Summary

If ranked solely as an independent Chatbot, Yuanbao has a weak presence. Viewed through distribution, data, and service entry points, Tencent is effectively “opening another table.” 庄明浩 argues that pure language-model gaps are narrowing, while Chatbot retention, time spent, and monetization remain underwhelming; 刘飞 notes that once Yuanbao is embedded in WeChat, it can connect to official accounts, QQ Music, and Tencent services through near-zero-friction actions such as pinning it or “@Yuanbao” in comment sections. The competitive objective is no longer simply to win one App.
AI’s main battlefield has shifted from model scores to closed-loop use cases, and the most practical path has shifted from “AI native” to “something plus AI.” As benchmarks lose relevance and model iteration slows, research, website and App creation, and even better meeting experiences have become contests in “language to action.” If a single AI feature in Tencent Meeting improves paid conversion and supports continued investment, it is already “ready” and “working.”
Tencent’s internal AI rollout has reached advertising, R&D, and game-production pipelines. The episode cites its Q3 disclosures: advertising revenue rose 21% year on year to more than RMB36B, 90% of its programmers use CodeBuddy, and roughly 50% of newly added code is AI-generated. Tencent Interactive Entertainment’s IEG is also providing extensive AIGC and AI-plus-gaming functions and modules; whether this is necessary and whether the ROI is positive no longer needs to wait for debate.
As model capabilities converge, proprietary data, search know-how, and existing user contexts are again becoming the assets that determine experience. 刘飞 now routes tasks by use case: Doubao for basic concepts, Yuanbao for Chinese research and official-account material, IMA for his personal content library, and Gemini more often for overseas sources. 庄明浩’s view is that WeChat’s “enormous lead” in the quality of Chinese content, especially official accounts, will affect model outputs.
The signal from Gemini 3 Pro and Nano Banana Pro is not another benchmark lead, but the beginning of language, multimodality, and Agent operating as one system. Generating a photo or standings table for the F1 race that just ended requires finding the latest race, checking drivers and teams, organizing the results, and rendering them correctly. Image generation thus moves from attractive showmanship to a credible scenario tool—and Scaling Law has not simply stopped working.
The most important reassessment for Hunyuan is not its pure language-model ranking, but the coupling of images, video, 3D, and world models with Tencent’s gaming business. 庄明浩 calls Hunyuan 3D one of the leading products in China and expects Hunyuan World Model 1.0 to have launched in September 2025. Pure language-model competition has “very poor cost-effectiveness,” but Tencent has the capital, resources, scenarios, and patience to remain at the main multimodal table.
Tencent’s capital strategy is to have the money without needing CapEx figures to prove its commitment, preserving the option to “attack on the front foot and defend on the back foot.” In the first half, the market scrutinized DeepSeek, compute spending, and the arms race; when that narrative mattered less in the second half, Tencent could move at its own pace. The private market remains firmly committed to AI applications, while public markets favor chips, energy, storage, and other hardware narratives. Software opportunities have not disappeared; they are simply quieter.
This “relaxed mindset” is not complacency; it is a return from FOMO to verifiable short- and medium-term goals. Two years out—even “two months” out—is hard to envision. The more durable framework is to steadily improve existing businesses, follow frontier-model capabilities, and explore new paradigms at the same time. 刘飞 also believes Tencent may already have planted 10 or 20 seeds through investments. It looks more like a complex system with abundant options than a single “stick” that can be wagered on only one path.

Deep dive

1. Chatbot’s dominance obscures the battlefield where Tencent is actually competing

庄明浩 first rewinds to the third anniversary of ChatGPT’s launch. Over the past 3 years, Chatbot has been so dominant in functionality and product definition that when many people talk about AI, they are really talking about Chat-based LUI rather than the broader AI product space.
On this narrow battlefield, Yuanbao does look more like a follower: Doubao was previously strong, DeepSeek broke out in early 2025, and Alibaba began heavily promoting 2 related products toward year-end. 庄明浩 cited a judgment from a Q3 earnings call: among pure Chatbots and natural-language models, “it is very difficult for anyone to be absolutely much stronger than another competitor.”
The more important caveat is that even after 3 years of development, Chatbot retention, time spent, and future monetization remain “not that ideal.” Yuanbao’s limited participation therefore need not be equated directly with Tencent falling behind in AI; it may instead mean Tencent sees the short-term milestones of a traditional standalone App as less important.

2. Yuanbao inside WeChat could become a low-friction connector, not another chat App

庄明浩’s preferred experience for ordinary users is not downloading a new application, but “pinning Yuanbao inside WeChat.” Users can also call AI in official-account comment sections. The steps are nearly frictionless, and “@Yuanbao” is already becoming habitual for some users.
刘飞 takes the idea further and defines Yuanbao as an ecosystem connector. On the surface it is a Chatbot; behind the scenes it may search official accounts, play QQ Music, read Tencent content, and continue into services inside WeChat. It is closer to the entry point WeChat’s mini-programs once tried to establish, and could even become a new operating-system interface for Tencent’s ecosystem.
This is what 庄明浩 means when he says Tencent “has the credentials to do this—it is effectively opening another table.” Instead of fighting by the old standalone-App and Web rules, Tencent can use WeChat’s existing social graph, content, and services to drive AI penetration. Yuanbao’s core value may therefore lie less in the answer itself than in turning an information request into a service action.

3. In the second half of 2025, model scores gave way to scenarios and behavior

庄明浩 summarizes this year’s shift as “from technology to applications, from language to behavior.” As benchmarks begin to fail, the industry no longer knows how to evaluate pure-model quality consistently, so scenarios and user experience naturally move up the priority list. The duration of this trend is uncertain, but it was already more visible in the second half of 2025.
“Behavior” is much broader than chat. Research is behavior; building websites and Apps is behavior; embedding AI into a specific scenario to complete a function is behavior. Industry discussion now includes the move from so-called L1 and L2 systems toward Agent, with the key question no longer whether a model can generate natural language, but whether it can call tools and deliver results.
刘飞 adds that large models have not hit a ceiling, but the acceleration of performance improvements is slowing. As the core bottleneck shifts to scenarios and applications, product breakthroughs matter more than another benchmark fight. Several smaller companies have already shown that Chatbot’s “high trial, low retention” and near-zero switching costs make it irrational to buy activity through marketing alone.

4. The positive feedback loop comes from “something plus AI,” not abstract AI native

庄明浩 uses Tencent Meeting to illustrate the smallest viable loop. If a new AI feature improves the product experience and raises revenue or paid-user conversion, the team receives clear positive feedback, keeps investing, and moves closer to user needs. For that team, the feature is “ready, which means it works.”
刘飞 acknowledges that in his first 2 years he focused more on which industries AI native products would disrupt from the ground up. He gradually shifted toward the idea that it is less “AI plus what” than “what plus AI.” Much of the value comes from better productivity and service experience, without requiring a radically different user interaction.
This explains why early wins are appearing more often at the application and scenario layers. Technology does not need to create an entirely new production setting first; improving efficiency inside existing workflows such as documents, PPTs, code, and meetings already has product and commercial significance.

5. AI conversion of existing internet businesses is enough to support the first wave of massive demand

庄明浩 relays 黄仁勋’s breakdown of the “three orders of magnitude of opportunity.” The first layer is the comprehensive AI conversion of existing internet systems including Google, Amazon, Meta, Microsoft, Tencent, Alibaba, and ByteDance. The second is extending those AI-enabled businesses forward. Only the third is the idealized AI-native layer.
The importance of this ordering is that even if the third layer takes years to arrive, the first already covers hardware, data centers, model capability, and internal enterprise ROI. 庄明浩 says bluntly, “let’s not consider that ideal AI native layer for now.” The existing conversion opportunity is already enormous.
A more direct piece of evidence is that token-consumption growth at leading cloud providers is being driven mainly by their own businesses, not third-party APIs or newly created AI-native companies. The slope of that curve was still rising over the past 2 or 3 quarters, implying that near- and medium-term demand does not need the next Killer App to become real.

6. Tencent’s internal AI rollout has reached advertising, R&D, and game-production pipelines

刘飞 cites Tencent’s latest Q3 information: supported by AI and the WeChat commercial ecosystem, advertising revenue rose 21% year on year to more than RMB36B. The logic is not selling one more AI product, but improving ad-delivery efficiency and advertiser ROI, thereby expanding an existing revenue pool.
The R&D figures make the “underwater rollout” even clearer: 90% of Tencent’s programmers now use CodeBuddy, and roughly 50% of newly added code is AI-generated. 庄明浩 says this trend is already “charging ahead and lying inside your production pipeline.” There is no longer a need to debate whether it is worth starting.
Gaming is not a sandbox either. 庄明浩 says Tencent Interactive Entertainment’s IEG is among the internet-company systems in China providing the most functions and production modules for AIGC and the combination of AI with gaming. Given the revenue scale of gaming, advertising, financial services, and To B services, internal efficiency gains could have especially large absolute benefits for Tencent.
刘飞 draws a boundary: small and midsize companies can benefit as well, but large companies have a larger business base, more data, and broader organizational coverage, so the absolute impact appears earlier and more clearly in financial statements. This is not technology exclusive to large companies; existing scale magnifies the same efficiency dividend.

7. To C is not a fake demand; the mobile-internet end-state metrics have simply stopped working for now

Responding to the criticism that there is no Killer App, 庄明浩 is direct: Chatbot is one of the fastest product forms in history to cross multiple user milestones, and many people around him already use it in place of search engines. Adding up Chatbot users globally already produces an “astronomical number.”
The real problem is that the industry has spent more than a decade using mobile-internet retention, time spent, and conversion to measure everything. These are “overly final answers.” When they failed to explain early AI applications, investors fell back to ARR, only to run into disputes over gross margin and revenue accounting.
庄明浩 argues for stepping back again and asking the most basic questions: what exactly is the user need, how is it being met, and which metric proves that it has been met? Current signals may be accelerating token consumption or reaching a very large narrow user group in a very short time, rather than immediately achieving the retention structure of a mature App.
A wave of new applications in the second half of 2025 also benefited from a change in founder capabilities. These entrepreneurs have deep experience in their original industries and have lived through 3 full years of the AI wave, giving them better entry points, initial features, and operating rhythms—and making it easier to win users and market recognition.

8. As models converge, data and search know-how become the experience differentiators

刘飞 no longer treats every model as a candidate answer to the same exam. He favors the fast Doubao for basic concepts, uses Yuanbao more for Chinese research because official-account material is relatively accurate, uses IMA to organize his own past content, and turns more often to Gemini for overseas sources because its breadth and perceived accuracy are better.
庄明浩 uses this to reassess the 3 elements of data, algorithms, and compute. Ordinary users struggle to perceive differences in algorithms and compute directly, but increasingly notice differences in data. Once models stop separating themselves by a wide margin, accumulated data, search preferences, and familiarity with user habits can produce stronger marginal effects.
In his view, Gemini 3 Pro’s reputation comes not only from the model itself, but from Google’s search know-how built over many years: understanding what a keyword or passage is really seeking, determining what counts as a good result, and gradually encoding that judgment into model outputs and AI search modes.
The corresponding domestic asset is WeChat content. 庄明浩 calls official accounts’ advantage in the quality of Chinese-internet content an “enormous lead,” and believes it affects Yuanbao’s research experience. He adds a time-based hedge: the difference is already clear to high-frequency users, but it may be too early for ordinary users to form a stable perception.

9. Gemini 3 Pro merged language, multimodality, and Agent from 3 tables into 1

庄明浩 reviews the release sequence this time: Gemini 3 Pro on day 1, Nano Banana Pro on day 2. People do not pay much attention to leaderboards anymore; the market relies more on the feel of actual use. The key message from Google’s official communications was that language, multimodality, and Agent are fundamentally one thing and should not be discussed at separate tables.
The example that best carries the argument is asking Gemini to generate “an award-ceremony photo from the F1 race that just ended, or the standings table from that race.” It must first search for the latest race, obtain the result, align driver names, faces, and teams, and then complete the layout and image generation. Miss any link and the result is merely attractive, not credible.
This is why the wave of infographics, explanations of Transformers using Doraemon, and photosynthesis handbooks is not merely social-media showmanship. Users previously would not have assigned such tasks to image models because they lacked basic world knowledge. Now that search, reasoning, organization, and generation form a chain, the tool has moved from “how pretty” to usable.
刘飞 compares this with the early end-to-end breakthroughs of large language models. NLP once studied tokenization, part of speech, and syntax separately; later, “scaling up produces miracles” was found to solve the subtasks simultaneously. 庄明浩 speculates that if the unreleased latest version of Veo 3 is strong enough, text, images, video, Coding, and Agent will converge further into a “unified system.” That is an expectation, not a verified conclusion.

10. Hunyuan’s underestimated battlefield is 3D and world models, not pure language rankings

庄明浩 believes attention allocation has created a cognitive bias. Overseas discussion concentrates on OpenAI, Claude, Google, and Elon Musk’s camp; domestic discussion centers on DeepSeek and Qwen. Progress by ByteDance, Baidu, Tencent, Xiaomi, Bilibili, Xiaohongshu, and Weibo—each selecting and extending models around its own business—receives far less attention.
In multimodal segments, he says Hunyuan’s image and video capabilities have consistently been solid, while 3D is one of the leading products in China. 3D is naturally adjacent to gaming, and leading game companies are more determined to invest than ordinary vendors. Tencent therefore has both a technical motive and a clear internal consumption scenario.
The next step is the world model. 庄明浩 recalls that Hunyuan World Model 1.0 should have launched in September 2025 and continued iterating. He expects that even if this path takes longer than language models to reach a “GPT moment,” Tencent could still go deeper through business integration and long-term patience.
“The cost-effectiveness of pure language-model competition is already very poor” is 庄明浩’s explicit tradeoff. But from voice, images, video, and 3D to world models, Tencent has the credentials to remain at the main table. 刘飞 adds the financial point: multimodality consumes more capital and resources, and Tencent’s profitability and free cash flow are sufficient to clear that threshold.

11. Google’s counterattack shows that AI has not wiped out the old resources of large companies

庄明浩 notes that during this round of competition, Sam Altman issued a temporary “red alert” and demanded renewed focus on improving ChatGPT and GPT. OpenAI had previously split its technology and product teams after GPT-5 and pushed multiple products, advertising, and other monetization efforts. Model competition has now pulled attention back to the core battlefield.
He had originally believed that by November 2025, OpenAI had temporarily resolved its Microsoft relationship, organizational structure, financing and IPO, talent flows, and the question of whether it was a product or technology company, allowing it to “leave the past behind and move forward.” Google’s condition in this round reminded it that the old battlefield is not over, and there is not yet enough capital to pursue every new experiment freely.
For the industry, this is good news: competition may accelerate model progress again. For OpenAI itself, it is “really too difficult.” Unlike a mobile-internet company that can secure its territory after a land grab, OpenAI remains on a rapidly shifting front involving models, multimodality, compute, data, and distribution.
刘飞 therefore revises his earlier view: AI may not invalidate the accumulated advantages of incumbent giants; it may reinforce their resources, compute, users, and scenarios. People once said, “OpenAI emerges and Google dies.” Now the reversal is that “Google is taking OpenAI’s life.”

12. WeChat’s all-in-one ecosystem lets Tencent move slowly instead of sprinting

庄明浩 interprets the application-calling capabilities shown at OpenAI’s developer conference as an attempt to build all-in-one functionality inside the ChatGPT dialogue box, with third-party services called up like mini-programs. This is a product structure that Meta, Google, and Microsoft failed to establish during the American internet era, while WeChat formed it long ago.
WeChat may therefore be the company least pressured to make “one giant leap.” Embedding Yuanbao into small scenarios such as comment sections is already enough to improve experience, retention, and AI penetration at modest cost. Tencent does not need to first replicate a super-Agent covering every service to prove the direction is right.
But 刘飞 stresses that an existing entry point is not an automatic victory. How the browser, Tencent Meeting, input method, QQ Music, and IMA each become AI-enabled are “extremely complex and personalized choices.” The same model can improve the overall product when integrated well and lack market competitiveness when integrated poorly. The question cannot be reduced to whether “Tencent AI is good or bad.”
庄明浩 uses his daily experience with IMA to illustrate the granularity of these decisions. Product managers must decide whether to follow NotebookLM’s new features, which roadmaps to adjust, and whether a competitor’s move merits a response. The work is already “attached to what you need to do every day,” rather than beginning with a grand AI blueprint handed down all at once.

13. Tencent replaces the CapEx arms race with the ability to attack and defend

In the first half of 2025, DeepSeek, compute investment, and the US arms race kept analysts focused on CapEx at Tencent, Alibaba, and Baidu. Chinese companies also needed to explain their AI commitment through narratives that listed-company analysts could understand. 庄明浩 observes that by the second half, this narrative had become less important for Tencent.
The reason is not a lack of money. Tencent has cash and no cash-flow crisis. As the efficiency of converting internal businesses to AI improves, it does not need to prove itself by stacking more GPUs every quarter than expected; it can set the investment pace according to actual scenarios, returns, and the stage of the technology.
庄明浩 calls this an “attack-on-the-front-foot, defend-on-the-back-foot” option and sees it as consistent with Tencent management’s usual style. Investors who value capital discipline will buy it; more aggressive investors may prefer larger spending figures. Tencent has the capital to take risks when it chooses, not indifference toward AI.
刘飞 compares the process with renovating an old house. For now, AI is more like infrastructure that continuously upgrades the experience of existing scenarios than a force immediately demolishing the entire building. If an AI operating system or genuinely AI-native paradigm appears in the future, Tencent can increase investment. Until then, there is no need to borrow against a distant future.

14. The “relaxed mindset” comes from replacing FOMO with verifiable short- and medium-term ROI

刘飞 recalls that when Yuanbao’s promotion was especially aggressive more than half a year ago, Tencent insiders were still visibly anxious about hoarding GPUs, building technical capabilities, and choosing a path. Recent conversations instead convey a “particularly top-down sense of relaxation.” This is not complacency; it is the company recovering its familiar product and operating rhythm.
庄明浩 believes Tencent, Alibaba, and other giants must still treat AI as the central issue, and continue tracking pure models even when their connection to the business is temporarily weak. But management must become increasingly explicit about what needs to be achieved in the short term, what is visible over the medium term, and whether periodic positive feedback and ROI meet expectations.
For Tencent, the more stable framework consists of 3 things: steadily improve existing businesses, follow the technical capabilities of leading models, and look for new AI paradigms in new businesses. It has not promised the most aggressive route to the summit, but it can adjust its path during the climb based on supplies, position, and feedback.

15. Applications still have room, but investors and product teams must abandon the illusion of an end state

The private market became more committed to AI applications in 2025. Model-investment barriers are rising and the number of eligible participants is shrinking, while some applications have achieved stage-specific positive feedback in users, revenue, or niche influence. Public markets rotated through chips, energy, storage, and other themes. Software is recognized, but its narrative is less direct than hardware demand.
Faced with the question “what if the giants build it too,” 庄明浩 points out that “the giants” now include Chinese incumbents, OpenAI, and Google. Models are like a rising tide: some companies will be swallowed, while others will build products into boats and rise with the water. The real difference in risk is whether a company is merely level with the horizon or has already gained enough elevation to stay safe.
One startup example he gives may originally have been a voice company. When image capabilities arrived, the team added image features and bundled them into a paid membership. Within a few months, it grew from 2 people to 10, while monthly revenue rose from $100K to $1M. Execution, creativity, recombination, and information asymmetry were enough to support the next financing round.
刘飞 and 庄明浩 ultimately break the “moat” into a bundle of capabilities. A founder’s adaptability, industry understanding, judgment on technical timing, business model, promotion, and personal brand can all become barriers to entry. The ultimate structure cannot be falsified in the short term, so it cannot be used to conclude that nobody should invest today. As the “AI product manager” paradox shows, what matters is understanding the real needs of meetings, comment sections, or agriculture—not adding AI for show.
Asked about the world 2 years from now, 庄明浩 first answers that even “2 months is hard to imagine.” Language, multimodality, Coding, and Agent may continue advancing toward L4, but nobody knows whether that is 5 years away, 10 years away, or permanently unreachable. For now, the more reliable path is his neutral “muddle through”: improve existing businesses, follow the technology, and explore innovation.
刘飞 believes Tencent may also have extra options created by its investments. Ten or 20 seeds can grow independently; once one direction matures, Tencent can add capital or “bring it back in.” Google’s revival likewise cannot be attributed to a single variable among DeepMind, organizational adjustments, founder pressure, or the acquisition of the key person behind Character AI. Complex systems produce results when a set of decisions happens to be right at a particular moment; how long that lead can last remains an open question.

Sovereign AI in Poland: Language Adaptation, Local Control & Cost Advantages with Marek Kozlowski OpenAI's Code Red, Sacks vs New York Times, New Poverty Line?