cxuan-ai-labs

Industry, Business & Companies ·

Yao Shunyu's Interview Explained Tencent's Second Half of AI

English translation of the Chinese original. This version is generated for international readers and may be refined over time.

Interactive reading view ↗

English | Chinese Original

English translation of the Chinese original. This version is generated for international readers and may be refined over time.

Date: 2026-06-07

It's been a while since I've been hearing from AI.

On the 5th of June, a public conversation was held between Tung Dosheng and Yao Sun Yu at the conference on the application of the AI industry.

On 6 June, Thomason gave another round of media interviews, continuing to spread the questions he had been asking.

After I read it, it was like, this time, I finally made myself clear about AI.

As for the phrase "at last" it was not so important. That's what they write at every big meeting. It's so boring to read.

The real thing to say this time is that the model is just one piece. The products and tools, such as Yuanbao, WorldBuddy, CodeBuddy, documents, conferences, are subject to the same task process. AI has to get the context, get the tools to move, and finally get it done.

Todo-sheng and Yao-soon-sun.

  • Source: Information on the Official Congress. ♪ I'm sorry ♪

This may be the right path to take.

I think that's the best thing to do.

Let's start with the easy mix.

This is Yao Sun-rain. Don't mess with yao Sun-woo.

Yao Sun Yao is Qinghua Yao Ban, Dr. Princeton CS, who is a researcher, Tree of Thoughts, language intelligence, and who was previously in OpenAi, and later joined as chief AI scientist.

Yao Sun-woo is the other, Qinghua Physics, Stanford Theoretical Physics, formerly in Anthropic and Google DeepMind.

The two names are only one word short, but the route is completely different.

Two Shunyu Yao photo comparison

  • Two Shunyu Yao photo contrasts. ♪ I'm sorry ♪

Let's get this clear, or we'll have a whole line back there.

The first question to be asked is:

Isn't that slow on AI?

When ChatGPT came out, OpenAI, Anthropic, Google, DeepSeek, KIMI, GLM, byte, Ali had models or products to remember. By contrast, it is true that the period of arraignment was not the most spectacular.

Yao Sun-un didn't hide hard this time.

He didn't say it wasn't slow.

He just dropped a question:

Is AI a short-term game or a long-term game?

If AI is a run-off in two years, the product is slow, modeling slow, markets slow, of course.

But if AI is more like the '70s when the PC first appeared, there are still a lot of products that are not stereotyped.

The easiest thing to remember now is ChatGPT and Claude Code. Chat and writing codes are just two product forms. Multimorphism, intelligence, office collaboration, business processes, etc., will continue to change in the coming years.

Of course there are elements in that that explain the rhythm of the past.

But it also speaks of the conditions under which it is to be questioned.

At the end of the day, such companies are not very good at a global model release every three months.

It's better at putting AI in the real product, allowing a lot of users to use it on a daily basis, and changing it on the basis of feedback.

If AI looks more at the product in the second half of the game, there's still a chance.

If you look at the model at the end of the day, it'll be hard. Put it straight, it's not the right thing to do.

So I think it's not just about looking for steps for myself to talk about a long-term game. It's making trade-offs clear.

It's more appropriate to keep changing in products. The next model release will attract attention, which will not be appropriate for it.

The most critical sentence of Yao Sun Yu this time is:

In the past, AI was more important to find a way, and now it's harder to find a problem.

Used to be AlphaGo, it's a design for chess.

Translation is the design of a model for translation.

Methods and problems are largely tied together.

However, the situation changed with pre-training, post-training, large models. The model has become stronger and the product team has to answer a question:

Which specific issue should be addressed first?

What can be done is first and foremost specific needs and context.

It has a lot of real products, and it has a lot of real problems.

Social, office, conference, documentation, games, finance, health, education, business collaboration are all problems.

And behind those questions is the context.

Users used the product in the past, what information was available within the enterprise, what was said at the meeting, what was changed in the document, and where the development process was most easily jammed.

These things don't grow in benchmark.

That's why Yao Sun has repeatedly raised contact.

To put it simply, it's the "precedent summaries" that the models get before they work.

As models become stronger, product gaps fall on specific materials. User information, minutes of meetings, document modification history, code repository, business privileges, whether or not these are used by AI, will directly affect the usefulness of the product.

To be honest, the second half of AI, who knows what the user wants to solve.

Co-Design is more important than a list

This interview also included a keyword: Co-Design.

It's about models and product design.

And you can understand that the model team should not just look at the chart, and the product team should not wait until the model is finished before it is put into the product process.

They have to set goals, read data and change experiences.

How's the model and product?

  • Source: Information on the Official Congress. ♪ I'm sorry ♪

It was easy to stare at benchmark.

Math scores, code scores, long text token, number one.

These certainly work. But real users do not ask questions like benchmark.

Real users often do this:

First, a word is asked, then there is talk about asking, changing, supplementing and mixing search, documentation, meetings, codes, forms.

At this point, if the model is able to catch the context, if the tools can be adjusted, it is closer to real experience than a single question and answer score.

So, in an interview on June 6th, Yudosheng said that after Yao Haun came, he pushed the hybrids away from being benchmark in the accident department, to use product user experience as the main measure.

This is more important than "how much has been added to a list".

It indicates that the modelling team has begun to move from a cross-fertilization exercise to a responsible product experience.

Thomason added that before the training, Hy3, Yao Hao had done a great deal of data quality work, cutting down many of the data that seemed capable of stacking data, which did not actually help or even harmful to training.

That's a pretty good move.

A lot of people talk about modeling, and as much data as possible.

But the truth may be that some data are not assets and are sources of pollution.

Cutting data is harder than stacking data.

This decision is hard and decisive.

I think it's a lot more telling than "a few points up a list of hybrids."

Because the list is up a few points, ordinary users may not feel it at all.

But if the answers in Yuanbao, WorldBuddy are more stable, the task is faster and the user is willing to use it more often, the change will remain in the product.

So far, Hy3 has been connected to these products.

According to official sources, the speed of the first response increased by 54 per cent and the average time taken to complete the mission decreased by 47 per cent after the WorldBuddy access to Hy3 Preview. Thomason also said that about 80% of the dollar users are already using Hy3, and retention rates have increased significantly.

Hy3 Preview related line screenshot

*Text source: Community of telecommunication cloud developers. ♪ I'm sorry ♪

Hy3 Preview Use Trends

*Text source: Community of telecommunication cloud developers. ♪ I'm sorry ♪

These figures do not mean that the call has won. The conclusion is too early.

At least it can be seen, however, that product feedback has begun to enter model training, assessment and product overlay.

The call also released the "Efficacy Smart Tool Set" covering more than 20 vertical scenarios.

Query Efficiency Smart Tool Set

  • Source: Information on the Official Congress. ♪ I'm sorry ♪

On the personal side are QClaw, WorkBuddy, Yuanbao, Ima, and Qatar.

On the side of the enterprise are the WorkBuddy Enterprise, ClawPro, the Information Cloud Smart Development Platform ADP 4.0, the Enterprise Marketing Cloud.

Look at the name.

But don't understand it as a sudden announcement of a bunch of Agent.

More precisely, it's the ability to intelulate the past in various products, and it's the tools that Agent can call.

Query document to become WorkBuddy-able Skill.

Following the conference, Agent understood, called, generated minutes and refined to-do.

Enterprise Wisdom continues to host connections between people, people and services.

WorkBuddy is more like a workstation for people and AI.

Director of the Quest Efficiency Smart Tool Figure

Simple source: New Wave reproduces relevant contributions to the tumble cloud

Tencent Cloud Smart Body Development Platform

  • Map source: ADP official screenshot. ♪ I'm sorry ♪

Agent's hard here.

Just one chat box, and then the model can't afford it.

You have to give it tools, memories, privileges, context and an enforceable environment.

Without these, Agent is a very speaking guest suit.

With this, it'll have a chance to work. Otherwise, only a seemingly correct answer would be given and the task would have to be left to others.

I think a lot of Agent's products are right here.

It looks good when you talk. Once you get to the task, the gap comes.

If there is an advantage in making a call, it should also be a "hand-on link". It doesn't make much sense to be a more chatable portal.

Thomason mentioned a detail in the interview, which was interesting.

He said that product development was typical of waterfalls:

Product manager writes PRD, interactive designer design process, visual designer interface, front-end backend realization, final test.

Flow of responsibility for traditional product development

  • Traditional R & D function flow map.*

After AI started generating a lot of codes, the line was no longer so good to go down.

Many engineers will make less of a keyboard and spend more time defining results, architecture design, evaluation and testing.

An engineer can even do a demand, code, running test with a few Coding Agents, like the team leader. This is no longer a simple " assistive code."

It's a lot like my recent experience with Codex, Claude Code.

When AI really wrote a lot of codes, the most important ability of a person went from "Can I write this paragraph" to "Do I know if it should be written?"

Previously, capacity realization was the threshold.

It is more important to determine whether needs should be done, how modules should be dismantled, how interfaces should be determined, and how they should be finally accepted.

It was also mentioned that CodeBuddy had covered over 95 per cent of the engineers, and the overall coding time had been reduced by 40 per cent.

WorkBuddy Network

  • Source: WorkBuddy official webshot.*

CodeBuddy Network

Text source: CodeBuddy official webshot

This figure may not be directly extrapolated to all companies.

But the direction is clear.

The programmer will not disappear immediately. And the bigger change is that a lot of people are going to change from the person who writes the code to the person who drives the AI who writes the code and is responsible for the results.

This is both an opportunity and a pressure for ordinary engineers.

In this interview, the commercialization of Tong Doson was more cautious.

Smarts like WorkBuddy are still in the strategic input stage, and the call does not target the commercialization of the team.

Because the biggest problem with the AI product has changed.

Users want to use only the first step. The more you use it, the higher the cost, the real stress.

In the age of mobile Internet, marginal costs are low, and advertising, trading and carrying have the opportunity to cover costs.

Unlike AI, the cost of reasoning is real money and silver.

Each complex mission consumes token, computing, tools and storage. The more complex the task, the greater the cost.

So Thomason says that it's hard for AI's native services to cover the cost of reasoning by advertising alone, and it's better to use it first in a high-value, well-calculated scenario.

And that's the problem that all Agent products have to face now.

Demo did a great job, just the first step.

Once in real business, the enterprise calculates the budget, the method of costing, the liability boundary and cost sharing. Each of them affects the long run of the product.

There is also a more specific pressure on the placard: arithmetic.

Yudo-sheng has repeatedly referred to the strain of computing power, and limited resources have given priority to internal products such as hybrid training, Twitter, teleconferences, Yuanbao, etc. Renting GPU to outside customers, priority back.

And that explains why it's not urgent to wrap AI into a high-profit story.

It's more like an infrastructure and product test.

Start by running the use, product retention, model feedback and tool chains.

When these data are stabilized, the methods of charging and profit space are discussed.

A lot of AI products now have needs.

The trouble is, demand comes up, costs come up.

The more active the users, the more painful the company is. It's not so obvious on the mobile Internet, but the AI era is a real problem.

This interview with Yao Sun-rain and Tom Doo-sheng doesn't prove that AI has won, that he's ahead.

But the news finally made it clear what it wanted to do.

A simple follow-up to the model list is not a viable path.

Just one chat robot, not enough.

It's not enough to put hope on some bad money, either.

The presentation was on the integration of models, products, context, tools, data feedback and internal collaborative approaches.

This road is more in line with the way in which the past was used as a product.

The advantages of telecommunication are the variety of scenes, products, users and internal tool chains.

The trouble with communicators happens to be too many, too complex and too difficult to coordinate.

So I don't think it'll be easy to talk about it.

Its strengths and difficulties are the same: too many products, too many scenes and too many organizations.

It doesn't have to be just about the next model release.

At the end of the day, it depends on whether or not the products of Yuanbao, WorldBuddy, CodeBuddy, QClaw can be used in the context, pull up tools, run the mission and allow users to continue using it.

AI can't just watch the model release in the second half.

It depends on who can put the model in the real business and let it do the job steadily.

That's why I thought this interview was valuable. At least on this occasion, the products, context, tools and costs were told, and not only slogans were given.

Source:

#Agent #WorldBuddy