By Jiang Feng, Deep Flow Research Institute 

NextFin News -- On June 5, at the 2026 Tencent Cloud AI Industry Application Conference, Tang Daosheng—Senior Executive Vice President of Tencent and CEO of the Cloud and Smart Industries Group—sat down with Tencent’s Chief AI Scientist Yao Shunyu for a conversation about Tencent AI.

The conversation took place at a particularly telling moment:

On the one hand, people outside the company were still pressing questions like “Is Tencent moving too slowly on AI?” and “Is Tencent investing enough in AI?”; on the other hand, Tencent AI was clearly accelerating—products in the “Lobster” series were rolling out, Hunyuan Hy3 preview topped the global usage leaderboard, WorkBuddy was growing rapidly, and tighter linkages were beginning to form across models, products, and real industrial deployment.

It is precisely in this contrast that a subtle perception gap has emerged between how the outside world judges Tencent AI and how Tencent AI is actually progressing.

This conversation, which lasted less than an hour, offered a timely window into Tencent’s thinking. Rarely does Tencent lay out, within a single dialogue, its strategic judgments on AI, its organizational approach, and its path to real-world implementation:

from how it interprets the “second half” of AI, to the relationship between models, products, and scenarios, and further to how real problems, context, agents, engineering systems, and organizational capabilities are folded into one unified framework.

If over the past year the industry was more focused on “what Tencent AI has done,” then what makes this conversation more worth watching is how Tencent understands AI, organizes around AI, and pushes AI into real scenarios.

From this dialogue, we can distill 10 key reflections on Tencent AI in the second half:

1. From “Finding Methods” to “Finding Problems”

“The second half of AI” is a concept that comes from a blog post Yao Shunyu published last year, which sparked broad discussion.

Back then, Yao tried to use the term to capture how AI’s development stage was changing. In today’s conversation, he offered further explanation and expanded on the idea.

In Yao’s view, for a long time AI’s most important task was finding research methods: build a translation model for translation, a Go model for Go, and design bespoke systems for specific tasks. In that phase, the methods themselves were the scarce resource.

But pre-training and post-training changed that. Large models started to look like a "universal hammer", with a relatively general ability to solve problems. AI’s breakthrough point was no longer just “is there a method,” but became “which problems are worth solving”.

This means the AI industry’s main battleground is undergoing a shift in focus: as general methods gradually mature, what’s truly scarce is no longer merely model-training tricks, but the ability to identify, define, and continuously solve real problems.

Whoever has high-frequency, complex, real problems is more likely to train useful AI; whoever better understands the user needs, business processes, and cost structures behind those problems is more likely to turn AI into a product—not just a demo.

This is also the point Yao Shunyu repeatedly emphasized when explaining why he chose to join Tencent——Tencent has a huge number of products, and a huge number of real problems; more importantly, those products form an environment in which models can take action.

2. Context Becomes the New Moat

If “finding problems” answers where AI should create value, then “Context” answers how a model can truly understand users within a specific scenario.

Yao Shunyu mentioned Context many times in the conversation. His view is that models are getting better and better at turning complex inputs into outputs—but only if they can get sufficiently good inputs.

For individual users, those inputs might be preferences, habits, and historical behavior; for enterprise users, they might include customer information, business workflows, organizational knowledge, permission systems, historical projects, and system data.

Without that context, no matter how strong the model is, it can only offer generic answers; with that context, the model can actually understand “who you are,” “what you’re doing,” and “which answer is valuable to you”.

This means future AI competition won’t play out only in model parameters and inference speed; it will also play out in the ability to organize Context. For Tencent, if entry points such as Yuanbao, WeCom, Tencent Meeting, Tencent Docs, coding tools, cloud services, and others can become sources of context for large models to understand users and businesses, they will form a competitive advantage distinct from the model capabilities themselves.

But context isn’t simply about shoveling data into a model. What information should be provided, what shouldn’t, how access should be controlled, how data security and user privacy should be ensured, and how to avoid noise and interference—all of these are engineering problems, and they’re also product problems. Context may look like a data asset, but in essence it is a combined reflection of product capability, engineering capability, and cross-team organizational coordination.

3. Symbiotic Relationship Between the Model and the Product

When a model needs to enter real-world scenarios, simply plugging the model into a product is no longer enough.

The outcomes of AI products are often open-ended, dynamic, and impossible to enumerate exhaustively, which means model teams and product teams must jointly define the problem earlier and at a deeper level. That is exactly why Tang Daosheng and Yao Shunyu repeatedly talked about Co-Design.

In the traditional software era, the relationship between product and engineering was usually linear: product defines requirements, R&D implements features, and testing verifies results. But AI products don’t work like that. Model capabilities shape product boundaries, product data in turn influences model training, and user feedback further reshapes the evaluation framework. Developing AI products is more like a closed-loop system than a waterfall process.

Yao Shunyu mentioned that pre-training is more foundational and general-purpose, with the goal of making the core capabilities solid; but once you move into the post-training stage, the questions become much closer to the product: what the model should reward, what it should penalize, what counts as a good answer, and what counts as bad behavior—all of this requires feedback and evaluation from real applications.

Tang Daosheng added from the perspective of product experience that, in AI products, “good experience” isn’t a naturally clear-cut standard. How data is labeled, how granular the labels should be, and which behaviors should be rewarded or penalized—if these can’t align with product goals, the product’s behavior may ultimately drift away from what was intended.

This is the core of Co-Design: it’s not merely that the model team “supports” the product team, nor that the product team “invokes” model capabilities; rather, both sides jointly define what a “good” outcome looks like.

That is also why Yao Shunyu placed special emphasis on Trust. The hardest part of Co-Design isn’t the technical interface; it’s whether the model team and the product team can build trust, put themselves in each other’s shoes, and acknowledge that their goals are both aligned in some ways and different in others.

4. From Leaderboard Chasing to Real-world Evaluation

The large-model industry once relied heavily on benchmarks. Whether a model was even worth discussing often depended on its scores on several public leaderboards.

Yao Shunyu stayed restrained on this point in the conversation. He didn’t deny the value of benchmarks, but he also emphasized that there is a huge gap between leaderboard questions and real user problems.

Questions on leaderboards are usually clearly described, information-complete, and tightly scoped; problems in real scenarios are often vague, multi-turn, and laden with implicit context. A user might ask only, “Can you take a look and tell me whether this proposal works?”—yet the model has to understand the document, prior discussions, the company’s style, the target customers, and the decision criteria.

Capabilities like this are hard to fully capture through traditional leaderboards.

That’s why real-world evaluation is becoming increasingly important. It can not only reveal a model’s baseline failure modes, but also help R&D teams understand the real distribution of prompts—sometimes even inspiring new directions for capability development.

The release of Hy3 preview reflected this thinking as well. Yao Shunyu mentioned that one of the key reasons for releasing a Preview model first was to gather real-world feedback and fix issues that leaderboards didn’t surface. Compared with focusing only on public benchmarks, real interactions from products like Yuanbao can help the team understand what users actually need and clarify the optimization direction.

This also means that model development is no longer just about optimizing for external leaderboards. Instead, it needs to be grounded in real business scenarios to build its own evaluation system. Which capabilities matter for search, which matter for chat, which matter for office collaboration, and which matter for Coding Agents—all of these need to be broken down, evaluated, and fed back within specific products.

5. The AI Product Development Paradigm is Changing

If models and evaluation define the foundational capabilities of AI products, then changes in how users interact are reshaping the products themselves.

Tang Daosheng used a vivid analogy: traditional products are like “pre-made meals”—users can only pick from the menu; AI products are more open, with users expressing needs in natural language, and the product doesn’t know in advance what the user will ask.

Behind this is a shift in the product development paradigm.

In the PC internet and mobile internet eras, a product manager’s core job was to design features, flows, and interfaces. Users completed tasks through buttons, menus, and pages. The goal of product design was to make features as clear as possible and workflows as smooth as possible.

In the AI era, users are no longer just clicking features—they are expressing intent. The product has to understand that intent, break down the task, call tools, leverage context, and generate the result.

This calls for product managers to shift from "feature design" to "intelligent behavior design".

An AI product not only needs to answer what it can do, but also: when to ask follow-up questions, when to refuse, when to call tools, when to cite sources, when to disclose uncertainty, and what tone matches the user’s expectations.

This is also what makes AI product development harder.

The boundaries of traditional products are defined by features, while the boundaries of AI products are jointly determined by model capabilities, the tool system, context quality, permission controls, and the evaluation framework. It’s not about dropping a chat box into a product; it’s about rebuilding the relationship between the product and the user.

6. A Coding Agent is Not Just a Vertical Tool

Yao Shunyu is the creator of the ReAct framework, and his doctoral research has long centered on language agents. In the conversation, he looked back on how he began thinking about agents back in the GPT-2 era:

How to turn a machine that only predicts the next token into an agent that can interact with the external environment, invoke tools, and complete tasks.

Today, one of the most representative implementations of this thread is the Coding Agent.

In Yao Shunyu’s view, Coding Agents matter not only because the software development market is huge, but because their capability structure is fundamental. When a model can control the file system, call tools, run code, observe errors, and revise its approach, it effectively enters a relatively complete task environment.

That makes Coding Agents a testbed for training and validating general Agent capabilities.

They require the model to have long-horizon planning, tool use, error fixing, multi-turn reasoning, context management, and result verification. Once these capabilities mature, they won’t serve only programmers—they’ll also carry over to many more scenarios, including office work, scientific research, enterprise workflows, data analysis, and business operations.

This also explains why the discussion repeatedly mentioned two Tencent products——CodeBuddy and WorkBuddy. The former is aimed at developers, the latter at office users, but the underlying direction of capability evolution is the same: moving the model from "answering questions" to "getting tasks done".

An agent’s real value isn’t in chatting like a human, but in being able to close the loop in a real environment.

7. The Token Anxiety Trap

As agents begin taking on complex tasks, token consumption rises quickly. Both users and enterprises have started paying close attention to credits, call volume, and inference costs.

In Yao Shunyu’s view, the cost-effectiveness of tokens depends first and foremost on capability. A stronger model that gets the job right in one go can end up cheaper than a low-cost model that fails repeatedly.

This reframes token cost from a “unit-price question” back to a “task question.”

For enterprises, what they really should be calculating isn’t the cost per million tokens, but the total cost of reliably completing one end-to-end business loop. That includes model invocation costs, but also the cost of manual corrections, failed retries, waiting time, and business risk.

If a model is cheap but often makes mistakes, it may not be cheap in the end. If a model is expensive but can consistently complete critical tasks, it may actually be the more cost-effective choice.

Of course, that doesn’t mean cost doesn’t matter. Yao Shunyu also noted that Chinese teams have advantages in cost optimization—such as using smaller models to handle high-value tasks well, architectural innovation, long-text management, agent scaffolding, and more.

All of these optimizations rest on one prerequisite: the model must be reliable enough first. In the second half of the AI race, the battle over costs may not be a simple price war, but a competition in system-level efficiency centered on “consistently getting tasks done.”

8. AI-Native Products Need New Organizations: Small Teams, Heavy Experimentation, a Low-Ego Culture

When discussing the product team behind Tencent’s WorkBuddy, Tang Daosheng mentioned a detail: the organization is extremely flat, with many small teams of three to five people tackling specific domains, running lots of experiments, validating quickly, and being willing to tolerate trial and error.

This is markedly different from traditional internet product development. Conventional products typically rely on more mature processes: requirements review, design and development, testing and launch, and continuous iteration. AI-native products come with far more uncertainty—model capabilities and user behavior are both changing rapidly, and many directions can only be judged by trying them to see whether they work.

At the same time, the role of engineers is also changing. As more and more code can be generated by AI, an engineer’s core value is no longer simply writing code personally, but understanding requirements, designing architecture, breaking down tasks, orchestrating multiple coding agents, and taking part in evaluation and quality assurance. Tang Daosheng even said that every engineer is becoming more like a leader with ideas.

This means that in the AI era, product teams will see stronger role convergence: product managers need to understand the boundaries of models, engineers need product judgment, testing needs to move upstream into evaluation design, and algorithm teams also need to understand user experience. Organizations can no longer be just functional divisions of labor; they need to be reassembled around closed-loop task delivery.

But changes in AI organizations aren’t just about smaller teams and faster experimentation. Yao Shunyu mentioned that one key reason he chose to join Tencent was the culture here: it places greater emphasis on trust and sincerity, rather than operating solely around short-term metrics; the team also has a low-ego, pragmatic, and solid side.

These may not look like concrete technical capabilities, but they are foundational conditions for an AI organization. Because AI R&D and product rollout are full of uncertainty—model training, product co-design, real-world evaluation, and agent trial-and-error all require cross-team trust, and also require the honesty to face failure feedback head-on.

In Yao Shunyu’s view, an AI organization that stays oriented toward AGI over the long term should be a balanced “triangle”: first, foundational capability—making pre-training and post-training truly solid; second, product capability—turning technology into real value for users and society; third, frontier exploration—continuously searching for new research paradigms and opportunities.

9. Has Tencent AI “fallen behind”? A Multi-front Contest has only Just Begun

This conversation also addressed outside doubts about “whether Tencent’s AI has been too slow.”

Yao Shunyu broke the question down into two judgments: is AI a short-term game or a long-term game? Will the future follow a single main storyline, or be a landscape of diversified competition?

In Silicon Valley, a strong sentiment had taken hold for a period of time: AI would replace a large amount of work within one or two years, and everyone needed to rush to get positioned in the short term. But Yao Shunyu didn’t agree with that assessment. In his view, AI isn’t a sprint that’s already nearing the finish line; it’s more like a long-cycle transformation that has only just begun.

He noted that ChatGPT and Claude Code shouldn’t—and won’t—be the only super apps. A more likely scenario is that, much like when PCs first emerged in the 1970s, the real product forms, business opportunities, and ways of using them are still far from being fully invented.

In the past few years, the industry seemed to have a relatively clear technological roadmap: pre-training, post-training, Agents, and Coding Agents. Everyone was chasing along roughly the same direction. But Yao Shunyu believes that the future won’t be left with only a single main track; instead, it will see diversified development.

Coding Agents will, of course, become increasingly important, but they are still only the beginning. Multimodality, embodied intelligence, enterprise Agents, office collaboration, and industry-specific applications—there are still many scenarios that haven’t truly been filled in.

That’s why the question “Is Tencent AI falling behind?” can’t be judged solely by a particular moment in time, a single product, or one launch event.

The more critical questions are: Can Tencent confront feedback honestly? Can it adjust quickly within a complex organization? Can it turn user feedback into model improvements, turn product experience into evaluation systems, turn engineering capabilities into reusable platforms, and turn multi-business scenarios into a Context network?

In the conversation, Yao Shunyu mentioned that in the past, both the models and the products went through a lot of exploration—and took plenty of detours. That’s not surprising. What truly matters is “whether you can Be Real, whether you can change after seeing feedback, and whether you can stay patient.”

This may also be the core proposition for Tencent AI in the second half of the game: how, in a long-term, diverse, and still rapidly evolving AI cycle, to translate complex scenarios and long-termism into the speed of continuous iteration.

10. Tencent’s AI path: Systematic Synergy among Scenarios, Engineering, and Models

For a company like Tencent—with complex businesses and a large portfolio of products—the challenge of an AI strategy isn’t just building a powerful model and launching multiple AI applications. It’s how to make dispersed scenarios, engineering capabilities, and model R&D work in concert, forming a system capability that can iterate continuously.

Tang Daosheng summed up Tencent AI’s three core capabilities: scenario connectivity, engineering mastery, and model-driven development. These three terms, in effect, also correspond to three key links in AI’s journey from technical demos to industrial-scale deployment.

First is scenario connectivity. Through high-frequency touchpoints such as WeChat, WeCom, and Yuanbao, Tencent embeds large models into real business workflows. This aligns with Tencent’s long-accumulated user scenarios and enterprise connectivity capabilities. Only when AI enters real workflows will it generate real feedback—and only then will it have the chance to be continuously improved.

The second is engineering mastery. Through the Harness system, AI Infra, Agent Runtime, high-speed networking, high-throughput storage, and GPU utilization optimization, agents can run in a stable, trustworthy, and sustainable way. This corresponds to the engineering challenges that must be solved as AI moves from demos into production. For enterprise customers, whether an agent is usable depends not only on how smart it is, but also on whether it is stable, secure, controllable, and sustainable.

The third is model-driven development. Built on the Hunyuan foundation model, and through co-design of models and products, it seeks a balance among practicality, cost-effectiveness, and ROI. A model is not a standalone capability; it needs to keep evolving through product feedback and be reliably delivered through an engineering system.

AI competition is not just a contest of individual models, but a multifaceted competition shaped jointly by models, products, engineering, scenarios, and ecosystems.

Tencent is not a company with only a single AI product; it is a company with complex businesses, many scenarios, and multiple organizational forms. Complexity may slow a company’s actions, but on the other hand it also provides a wealth of real problems and context. Whether Tencent can turn complexity into a systematic advantage is the key to its “second half” of AI.