Summary

OPPO is betting not on any single AI feature, but on rebuilding the mobile experience around “new compute, new perception and a new ecosystem.” Xiaobu’s monthly active users have risen from roughly 150M in June to nearly 170M, spanning phones, watches, tablets and earbuds; hybrid edge-cloud computing lowers inference costs and improves response times, system-level perception fills in context, and third-party services connect through Agents and are orchestrated by Xiaobu.
A handset maker’s most distinctive assets are on-device compute, data stored on the device and the system entry point closest to the user. OPPO puts chat history, information users explicitly ask it to remember, local files and photos into memory, then connects them through search and RAG to Q&A, search and recommendations; sensitive information such as bills, along with the desensitization process, can be handled locally by small on-device models. The product credo is: “The more it remembers about you, the better it understands you.”
Nearly 170M monthly active users are still below Wan Yulong’s expectations, and AI is currently “not a purchase driver.” The negative experience users internalized from Siri since 2011, combined with the need for a phone assistant to handle general Q&A, personalization, thousands of device controls and third-party execution, puts its bar above that of a standalone AI app. For now, OPPO values AI more as a source of post-purchase satisfaction, word of mouth, stickiness and repurchase than as an immediate sales catalyst.
Wan Yulong believes that controlling an app simply by understanding its content is not AI-native enough. A more AI-native direction is to atomize capabilities and expose them to models through some standardized mechanism.
As foundation models converge, handset makers’ key control points are shifting to context, interaction and a closed loop of real-user feedback. Around 2023, OPPO considered pre-training its own foundation model; by the first half of last year, it concluded that this was no longer a capability it had to own. Its in-house work is concentrated on SFT and reinforcement learning atop open-source models, context engineering, and perception-oriented models and small models. “Going alone gets you there faster; going together gets you farther.” For general capabilities, it works with DeepSeek, StepFun, Qwen and others.
Agents will rewrite mobile internet traffic measurement, advertising and value distribution, but the industry has no settled answer yet. If a user invokes an Alipay service through a Xiaobu conversational card, does that count toward Alipay DAU, and how should time spent be measured? When merchants start writing “recommended by DeepSeek,” the answer itself could become a new ad format—but only if recommendation quality and user trust hold.
The long-term form may be one personal “butler” coordinating multiple vertical “advisers,” while a truly AI-native OS has yet to emerge. The butler knows the user best and handles intent understanding and service distribution; experts in law, medicine, travel and other fields execute vertical tasks. Given its compute, usage time and role as the device hub, the phone could remain a major butler platform. Wan Yulong’s unanswered question is whether the industry can move beyond layering Agents onto traditional Android and existing app interfaces, and grow a new generation of operating system and application ecosystems from the ground up.

Deep dive

1. After nearing 170M MAUs, OPPO defines its next destination as a personalized AI OS

Wan Yulong has worked in AI since 2008. Xiaobu now spans OPPO phones, watches, tablets and even earbuds, with monthly active users nearing 170M; he believes that puts it among the more active Android voice assistants by scale.
Ahead of the 2025 ODC on October 15, OPPO introduced the “personalized AI OS,” organized around three strategic directions: “new compute, new perception and a new ecosystem.”
The industry consensus is straightforward: every device maker has on-device compute, personalized data and a user entry point. The real differentiation is tactical—how to collect context, build user habits and bring services into the system.

2. The value of new compute is lower cost, faster response and keeping sensitive data off the cloud

Wan Yulong notes that models have become both more capable and smaller over the past 2 years. Inference that once required the cloud is gradually moving onto phones, allowing edge-cloud collaboration to lower inference costs and shorten response times at the same time.
When Cheng Manqi asked about payment and financial privacy in expense tracking, Wan Yulong pointed to small on-device models: sensitive information and the desensitization process can both be handled locally.
This is a structural advantage for device makers over Web applications: the phone provides compute while holding hardware and software sensor signals, allowing it to fill in model inputs without sending all raw information to the cloud.

3. OPPO uses a physical button to build the muscle memory of “remembering”

Wan Yulong says the entire industry is working on memory, but OPPO “should be the first” to give it a dedicated hardware entry point. A physical button lowers the activation threshold and turns “remember anything, anywhere” into muscle memory.
One-Tap Flash Memory 1.0 mainly understood screens and images; the new version extends to video, voice, bills and pickup codes. Long videos can be distilled into key points, to-dos and calendar entries, while fleeting ideas or family reminders can be converted from audio into structured information.
Expense tracking can aggregate and analyze bills from different apps, eliminating manual entry into Excel or an expense-tracking app. Food-delivery pickup codes can appear on the Fluid Cloud display, after which users can simply ask Xiaobu, “What’s my pickup code?”
OPPO wants to connect more than the act of saving something. It wants a complete chain linking remembering, asking, searching and recommending. Wan Yulong summarizes the product logic as: “The more it remembers about you, the better it understands you.”

4. The system-level assistant now sees screens, sees the physical world and proactively produces briefings

One-Tap Ask Screen lets users ask directly about the current app, article or photo. Real-world Q&A is closer to the Astra Google previously demonstrated: users open the camera and ask about what is in front of them.
To work in noisy environments such as museums and streets, OPPO added voiceprint recognition so the assistant responds only to the person who initiated the interaction. “Tap where you don’t know what to tap” lets users select an object on screen and have a large model explain what it is.
Personalized briefings combine the day’s schedule, weather and saved memories to surface agenda highlights and unexpected reminders. Wan Yulong groups screen Q&A, memory search and proactive recommendations under the label “system-level AI assistant.”
On the photo side, AI relighting, glare removal and passerby removal represent another path: operations that once required professional image-editing software can now be completed with a single sentence.

5. OPPO’s definition of memory goes well beyond chat history

Wan Yulong divides memory into 3 categories: chat history with Xiaobu, content the user explicitly asks AI to remember, and files and photos stored on the phone. The last category is content already saved on the device, and in Chinese can also be understood as “memory” in the storage sense.
That means questions such as “How high was my blood lipid level in my medical report?” and “What was my weight at my last checkup?” can become personalized Q&A, rather than merely repetitions of things the user previously said in conversation.
For “Find the photo of me and my child,” his technical explanation is on-device search plus RAG. The photos are already memories captured by the user’s camera; the system needs to retrieve and understand them, not create another chat memory.
Version 1.6.0’s personalized Q&A brings these sources together. Compared with ChatGPT’s heavier emphasis on conversational memory, OPPO wants to make use of the richer data on a phone—data that is also closer to everyday life.

6. Context engineering turns a phone’s “How do I do this?” into a complete intent

Models are becoming more powerful and context windows longer—key technical drivers behind AI’s acceleration in Wan Yulong’s view. Over the past 6 months, industry discussion has also shifted markedly from prompt engineering to context engineering.
His central analogy is that a model remains an information-processing unit, and the quality of its output depends on what information it receives. The query is only part of the input; the app currently open, the screen content and historical memory also determine the answer.
The same “How do I do this?” means “What’s the recipe?” when a dish is on screen, but “How do I solve it?” when a problem is displayed. Cheng Manqi added that she sometimes screenshots a Duolingo question and asks only, “How would you do this?”—and the model can still recover the intent from the image.
For handset makers, hardware sensors, software state and personal data together form context. That is why “new perception” means more than simply adding cameras or microphones.

7. As model costs fall, competition is shifting from pre-training to application innovation

Looking back at 2023’s “hundred-model race,” Wan Yulong says many AI companies wanted to pre-train models with tens or hundreds of billions of parameters. Since then, model suppliers have gradually converged, while costs have fallen “by dozens or hundreds of times every year.”
As the models used behind startups consolidate around a few providers, differentiation is increasingly coming from interaction, context engineering and understanding user scenarios. “How to use models well” is replacing “whether you own a model” as the core question.
With model counts converging and costs falling, more innovation is now centered on interaction, context engineering and pain points in specific scenarios.

8. The rise from 150M to 170M still falls short of expectations, while legacy voice assistants have left a mental debt

Cheng Manqi noted that Xiaobu’s publicly disclosed MAUs rose from roughly 150M in June to nearly 170M 4 months later. Wan Yulong, however, was blunt: “Objectively speaking, this figure is actually below my expectations.”
He used the DAU ratio to illustrate the gap: if a product is installed or activated on 100M devices and 10M are active each day, that is 10%. Phone makers use all activated buyers as the denominator, leaving their DAU ratio less impressive than that of leading AI applications.
Siri launched in 2011 and Xiaobu’s first version in 2018, but assistants delivered poor experiences for years before large models arrived. Wan Yulong did not put the blame solely on Siri: “We didn’t do a good job ourselves either.” Users consequently developed the habit of “If you’re not useful, I won’t use you.”
He believes Xiaobu is now fairly close to standalone AI apps with strong word of mouth, though hallucinations and wrong answers remain. Many users still instinctively install and open a separate app, sometimes forgetting that their phone already has an assistant.

9. A phone assistant has to pass 4 exams at once, so its passing grade is naturally higher

Wan Yulong breaks the task into 4 categories: general Q&A, personalized Q&A based on device information, on-device control, and execution across third-party apps. The device-control commands alone may number in the thousands.
“Where should I stay in Chengdu?” marks the boundary between personalized and general Q&A. If the user previously saved an itinerary and hotel for Chengdu, the assistant should retrieve that hotel rather than recommend another list of accommodations.
“Navigate home” requires the system to launch Amap or Baidu Maps, enter the address and trigger navigation. A standalone chat app does not have to manage this chain, but a phone assistant can be judged “really stupid” if any link fails.
As a result, incomplete capabilities directly suppress penetration. Once users view it as the phone’s intelligent assistant, they will not accept one that is good only at chatting but cannot set an alarm or turn on Bluetooth.

10. AI products are not designed into existence; user feedback pushes them toward expectations

The boundaries of a traditional calendar app are clear: it is easy to evaluate whether adding or deleting an event works well. An AI product, by contrast, is “hard to be by design” at the start because the same input can correspond to entirely different real-world needs.
Wan Yulong used “Lin Daiyu pulling up a weeping willow” to illustrate ambiguity. The user might want to write a novel, fact-check the claim or simply make a joke. A model can answer the proposition without understanding why the user raised it.
His product judgment is that “omnipotent” does not mean “all-understanding.” Launch first, observe how users actually use it, then refine interaction, visual design and strategy around high-frequency needs—that is how an effective feedback loop forms.
This makes AI experience like “chasing a horizon.” A product can iterate from 60 or 70 points to 80 or 90, but the user’s expectation of 100 keeps moving forward.

11. The more AI-native direction is to atomize capabilities

Wan Yulong believes that controlling an app simply by understanding its content is not AI-native enough.
The direction he favors is to atomize capabilities and expose them to models through some mechanism. The view is consistent with his broader understanding of AI products: systems need to reorganize services around user scenarios and model capabilities, rather than simply preserve existing software boundaries.

12. OPPO uses productivity, lifestyle and imaging scenarios to constrain technical impulses

Wan Yulong starts from the premise that “AI is just technology”; the core remains the user problems on a phone. Since 2023, OPPO has repeatedly identified which experiences AI can genuinely change, rather than starting with a model and forcing it to find a use case.
The 3 scenario groups map to 3 product positions: study and work as a “productivity assistant,” information, schedules, services and expense tracking as a “life butler,” and taking, managing and editing photos as an “imaging companion.”
Technology is evolving in parallel: generative capabilities boost productivity, RAG and search improve information access, and multimodal models lower the barrier to image and video creation.
Cheng Manqi cited Sora 2 to illustrate the mobile opportunity: phones can support richer queries, creation and content consumption. Wan Yulong emphasized that unified generation of video, voiceover and sound effects allows the phone to support more complete, immediate creation.

13. Expense tracking was not invented in a conference room; user research corrected the leader’s intuition

Since early last year, OPPO has organized “origin trips” to “go into the front line and go deep into the front line,” bringing real users, sales staff and product teams together to discuss phone pain points, high-frequency AI scenarios and unmet needs.
The demand for expense tracking came mainly from college students and young people just entering the workforce. Wan Yulong admits he does not track expenses and “didn’t really understand it” at first, but repeated common requests across multiple interviews changed his view.
The team also watched users request features under the company spokesperson’s Weibo posts, then co-created with users: how they track expenses, where they get stuck and whether proposed implementations would be acceptable. Small features entered repeated validation instead of being approved as one-off projects.

14. Strategy product managers and a weekly release cadence are becoming new infrastructure for AI R&D

Traditional user product managers still own requirements and interfaces, but OPPO has added “strategy product managers” who continuously analyze online logs and implicit feedback such as likes, dislikes, copies, shares and multi-turn conversations.
These data help determine which answers are more worth sharing, what strategies extend conversations and which questions users actually ask in sequence. Product design is therefore shifting from a static feature list to a behavioral loop.
Wan Yulong says the team moved from releases roughly every month and a half around 2021 to a fastest cadence of once a week by the beginning of this year. After DeepSeek launched, OPPO was among the earlier handset makers to integrate DeepSeek R1 and rapidly rebuild its online conversation and multimodal Q&A experience.
“Resources are always limited.” The response includes breaking requirements into smaller pieces, compressing schedules and concentrating effort on key battles. He estimates that AI teams at mainstream handset makers typically number in the high hundreds, and are still catching up with mature imaging organizations.

15. AI has not yet sold more phones; for now it looks more like a long-term retention investment

Wan Yulong’s judgment is clear: at least for now, AI “does not yet constitute a purchase driver.” Most users will not choose one phone over another simply because its AI is better.
He calls AI a form of “after-sales experience”: only if users find the device useful after purchase can it generate word of mouth, usage stickiness, brand affinity and willingness to buy again.
OPPO therefore evaluates AI on a long-term horizon rather than demanding that every capability convert immediately into sales. What needs to accumulate is sustained usage of the brand, device and multi-device ecosystem.

16. OPPO abandoned the full-stack fixation, keeping in-house work for areas that can create a feedback moat

Wan Yulong explains the partnership strategy with “Going alone gets you there faster; going together gets you farther.” A phone is already an integration of chips, cameras, batteries and other capabilities; in the AI era, there is likewise no need to build every layer in-house.
OPPO did consider pre-training at the time, but by the first half of last year concluded that it was “not a key control point” for a handset maker. General-purpose models can come from industry solutions such as DeepSeek, StepFun and Qwen.
In-house work remains concentrated in 3 layers: SFT and reinforcement learning on open-source models, context engineering around products, and proprietary work on perception-oriented models and small models.
The test is whether the work can use real online data to create a feedback loop. If OPPO’s data and distribution can keep a model ahead of general-purpose solutions over time, while training costs remain manageable, that layer is worth continuing to build internally.

17. Wan Yulong chose phones because AI needs a long-cycle feedback field at hundred-million scale

From 2015 to 2021, Wan Yulong worked at Alibaba across iDST and DAMO Academy, including in a speech lab. He worked on AliOS phones, intelligent cockpits, robots, far-field TVs and Amap’s voice assistant.
After working across multiple device types, he concluded around late 2020 and early 2021 that AI is not software that ends when it goes live. It needs to absorb feedback over time and move continuously from 60 or 70 points toward 80 or 90.
Phones are closest to people, accompany them for the longest and offer relatively strong compute; users also complete the widest range of tasks on them. These traits support the continuous accumulation of data and feedback.
The phone is also the information hub for watches, earbuds, tablets and even car keys. Beyond upgrading the experience on one device, AI can start from the phone to coordinate across devices and extend system and sensor capabilities to new hardware.

18. Different players appear to be fighting for the same entry point, but their underlying “why” is different

Wan Yulong divides the competitors into super-app companies, handset makers and foundation-model companies. His basic view is that the AI market is large enough for each to create different value, just as phones and apps did in the mobile internet era.
Internet companies exploring phones may not actually want to become phone brands; they may see phones as a “training ground” for OS, hardware-software integration and new devices. He uses AliOS as an example: its goal at the time was to enable handset makers to port the system, not to create an “Alibaba phone.”
OPPO’s why for AI is improving the phone and multi-device experience. ByteDance’s motivation is “most likely” closer to content, services and the CapCut ecosystem. DeepSeek, meanwhile, rarely emphasizes its own app.
Similar whats—an assistant, a model or hardware—therefore do not prove that the end states will be the same. Understanding the value chain each company truly wants to control matters more than looking at what it has demoed.

19. After Alipay connected to Xiaobu, DAU, time spent and advertising all need new definitions

Wan Yulong uses commercial history to reason through service boundaries: consumers once bought from individual stores, then Taobao and Tmall aggregated merchants, and phones put every app onto one device. AI may compress the service entry point further into a single sentence.
OPPO has already partnered with Alipay and others through Agents. Xiaobu identifies the user’s intent, hands the request to the partner’s Agent and returns the service in a conversational card. “The assistant doesn’t necessarily need to know how to make hotpot”; it only needs to find and orchestrate someone who does.
That immediately prompted Cheng Manqi’s question: if a user does not tap the Alipay icon but uses an Alipay service inside Xiaobu, does it count toward Alipay DAU? How should time spent be measured? Wan Yulong acknowledged that the measurement logic may change in the Agent era.
Value distribution and business models remain unsettled. A merchant writing “recommended by DeepSeek” suggests that a trusted answer itself could become an ad. But if paid results damage quality and users stop trusting the assistant, the model cannot work.

20. Mobile Agent protocols remain on multiple tracks; Android standardization is still only a prediction

API and A2A are not mutually exclusive. Wan Yulong sees an API as one interface for Agent communication, while A2A adds a standardized information-transfer method on top. MCP is relatively mainstream on the Web, but mobile has yet to reach a similar consensus.
For now, each handset maker is largely building its own system. Apple has its own App Intents Framework; Google has 2 or 3 protocol stacks, including an app-intent framework and A2A, but not every Android developer will follow them.
Wan Yulong believes the Android camp will at least eventually form a unified protocol, perhaps through an alliance of multiple manufacturers similar to the quick-app era, alongside policy-level standardization. For now, however, “several systems are all running.”

21. The end state looks more like one butler plus multiple advisers, with the phone still the primary entry point

Wan Yulong divides intelligent assistants into 2 types: the butler, which knows the user best, accompanies them the longest and handles intent understanding and distribution; and expert advisers in law, medicine, travel and other fields, which go deep on vertical tasks.
If a user says, “I want hotpot,” the butler knows they do not eat cilantro or spicy food, then passes those preferences to the dining service. In theory, a frequently used device needs only one butler plus multiple advisers the user chooses to subscribe to.
He does not expect one “unified” butler to span every device and brand. Phones, cars and other devices may continue to have their own assistants, coordinating only on some tasks. The phone can close the loop on more services and serves as the hub for other devices, so it could become an important butler platform.
In the near term, glasses remain sensory extensions of the phone. Their app ecosystems, compute, battery life and input are not sufficient to support heavy-use activities such as video feeds, gaming and communications. They can strengthen the connection between the digital and physical worlds, but have yet to challenge the phone’s position.

22. The biggest obstacle to voice interaction may be generational habit, not technology

Wan Yulong notes that some people feel awkward giving AI commands in public but do not mind making a phone call. Bystanders may not even hear the difference; the barrier is more psychological—the feeling that “I’m talking to a machine.”
Children already use voice naturally when watching TV and searching for content. He calls them “the AI-native generation.” As they grow up, a more natural mode of expression may no longer feel unusual.
Input methods offer a useful analogy: even with a nine-key keypad or full keyboard, some parents may still insist on handwriting. This generation is accustomed to keyboards; the next may be accustomed to voice. Interaction habits typically migrate more slowly than capabilities come online.

23. AI is changing interaction and the app ecosystem, but it has not had its App Store moment

Wan Yulong summarizes the first iPhone’s leap as 2 things: the “revolutionary UI” enabled by multitouch, and an app ecosystem that put previously impossible services onto the phone.
Large models are likewise changing 2 layers: a sentence can retrieve information, create content or reach a service, while application production may shift from coding to prompting. What Sora 2 demonstrates is a highly personalized content loop in which models create content and users consume it.
But he believes the Agent ecosystem has not reached App Store maturity. It needs a genuine breakout product—something like WeChat red packets or Taobao’s low-price strategy—to create a clear mechanism that reinforces developers, consumers and traffic.
The biggest question is “What does a truly AI-native OS actually look like?” Today’s browser use, mobile use and GUI Agents still sit on traditional Android or existing interfaces. In Wan Yulong’s view, an OS for the AI era must be different at the foundation in order to produce a different product ecosystem.
OpenAI’s shift from a single model toward a system, with GPT-5 using a router to select different models, offers him one possible direction: the future core may be a system mounting multiple expert models. But he admits he has “not seen any particularly clear signs” yet, and hopes more people will build a new underlying ecosystem the way Android was built back then.

E211 | At the Intersection of Content Creators and Robots: The Evolution of 3D Digital Humans 20VC: Why VC Today Is Worse Than 2021 & Why Vertical SaaS Is a Bad Bet