. I had never seen a Chinese supermarket styled with such deliberate calm, such Apple Store-like reverence. It was clean. It was bright. It felt almost sacred. Imagine walking through the usual chaos of a city plaza dodging electric scooters and street vendors and then stepping into a space with no promotions, no clashing advertisements, not even background music. What greets you instead is an unexpected sense of serenity.

Only later did I learn that Pangdonglai has a strict policy: no promotional booths at the entrance. Most supermarkets in China place the exit far from the entrance and line the path with vendors hawking everything from credit cards to shampoo. It is common. It is smart business. But who goes to a supermarket expecting something like spiritual clarity? And yet, standing in that atrium, I realized maybe that was exactly what I had come for.

The interior is not lavish, but it feels unmistakably high-end. The floor tiles are immaculately laid, with no black grout lines like those in most malls or markets, just subtle filler that matches the pattern. Perhaps the most charming detail of all: at the entrance, Pangdonglai posts the sourcing information for every material used in the store, including tiles, lighting, and refrigeration units, complete with brand names and contact numbers.

As far as I know, it is the only supermarket in the world where the hardware gets its own credits.

But what impressed me most, more than decorations or the tiles, were the people. The Jinhui branch we visited is the smallest of Pangdonglai’s stores in Xuchang. It was busy, but not overwhelming, and there was enough parking space that made the experience feel local.

What struck me at Jinhui was that every employee I encountered looked genuinely happy. It was not the forced cheer you often see from highway tollbooth attendants, but something more relaxed, more sincere, as if they actually liked being there.

One small but telling detail: all the cashiers were seated. This might be standard in many countries, but in China it is still rare. Even at high-end chains, checkout clerks are expected to stand for their entire shift. At Pangdonglai, the simple act of letting staff sit felt quietly humane.

And the difference did not stop there. I noticed that employees moved with ease, without the mechanical tension that often accompanies low-wage service work. I saw a security guard helping an elderly customer with her groceries. When another customer’s electric scooter got stuck in the lot outside the supermarket, he stepped in to assist, even though the parking area technically was not part of the store.

At Pangdonglai’s flagship location, Angel City, I had another unexpected encounter. In the TV section, I asked a young saleswoman whether they had LG models. She told me they did not. I then asked about Samsung. Before launching into the usual skit about features and specifications, she paused and said, quite plainly, “But Samsung TVs do not support Dolby Vision.”

That single sentence broke every rule of conventional Chinese retail logic. She began with a limitation, not a selling point. What surprised me even more was that she seemed to say it because I had asked about LG first. She knew the difference, even though LG was not sold in the store at all.

There were so many other small details, so many quiet interactions, that stayed with me. By the time we left Xuchang, the trunk was filled with Pangdonglai orange juice and premade fruit cereal, and we carried something else too: a lingering sense of wonder.

In 2024, a wave of food safety scandals swept across China. Almost overnight, Pangdonglai became a national symbol of trust. Demand surged. Their online store began selling out within minutes of each restock, and an informal gray market of resellers sprang up. What had once been a beloved local supermarket was now something else entirely: a pilgrimage site, a hotspot for speculators hoarding Pangdonglai-branded goods, a destination where out-of-towners lined up just to say they had been.

From what I have seen in recent videos, crowd control is now a daily necessity at many of their stores. The experience I had in 2023 may no longer be possible.

At the same time, supermarkets across China began taking notes. Many started copying Pangdonglai, or at least trying to. Among them was Yonghui Superstores, one of China’s largest grocery chains and a company several times Pangdonglai’s size.

When Yonghui, sometimes called the Chinese Walmart, announced in 2024 that it would undertake a Pangdonglai-style transformation, the move was both a declaration of ambition and a quiet admission of crisis.

Yonghui’s 2024 financials made the stakes clear. Total revenue fell 14.07 percent year-on-year to 67.574 billion yuan (about 9.31 billion USD), while net losses widened to 1.465 billion yuan (around 201.7 million USD).

As of June 2025, Yonghui has completed the revamp of 100 stores. Each of its 28 national regions has now delivered at least one remodeled store on its own.

The company plans to accelerate. By the end of July, it aims to have 150 upgraded stores. By September, 200. Before the Spring Festival in 2026, more than 300. The ultimate goal is to complete the overhaul of all 600+ stores by June 2026.

That transformation is happening at every level. In parallel with the store revamps, Yonghui is developing over 60 private-label products this year alone, with more to come in 2025 focused on quality, clean ingredients, and regional character.

Yonghui is also experimenting with profit-sharing for frontline staff. In the first quarter of 2025, 41 transformed stores distributed over 26 million yuan in bonuses and dividends to employees. The company has launched an internal “Craftsman Plan” to train 10,000 skilled workers and aims to offer top-performing store managers some of the best compensation in the industry.

Despite real financial pain, the company argues that much of it is temporary. Clearance discounts on legacy inventory, shutdown costs for hundreds of old stores, and upfront investments in new systems and people are short-term hits.

Still, insiders say profitability is already returning. Of the transformed stores that have been open for three months, most are back in the black. Sales have, in many cases, doubled. During the recent Dragon Boat Festival holiday, sales at over 80 stores grew 87 percent year-on-year.

“Do not rush,” Pangdonglai’s founder Yu Donglai reportedly told Yonghui leadership. “Take two or three years to get healthy again.” If the trajectory holds, Yonghui may just do that, and emerge on the other side not only leaner, but more human.

Near where I live, there is an RT-Mart. The supermarket brand originated in Taiwan, and Sun Art Retail runs the chain in mainland China. Formerly under Alibaba’s retail arm, it was sold to DCP Capital in early 2025. It is also beginning to absorb Pangdonglai’s influence, at least in part.

From my observation, changes became visible in recent months. Free water dispensers were installed, shelf heights adjusted for easier access, displays looked more intentional, and customer interaction zones began to appear. The mood was changing.

On February 27, 2025, Alibaba officially sold Sun Art Retail Group to DCP Capital. Just weeks later, Sun Art Retail CEO Shen Hui visited Pangdonglai, toured its central kitchen and employee quarters, and joined in discussions on the future of retail. Initial plans suggest that RT-Mart will pilot Pangdonglai-style revamp in select stores across Shanghai, Guangdong, and Jiangsu Provinces.

By September 2024, RT-Mart operated 502 stores nationwide, including six membership-format outlets. Its mid-year financials showed a 2.96 percent decline in revenue to 34.7 billion yuan, but a sharp profit rebound to 206 million yuan, up over 150 percent year-on-year.

Shen Hui has stated that RT-Mart’s future depends not only on reducing costs but on rebuilding consumer trust through value, service, and consistency. The company plans to increase its share of private-label goods, especially in its membership stores, and is restructuring its supply chain to focus on freshness and tighter vendor relationships.

China’s supermarket sector continues to face fierce competition and internal pressure. Membership retails such as Walmart’s Sam’s Club and Costco are expanding rapidly, using bulk pricing and limited-time products to build loyal followings. Yet it is Pangdonglai, with just 13 stores in the cities of Xuchang and Xinxiang, that may be offering the most compelling blueprint for what comes next.

In 2024 alone, those 13 locations generated nearly 17 billion yuan ($2.34 billion) in sales and more than 800 million yuan ($110.2 million) in profit. The average monthly income for employees exceeded 9,000 yuan ($1,240), which is more than three times the local average in Xuchang.

For an industry long plagued by burnout, price wars, and razor-thin margins, these numbers feel almost mythical. Pangdonglai has also made it clear that it does not intend to scale quickly. In its 30th year, the company announced plans to enter only its third city: Zhengzhou, the capital of Henan Province.

The Panglonglai-ism is deceptively simple. Serve customers well. Treat employees with dignity. Offer a few unforgettable products. Whether China’s supermarkets can evolve into something uniquely reflective of the country’s values remains an open question.

But if they do, it will require more than redesigned floor plans or a handful of viral items. It will mean a transformation in how we think about the everyday economy, about human-centered business, and about the role of labor in modern life. The consequences may prove far more profound than they appear now. I will be watching.

It was a wild, wild May. One of the wildest I can remember.

When I restarted this newsletter, I had planned to keep it weekly. That didn’t happen last month. I got derailed, not by laziness but by an unlikely collision of projects—all demanding and, somehow, all positive.

So basically, I found myself editing China tech news stories every day for a tech blog; running a Xiaohongshu account about wireless audio and gear; flying to Macau to catch up with old friends I hadn’t seen in six plus years and moderating a panel on sports tech; researching a story about Pangdonglai and YH Supermarket; and, most improbably, learning to “vibe code.”

Yes, code. For an English lit major whose last math class was 19 years ago, I never thought I had it in me. But with the help of AI, I managed to build, in just three weeks, not one but two fully functional full-stack web apps that had been lingering in my head since last year.

The code story, I’ll save for another post. This one is about the moment I realized I needed to think about money.

The app I made is, in short, a Brickset for Dolby Atmos music. A neat little WeChat mini program: clean, focused, and (so I’m told) “not like a mini program at all.” Many WeChat mini programs people encounter every day are cluttered and ad-riddled. I designed mine to feel elegant, but the downside is that no tip jars and no sponsored banners also mean no revenue.

So I thought: why not make an iOS app? It could be the most gently dignified way to ask for money. I can add exclusive features like in-app music playback in return for their maybe 9.99 dollar favor. Not just for the utility, but as a kind of patronage.

Then a CTO friend talked me out of it.

First, there’s the $99-a-year developer fee. But more significantly, there’s the infamous Apple Tax, a commission that goes straight to Cupertino for every dollar an app earns. That means if I want to keep prices reasonable, I either take the hit myself, or raise the price and risk losing exactly the users I’m trying to reach.

To sustain a cashflow, I’d need to market harder, cut price tags, and even compromise on design. The math doesn’t add up.

It made me think about the Epic Games vs. Apple case, when the US regulators finally began to push back. After years of litigation, the April 2025 ruling found Apple in contempt for undermining previous court orders. The court concluded that external payment links must not be subject to Apple’s commission. Spotify moved fast. Within days, its US app included external links for premium subscriptions.

In Europe, under the Digital Markets Act, Apple was fined €1.84 billion for anti-competitive behavior. Korea and Japan also passed laws targeting app store gatekeeping. Globally, Apple’s garden walls began to show cracks.

But in China, things remain unchanged.

I’ve acknowledged that if you are a small business, you’re eligible for the App Store Business Program , cutting the commission to 15%, which is nice until you realize that Apple charges different rates in different regions.

In the US it’s 27% for standard enterprises and 12% for small businesses; in the EU, 17% and 10%; in South Korea, 26% and 11%.

China’s 30% and 15% are the highest tier across all markets. Moreover, unlike in those markets, Chinese developers are still not permitted to implement external payment systems or link to third-party payment services.

In fact, Apple collects more from developers in China than almost anywhere else. In fiscal 2024, its estimated commission revenue from China alone reached 49 billion yuan (roughly $6.7 billion), calculated from Apple’s $66.95 billion revenue in Greater China, with App Store commissions accounting for a significant 10% share of its services segment, roughly a third of the country’s total investment in AI chips that year.

ByteDance, WeChat, even gaming giant miHoYo have found themselves boxed in. Before the 618 shopping festival last year, Apple reportedly delayed Douyin’s app updates until it agreed to route payments through Apple-approved channels. To many, this wasn’t negotiation. It was pressure.

The irony stings. And yes, I envy that in Europe, developers can now use third-party app stores, alternative payment systems, and benefit from lower commission rates. In China, we still pay full freight.

But I’m also happy to see that in Asia, things are starting to change. The Japanese legislative momentum accelerated in 2024 when the Diet passed the Act on Promotion of Competition for Specified Smartphone Software, targeting dominant app store operators like Apple. The law, set to take effect in December 2025, mandates that users be allowed to install third-party app stores and use alternative payment systems on iOS.

In South Korea, change came earlier. As far back as 2021, the country passed a landmark amendment to its Telecommunications Business Act, becoming the first in the world to ban app store operators from forcing developers to use their own in-app billing systems. Enforcement began in earnest in 2022, and though Apple introduced an alternative payment option, it also imposed a 26% commission—prompting further pushback from regulators. These cases helped set global precedent for app store antitrust scrutiny.

With Beijing’s increasingly assertive stance toward American tech firms, Apple may soon face its first formal reckoning in its second-largest market. China’s State Administration for Market Regulation (SAMR) is now preparing to launch an antitrust investigation into these exact issues.

Which brings me back to my modest idea. It’s one thing to say “don’t bother with an iOS app.” It’s another to realize that the platform owner takes more from my software than anyone else in the chain. Apple provides no hosting, no promotion, no payment risk—just a rule: 30%/15% goes to them, because they say so.

I’m not calling for a boycott, and I’m not suggesting China should replicate Brussels overnight. But the silence has a cost. With no regulatory pressure, Apple’s China App Store remains its most lucrative and least contested market.

That’s why I’m thinking twice about submitting my app. It’s not just about $9.99. It’s about a platform that sets its rules in the dark, for a game I can’t shape.

Feature image generated by ChatGPT

As AI agents take center stage in 2025, a quiet yet profound tension is emerging between the promise of smarter devices, systems, and applications and the peril of eroding user privacy.

Unlike LLMs or other foundational models, AI agents represent a more advanced form of intelligence. It is no longer just about conversations. It is integration, orchestration, and decision making.

“The takeover, the break's over.”

The moment crystallized when the then Honor CEO George Zhao reenacted Steve Jobs’ 2007 latte ordering prank. Except this time, it was real. The AI agent took over the phone and placed an order of 2000 tea drinks on its own, echoing that viral TikTok where someone buys a Porsche by accidentally splashing water on the screen.

@user90tymlsymv赵明荣耀发布会现场用手机AI给观众点奶茶，网友： 3万块上个热搜，这广告费划段。

Tiktok failed to load.

Enable 3rd party cookies or use another browser

In China, this AI agent landscape is more than that, it takes shape along three major lines:

• Device-anchored agents embedded in smartphones, robots, and glasses (just like the Honor phone does);

• Platform-based agents developed by major tech players such as ByteDance, Alibaba, DeepSeek and so on;

• A vibrant ecosystem of startup-driven applications, with Manus among the more visible examples.

Yet the road to a mature agent economy is far from smooth. Technical bottlenecks, uneven development paths, and growing concerns over fairness and privacy have all begun to weigh on the sector's momentum.

In particular, the tension between innovation and individual rights is fast becoming impossible to ignore.

The privacy boomerang in the AI era

As AI agents grow more sophisticated, their appetite for data has expanded exponentially. Tasks that require nuanced contextual understanding have pushed many developers toward increasingly invasive data collection practices.

US based efforts are shifting toward more standardized, protocol driven methods, with an emphasis on privacy preservation.

In China, the trend is particularly stark: rather than building robust, consent-based architectures, many agents have opted to take the leanway in operating system permissions, especially through the abuse of Accessibility Services.

The Chinese reliance on GUI-based screen-reading techniques not only heightens user risks but also signals a broader strategic divergence that may have lasting consequences.

Screen-reading agents: a chilling privacy faultline

The smartphone has become the frontline of this silent privacy war. Beginning in 2024, companies like Honor have adopted practices that allow their AI agents to read screens and simulate user actions without the explicit consent of third-party apps (buy me 2000 cuppa tea and it did).

By misusing Accessibility Services, these agents gain near-total visibility into users' digital lives—from private conversations to banking credentials.

Originally intended to aid users with disabilities, Accessibility Services offer a powerful—and dangerously underregulated—gateway into user data. The risks are not theoretical: the same mechanisms have historically been exploited by malicious actors to siphon billions of personal records.

On desktops, ByteDance's Doubao AI assistant introduced a "Screen Sharing Call" feature in early 2025. Although pitched as an enhancement for user convenience, the feature—allowing AI to read and process live screen content—evokes clear parallels to Microsoft's controversial Recall feature.

Recall, unveiled in 2024, promised to create a searchable visual memory of user activity. Yet it immediately drew widespread backlash over its sweeping scope and potential for abuse. Microsoft was forced to delay its launch, ultimately restricting Recall to an opt in setting with strict local data controls. Nearly a year later, the company has finally begun rolling Recall out to the public, though only on high spec Copilot Plus PCs equipped with powerful neural processing units. The updated version of Recall now includes automated content filtering, enhanced local storage protections, and a clear opt in requirement—a marked improvement over its original form, though still not immune to criticism.

The echoes are unmistakable: China's new generation of smart AI assistants appears poised to repeat the very missteps that provoked such controversy abroad, with even fewer safeguards in place. On PCs, this takes the form of explicit screen sharing and content parsing. On smartphones, it is often less visible but no less invasive, with preinstalled AI features like Xiaomi's XiaoAi or OPPO's Breeno quietly capturing screen content under the guise of convenience.

In some cases, such as OPPO's "Breeno Screen Recognition," users are not even asked for consent before sensitive permissions are activated—a practice explicitly acknowledged in the platform's privacy policies.

What unites these trends is not merely the technical method but a broader disregard for user agency. Whether through overt consent or buried terms, the result is the same: the boundary between user and machine is eroding, often without meaningful oversight.

The new approach

Xiaomi CEO Lei Jun, among others, has called for reform, urging the industry to move toward standardized, auditable interfaces.

In the context of AI agents, this includes two key technical enablers: intent frameworks, which help agents parse user goals and translate them into structured actions within apps or operating systems; and function calls, which allow agents to interact directly with backend services or APIs—bypassing the need for brittle screen scraping or simulated UI interactions.

Yet Chinese tech players have made visible progress in intent frameworks and function calls, many still pair these with accessibility-based workarounds.

Meantime in the US, protocols like Anthropic’s MCP and Google’s A2A are laying the groundwork for a scalable, interoperable agent ecosystem—moving from solo tools to coordinated networks. MCP gives agents the ability to use tools safely; A2A gives them a common language to collaborate. Together, MCP and A2A push agent design from isolated assistants toward an extensible, networked architecture—an essential leap for long-term scalability.

While companies like Baidu and Alibaba have begun exploring MCP style protocols, through initiatives like Baidu’s Qianfan and Alibaba's Model Studio, the pace as a whole remains slow and fragmented. Most implementations so far are partial and exploratory, focusing on peripheral services rather than core transactional functions.

Key consumer-facing platforms such as WeChat, Meituan, and Douyin have yet to meaningfully participate. In many cases, what is labeled an MCP server is simply a repackaged API with limited capabilities, often read-only or narrowly scoped. This reflects not just technical immaturity, but also a deeper reluctance among major platforms to loosen their grip on data, traffic flows, and monetization models.

China’s current reliance on GUI manipulation and system-level shortcuts puts it at risk of missing this deeper infrastructure wave. If Chinese developers want a seat at the table when global agent standards are written, the time to shift course is now.