Ali remakes Ali
Author | Chai Xuchen
Editor | Zhou Zhiyu
Ali's campus is surrounded by AI-ordered milk tea.
On January 15, Wu Jia, President of the Qianwen C-end Business Group, stood on stage and addressed the Qianwen App, saying, "Order me 40 cups of Bawang Tea Ji's Boya Juexian." Qianwen directly placed the order and completed the payment. Soon after, a Taobao Flash Delivery rider delivered the milk tea to the scene. There was no jumping between apps, no switching back and forth—everything was accomplished in one go within the Qianwen App.
This is a top-down expression of intent. Alibaba is trying to use AI as a lever to reshape the previously "fragmented Alibaba" into a highly efficient and collaborative "New Alibaba Kingdom."
At the venue, Wu Jia announced that the Qianwen App was fully integrated with Taobao, Alipay, Fliggy, AutoNavi, Taobao Flash Delivery, and other core Alibaba ecosystem businesses. The AI that used to only chat or draw with you suddenly gained capabilities like ordering takeout and booking flights.
"Qianwen is the first real AI that can get things done for you." Wu Jia told Wallstreetcn that the age of AI handling tasks has just begun; he wants to make Qianwen App the most powerful human AI assistant—eventually covering everyone.
And this is not just a functional iteration of an app, but more like Alibaba’s “self-reconstruction” in the AI era.
Alibaba's ecosystem is vast, from shopping and food delivery to maps and travel; but for users, forming synergy is hard, and in the mobile internet era, entry points only became more fragmented.
If Qianwen truly becomes "the most powerful human AI assistant," Alibaba will leap from "shelf e-commerce" to "command e-commerce." Not only would this fill Alibaba's gap in high-frequency C-end traffic entry, but it would also create a closed loop centered on itself: users can ask for directions, hail rides, eat out, and shop all within Qianwen, with all data and transactions kept inside the Alibaba system, no longer relying on external traffic sources.
But it’s not easy. The logic of "shelf e-commerce" is to let users browse, offer vast choices, and tolerate errors; while the logic of "command e-commerce" is to help users choose, pursue precision, and have low fault tolerance.
Letting AI “handle tasks” is essentially letting AI take on decision-making responsibility. This demands higher reasoning ability from Alibaba’s large model and real-time response from its service supply chain.
The Qianwen C-end Business Group was born into fierce competition. In a post-event interview, Wu Jia candidly told Wallstreetcn the endgame may consist of just two or three players.
Nevertheless, in his view, the old internet approach of burning cash and flooding the market with traffic to win no longer works. Only companies that truly raise the intellectual level of their models and have resources to invest in their ecosystems can reach the final table. There are very few such major players.
Thus, under Wu Jia's lead, Alibaba’s “second entrepreneurship” is unfolding with vigor. Can Qianwen become the "Jarvis" in every Chinese person’s phone? The answer remains unknown.
What is certain: When AI starts to pay for you, a new business era begins. In this era, whoever lets AI "get things done beautifully" first will be the new king.
Below is an excerpted transcript of the conversation with Wu Jia, President of Qianwen C-end Business Group:
Q: What are the core directions for general intelligence and Qianwen's iterations in the next six months?
Wu Jia: In the next half-year, integrating into the Alibaba ecosystem, expanding the boundaries of task-handling capabilities, strengthening the model’s understanding ability in daily-life scenarios—these are our very important main lines.
In life scenarios, achieving satisfaction for everyone is quite difficult. Consistency is better in work and study scenarios. We still want to leverage Alibaba's rich ecosystem alongside our model’s capabilities and user understanding to make Qianwen a global-leading product in life scenarios.
Q: How does Qianwen balance the conflict between efficiency and depth of thought?
Wu Jia: Internally we have a term: “appropriate.” I think AI shouldn’t equal absolute simplicity. For example, when I need to write a research report, I don’t just want an AI draft—I want co-creation. Communication between AI and people is essential, and that’s also true for life scenarios: sometimes you don’t know your real needs and need AI to talk with you, or even proactively suggest ideas.
I believe the key for AI lies in intelligence, not just efficiency. At this stage of intelligence, it often shows as higher efficiency. But more advanced intelligence won’t just be about efficiency; AI should think like a person.
We’re mainly using the model to raise its “IQ” and increase satisfaction across different types of needs.
Q: Qianwen is now the unified intelligent entry point for Alibaba’s ecosystem. What is the collaboration architecture, resource allocation, and coordination challenge between Qianwen and other Alibaba Group BU’s?
Wu Jia: In the AI era, we won’t separately create a lot of independent efforts for integrating specific services. It’s actually one model connecting many services. BU’s only need to register their tool capabilities with the AI. In this process, we need tuning, so the current approach is—we set up virtual teams with all group business units.
This is win-win. The bigger Qianwen gets, the more incremental life services it creates. Future AI-generated services will be incremental, not just existing ones.
Q: Will Qianwen only open the Alibaba ecosystem?
Wu Jia: No, we will open up for official cooperation after Spring Festival. We’ll select a time, because as you can see, internationally and domestically there are many different approaches—how to build a good methodology and blend with the Chinese ecosystem, given the difference from international internet ecosystems. So, we need to figure this out. Also, model usability and convenience; by bridging Alibaba’s ecosystem we can sum up better methods. But directionally, we will definitely open up.
Q: On the capability side, roughly how long is the cycle for Qianwen to go from novelty to daily necessity?
Wu Jia: Right now we see retention rates are pretty good. I think today’s AI functions we released are all user essentials. We didn’t heavily invest in entertainment and creative features (though we do some). What we invest in is essentials: work, study, food delivery are all daily-use, so retention looks ok. If people leave, we need to do better.
Q: What’s your logic in choosing what to do or not to do?
Wu Jia: We focus on high-frequency essential needs right now. Second, we focus on abilities AI can currently deliver, especially to C-end. The scope of AI's capabilities is fine, but not everything can be handled.
As a leader in China’s market, we’re relatively open. We haven’t ruled out any functions. The issues people wrestle with often center on doing AI products vs. traditional products—there is a real difference. For traditional products, we’d break things into dozens of tasks, each dealt with separately. Now, it’s the model doing 80% of the product with 90% satisfaction, so we spend 70% of our effort on that. Rather than optimizing one piece, we uplift coding, execution, planning abilities—all of it.
So the key is: we abstract user needs out of high-frequency, essential, model capability, Alibaba ecosystem, translate them into model/agent iteration directions, and execute. Long-tail needs that are really tough (even for humans)—we just wait for the next phase.
Q: How can we further leverage ecological advantages for deeper cultivation?
Wu Jia: Now we have three lines: one main long-term line is model and agent—making up for weaknesses. Today, looking at the entire market (not just AI products), all products’ satisfaction rates for such user needs are ranked. These two lines iterate by level—basically each quarter there’s a major model iteration (in conjunction with Tongyi), which is important.
Based on this, the agent's ability upgrades require further work—another line. Then, looking at the third line. So all three advance rhythmically.
Generally, today’s core is still driven by tech+data+ecosystem, not just a bug-fix iteration mode, since capability growth is still accelerating. So, for the next version—e.g., life assistant—we hope it’s personalized.
But translating this, it causes model problems. Next version probably won’t perfect takeout ordering, but we’ll keep abstracting out certain abilities. Takeout features, like showing pickup codes on orders, are experience-focused.
Q: How does Alibaba evaluate Qianwen’s impact on existing retail or e-commerce businesses?
Wu Jia: We haven’t seen people stop using Taobao just because they opened Qianwen. We’ll create incremental growth—easier, lower barrier, habits create increments.
But we can't rule out that some people will get used to ordering takeout in Qianwen instead of traditional platforms—it becomes a habit.
People aren't overly concerned, but we do look at rising visit frequency and session length.
Q: Is Qianwen’s model iteration target completely different from the base model?
Wu Jia: Our iteration target is a subset of the base model’s target since we build our business on top of the base model. Of course we do some post-training for certain applications, but still on their model.
Our major versions, every three months, basically feed new demands to the base model, which updates a new version accordingly. So, we’re a child account of their iteration target, because the base model serves the whole Alibaba ecosystem.
Q: For a generic AI assistant, on the engineering side vs. base model side, which side brings higher efficiency?
Wu Jia: Online, there's lots of talk: does bigger models/data help? Does C-end not need so much intelligence? In my area, Chinese business has passed that phase—it's all one model, not many models.
It's not that lower-IQ area needs a smaller model, higher-IQ needs a bigger one—it’s all large models. The model should be smart enough to know when to use simple reasoning for simple questions, smart models for smart questions. As the model count drops, our interface with the base model gets clearer.
The second is data capability: is more data important? Absolutely. Extra data is key, especially for life scenarios. Training data is from a set period, but China's supply is huge, and changes fast; you must solve timeliness, which is central to capability.
So, long-term benefit comes from the base model; if you look at efficiency over a year or two, the base model matters most. For short-term function iteration, post-training becomes more obvious.
Q: Within Alibaba’s AI-to-C entry points, how do Qianwen and Quark differentiate?
Wu Jia: They’re different. Quark is an AI browser and search; some people need an AI browser. Qianwen is an AI assistant—more like a person; some need an AI assistant.
The commonality: all AI features are in Qianwen, Quark’s AI features are in Qianwen too. This is a path issue; as time goes on, AI browsers/search won't disappear, but their share will shrink compared to dialogue in the AI era, though user habits remain—summoning Qianwen from browser/search.
They are different user-facing interfaces, but all AI is in Qianwen. On PC, it's 50/50: half prefer Qianwen, half prefer Qianwen in browser—still Qianwen. We don’t fuss over this. On mobile, though, more will use Qianwen.
Q: There's a lot of battle over AI entry points now. What’s our C-end strategic layout?
Wu Jia: The reality is, many endpoints have the same issue. I don’t think there’ll be as many players in the end: only those truly raising model intelligence and investing in ecosystem will survive—just a few, maybe 1-3. But does this mean only 1-2 front-end interfaces will remain? Hard to say.
In the early AI stage, everyone thinks there are opportunities. At this stage, making different-style AI assistants is still easy; each feels unique.
But I think from early 2026, that won't be the case. We’re running a lot of online tests, simulating different agent styles—main style now, but it’ll change. Like real life: some people become friends, some don't. You like my style, not theirs, even if we say the same thing.
So, when we reach true personalization and anthropomorphism, many companies won’t do AI anymore. There’ll be just a handful left in the market.
Q: This API integration across the group—do you foresee a big leap in model or agent capabilities in the next 3-6 months?
Wu Jia: We always considered this the core of AI product development—not a sudden revelation. Second, we saw the agent trend in life scenarios from July last year—after model launches, plus VRL advances. We saw the trend then and have invested even more since.
Q: Now big companies’ chat tools compete fiercely—some have DAU over 100 million. Does Qianwen care about DAU or MAU data?
Wu Jia: Even in traditional internet, 80 million vs 100 million isn’t a big difference. I truly believe: in the intelligence age, the product's key is whether it's passed the "intelligence threshold"—can it really serve and execute like a person, with high accuracy and satisfaction in a digital world. Achieving this depends on model and training paradigms etc., not just traffic as in old internet. For Alibaba, which invests heavily in models, we care about crossing the intelligence threshold, and global AGI progress.
Q: Prioritize user experience?
Wu Jia: No one in the world ever used this before—like when the first iPhone came out: it wasn’t great, but by iPhone 4 it was different. Maybe AI is still in the iPhone 1 era today.
AI making good decisions—we’re working on it with various methods. It’s not about exact product recommendations yet, but AI replacing traditional “scrolling” is promising—give us some time.
Q: AI is disrupting traditional traffic models, mobile OS, and APP commercial ecosystems. As more AI appears and is used, how will APP ecosystems change? Will the boundary between Qianwen and Taobao blur?
Wu Jia: I think it’ll take time to get there. First, I don’t think many agents will become major entry points as tech evolves. Today, comprehensive agent competitiveness is strong. Agents are now a stage product, different from app-integrated agents, like mini-programs to WeChat, merchants to Taobao. I think agents as standalone entry points won’t evolve much; they’ll be comprehensive. All-in-One is the trend, demand-wise.
Q: Forecast some likely changes for 2026?
Wu Jia: On the AI front, I think it’ll be more and more human-like—the ways it thinks and gets things done will feel more and more like using human services for your needs.
Risk Warning and DisclaimerThe market has risks; investment needs caution. This article does not constitute personal investment advice, nor does it take into account individual users' specific investment goals, financial circumstances, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their particular situation. Investments based on this are at one's own risk.
