Author | Zhang Xiaojun

Production | Tencent News "Underwater Eye"

In the past year, a new important corporate WeChat group chat has been added among Tencent's senior executives.

Tencent Chairman and CEO Ma Huateng, President Liu Chiping, Senior Executive Vice President Lu Shan, and Tang Daosheng have entered the group. The group is dedicated to large-scale model technology transformation and is used for sharing and discussing the latest frontier and following up on business progress.

Many people said that Tencent "is not in a hurry" for large-scale models. In this regard, Tang Daosheng expressed a different view: "We are anxious. We often discuss it in the group and can't say it's not anxious."

Tang Daosheng (Dowson Tong) is the CEO of Tencent Cloud and Smart Industries Business Group, and a member of the Executive Office. He was born in 1973 and graduated from the University of Michigan and Stanford University. In his early years, he worked at Oracle and Sendmail. He returned to China and joined Tencent in 2005, from a systems architect to rising to become the CEO of the Social Network Business Group and the Cloud and Smart Industries Business Group, the former responsible for the QQ business and the latter responsible for the Cloud and to B business. In the past 19 years, Tang Daosheng has witnessed Tencent's key battles in social, music, to B and other key battlefields. Now, he is one of the core members of Tencent in the AI ​​war.

This is a cross-sectional view of how China's large technology companies think about the AI ​​technology revolution. Different from the perspectives of entrepreneurs and investors such as Yang Zhilin, Zhu Xiaohu, and Wang Xiaochuan we reported earlier, Tang Daosheng provides a perspective of artificial intelligence from the perspective of a large corporate operator.

During the nearly 4-hour interview, Tang Daosheng answered fresh topics such as "how to view price wars", "investment thinking on large-scale models", and "the relationship between platform-type enterprises and large-scale model startups"; at the same time, explore from a longer-term perspective. Discussing how Tencent views competition, business rhythm, and organizational culture in the two races of to C and to B. More importantly, how should Tencent, born in the last technological revolution and now 25 years old, respond to the new technological changes.

In Tang Daosheng's view, for a company with such a size as Tencent, resources must ensure investment and follow-up research and development, and do research and development similar to OpenAI; but at the same time, it must remain clear-minded, and "don’t equate AI with just large-scale models, we must look at it more comprehensively."

"AI is not just for large-scale models. This is equivalent to thinking that only companies that make mobile phones are important in the mobile era, which is too narrow." he said.

You will see that this business executive who has been on the Internet for a long time rarely uses expressions like "Get a feeling". He said: "The high-level feelings of Tencent for the market and products are closely related to the feedback from use." For the latest technology, they will obtain information from different angles, but the most important thing is the feedback from use." Our feelings come from experience, use the products on the market myself, Get a feeling of what stage this technology has developed? What is the problem? "

The following is the interview with Tang Daosheng. (For ease of reading, the author has done some text optimization.)

1. Talking about platform-based enterprises doing AI-"AI is not just a large model, and keeping options open is a reasonable approach"

"Underwater Eye": How has your state been in the past year? Do you lose sleep because large models are progressing too fast?

Tang Daosheng: Emm, hahaha, maybe a little bit. I am very interested in technology. Even on a busy day after a meeting, at home or on weekends, I will open some technical explanation videos and watch them until I fall asleep-from another angle. Describe this state.

"Underwater Eye": An industry evaluation that is often heard is that in the large model wave, Tencent "is not in a hurry." Are they right? Why is Tencent not in a hurry?

Tang Daosheng: Pony (Ma Huateng) said before, "Don't be in a hurry to come up with half-finished products" It is most important to find the direction, think clearly, and maintain iteration. Is it important to get up early or stay up late? I think that it is not only important to get up early, but also to stay up late - sometimes excessive investment in the early stage, cannot persist, and cannot laugh until the end.

Look at Apple's investment in mobile devices: from the earliest Newton (Apple's touch palm computer launched in 1993), which was a big palm computer that I used a stylus to operate when I was in college; to PDAs like Palm (also a handwritten palm computer); the next stage is iPod music player; then to the iPhone. Even Apple, it took more than 10 years of iteration of different product forms to seize the opportunity in mobile devices-so the most important thing is not who made the PDA the earliest, but who eventually occupied the market.

"Underwater Eye": Are Tencent FOMO (afraid of missing out) on large models?

Tang Daosheng: Of course, we are afraid of missing opportunities, and companies of this size must be prepared for every technological change or paradigm shift to ensure that resources are invested.

But large models are only a part of the AI industry. Many other technical paths in the field of artificial intelligence are also worth paying attention to. To build a useful intelligent system, large models may only be one module in it.

Not only players who do large models are doing AI. This is equivalent to thinking that only companies that make phones are important in the mobile era, which is narrow-minded. At least I try to let the team see things more comprehensively.

Tencent's mechanism is still quite vital, and each team is paying attention to the relationship between AI and their own business and has multiple layouts. Therefore, it is normal to be afraid of missing out, especially for us, keeping options open is a reasonable practice.

"Qian Wang": What is the overall layout of Tencent's CSIG (Cloud and Smart Industries Group) in AI?

Tang Daosheng: Our AI layout is roughly divided into four layers: infrastructure, model management tools and engine frameworks, large models and model stores, and application scenarios.

The bottom layer is infrastructure, including building high-performance computing, connecting clusters with high-speed networks, and matching high-performance, cost-effective storage to make training faster. This is the foundation of cloud business.

Above that is the tool layer. To train large models, you need a set of tools, including tag management and inference acceleration, as well as an engine framework for building a RAG (retrieval augmented generation) system, which is also an ability we will provide.

Then there is the model layer that everyone is concerned about. The company's focus is on the mixed-element large model, which has supported more than 600 internal applications. We also cooperate closely with the mixed-element team to explore more industrial scenarios that call for large models to meet customer needs for drawing, graphics, videos, and 3D. Different customer scenarios also have different requirements for models. In addition to mixed elements, Tencent Cloud also supports customers to choose other models, and some customers also use our model tools to fine-tune other open source models such as Baichuan and GLM.

Then there is the application layer. Tencent already has many to C and to B products, each with user service scenarios. They are all thinking about how to use AI to improve user experience and increase usage efficiency. For example, in CSIG's corporate applications, Tencent Meeting uses a large model to generate meeting summaries, Tencent LeXiang uses a large model to build an intelligent assistant familiar with corporate document libraries, and the Coding R&D platform uses a large model to generate code.

"Qian Wang": What is the most frequently considered question about AI for you right now?

Tang Daosheng: In addition to general large models, I attach great importance to how everyone can use AI in industrial scenarios and what tools and capabilities need to be provided. At present, I feel that it may still largely fall back to the mode of RAG (retrieval augmented generation) to reduce the probability of illusion.

At the same time, I also attach great importance to how to grasp the direction of the "intelligent entity," how to combine different technical modules, build more valuable applications, solve specific problems, and establish a reasonable business model.

In addition, many AI companies in the previous stage did not produce standard products and could not achieve rapid replication in similar scenarios. They could only rely on stacking manpower to continuously do projects and personalized delivery, which makes it difficult to establish a large-scale business model and leads to perennial losses and inability to make a profit. The new generation of AI companies should learn from this and use relatively standard products to establish a business model and reduce personalized delivery.

"Qian Wang": What do you think of the saying "model is product" or "model is application"?

Tang Daosheng: I don't care too much about some conceptual packaging. Products are products, and the model itself is not a complete product, but needs a lot of capabilities. Users have to use it comfortably, not just "models spit things out."

"Qian Wang": Recently, the industry has paid much attention to the news of the large-scale model's price cut by domestic companies. ByteDance started it, and Alibaba, Baidu, and Tencent followed. How did Tencent make the decision to follow?

Tang Daosheng: The actual industry price cuts are not as significant as imagined. (Many companies' actions are) free input tokens and do not affect output tokens much. Initially, I was more opposed to price wars, but when the whole industry is doing it, we also need to take action. Chinese companies are too fond of using price wars to compete, often using losses to exchange market share, rather than winning by ability or effect.

In the case of ample capital, some companies like to use the subsidy tactic to grab the market, but Tencent attaches more importance to sustainable development. I pay close attention to the P&L statement and each business should calculate the cost and set a reasonable price to avoid relying on other companies' profits to subsidize our own losses. When doing any business, I would think: What is the reasonable business model? What kind of input and output should there be for long-term and healthy development?

Cloud services are the same. Early on, everyone only cared about market size and share, and we used losses to gain market share and get projects. Later, we found that these losses did not bring customer recognition and loyalty and even caused many management actions to deform. Therefore, we insist on sustainable development, focus on products, solidify technology, manage prices, continuously optimize costs, and allow each product and sales unit to ultimately support itself.

Sometimes, the environment is not so good, but it is more favorable for our practices. You don't need to be more influenced by the fickleness of the capital markets and return to the essence of the business. The foundation of the business should look at how much your costs are, how much you earn, how large is your team based on the profits, so that you can achieve long-term, healthy, and sustainable development. If you carefully look at business and financial data and clearly see each input and output, you will be able to see the truth.

Generally speaking, we are steady and persistent, and we will stick to it once we have it figured out. We're not the kind of people who are very reckless.

2. Discussing the relationship between platform-type companies and large-scale venture companies – who will have the opportunity to create the next generation of C-end user portals?

Perspective: Globally, large-scale venture companies still rely on platform-type companies today. At this stage, what specific demands do they have of platform-type companies?

Tang Daosheng: For venture companies, the first thing is to be able to consistently receive investment. Building a large model is quite expensive.

The second is computing power. Computing power was very scarce early on. Whoever could supply it, they would be willing to partner with them. As the supply gradually becomes more abundant, enterprises will begin to be more choosy and consider which platforms offer more efficient resources. When the market isn't being constrained, they'll want everything. They may now shed scattered resources and concentrate on resources with the highest ROI (return on investment). People don't have much cost difference when it comes to bare resource costs, but different cloud enterprises and platform vendors have different accumulations and may have differences in training efficiency, service capabilities, and pricing.

Thirdly, technology. If you're running a large cluster for training, you're heavily dependent on the underlying cloud platform technology, such as what to do if a GPU server crashes during checkpoint (checkpoint) and how to quickly restore training without starting from zero, as well as network bandwidth and speed of the GPU card interconnect and the performance and cost of model storage. So our Xingmai network and GooseFS are quite popular.

I think several large-scale companies are using resources from multiple cloud vendors. Some vendors' resources are used for training, some are used for inference, and some are used for various data processing tasks before training.

Perspective: In terms of investment strategy and cloud computing strategy, some platform-type enterprises are more aggressive. For example, they will invest heavily in large-scale venture start-ups and use computing power to exchange for investment. What do you think of this approach?

Tang Daosheng: If the invested enterprise succeeds, cloud consumption needs to be sustained. This is of course the best source of revenue for the cloud business. However, the current GPU resources used to train large models are very scarce, and enough GPU cards must be used for investment.

Whether using cloud resources or funds to invest, you still have to spend money. With so many players involved, how many winners can really be picked in the end? This is a test of the investor's perspective, not of their infrastructure capabilities.

Last year, there was a GPU computing power famine, and large-scale venture companies had to buy GPUs with the money they raised. During the investment process, if computing power is packaged along with cash to ensure supply, the entrepreneur doesn't care. This has led to some companies sharing resources to make investments.

But in a situation where computing power is abundant and the market is running efficiently, truly viable enterprises and their original investors will definitely want cash. This ensures that their resources are allocated more flexibly, and choosing cloud vendors can also be based on cost-effectiveness and the quality of their services and technologies. This is the most efficient market behavior. Previously, bundling occurred because of supply-side problems.

Perspective: Is using large models to drive cloud revenue a sustainable long-term business model?

Tang Daosheng: In the early stages of a new technology, there is a boom, and a large amount of capital drives the frenzied growth of venture companies. It may involve excessive investment and many of these players may be part of the bubble. If cloud revenue is too reliant on capital-led venture companies to consume, once the bubble bursts, some clients will disappear and revenue will decrease. This will be painful when it drops.

You look at the 1990s, when internet venture companies were flourishing. When I was at Oracle, the company made a lot of money selling databases to internet companies. But later, 90% of internet enterprises disappeared, which put a lot of pressure on performance in the early 2000s.

Perspective: What is the current state of the large model industry today?

Tang Daosheng: It's very early. Everyone is staking their claims and trying out different business models. Some are pursuing scaling law, some are creating new B-to-C market entry points, and some are working on industrial landings. It's very lively.

So bundling cloud resources to make investments is a special phenomenon that occurs in a special situation?

Tang Daosheng: Yes. Also, if you invest in a large model enterprise, you might be investing $500 million in it. Suppose it consumes $300 million of computing power in one year, and suppose it's a 0-margin and 5-year-amortized cost. The cloud vendor's capital expenditure could be five times this. Although you invested 300 million computing power for one year's consumption, you actually spent 15 billion in actual cash to buy GPUs.

I don't know the details of how other investors invest in each enterprise, and I don't know if they use cloud vouchers or lock certain GPU card resources. There is great uncertainty in these 5 years. Perhaps GPU technology will be upgraded, and the originally invested GPU card will depreciate greatly. Perhaps the invested enterprise's development is not ideal, and they won't be able to consume all of the cloud vouchers, so there may be a need to find a way to exchange them back for cash.

There are many complex problems here, and there are many accounts to be calculated in the operation.

"Qian Wang": In your opinion, is the basic large-scale model an opportunity for giants or start-ups?

Tang Daosheng: In the basic large-scale model, in China, big companies have more resources, and have invested more or less time in AI; however, start-ups are more flexible and have their advantages in attracting young talents.

Once entering the field of model application, start-ups are best differentiated from large companies, and enter the market that big companies are not particularly concerned about, or business models that big companies are not familiar with, such as Didi's taxi service at the time.

Start-ups like Dark Horse, Baichuan and MiniMax have the opportunity to win the market. In addition to their excellent large-scale model technology, they also pay attention to investment in C-end applications and B-end vertical tracks.

For example, everyone is trying to use chatbot apps to meet the search needs of users - who has the opportunity to grab more market share in search? Or build a new generation of C-end user access?

3. Talking about Tencent's speed and business style - "There are many right things, which may seem stupid in the short term"

"Qian Wang": Tencent released the C-terminal product Yuanbao on May 30th, why at this time?

Tang Daosheng: The product needs to be ready. If you feel that the preparation is not sufficient, it is okay to wait. Not everything is about speed. We need to master the technology thoroughly, ensure a good user experience, and the market has not yet reached the stage of "life and death" or "fighting every second". As the market gradually takes shape, user needs will become more clear and we will accelerate our pace. In the past six months, there have been significant changes, and the frequency of new product and new version releases has become higher and higher - we have shifted from long-term technology deployment to attacking the market with products.

"Qian Wang": Some people ask, is Tencent slow on the big model?

Tang Daosheng: Actually, it's okay, at least I don't think so.

But even if it's slow, so what? You can say that Google, you are slow, and you didn't make a search engine in 1995. At that time, isn't AltaVista more powerful? (Note: AltaVista was founded in 1995 and was one of the most famous search engine companies in the early days of the Internet. Google was founded in 1998.) But who cares? They are the winners now and their technology is very advanced.

"Qian Wang": There is an evaluation of Tencent's business style in the outside world, "Follow, follow, and gradually lead." What do you think?

Tang Daosheng: The industry often underestimates Tencent's ability to make long-term products. If you want to win, can you just win by following? (laughs)

In fact, our style of doing things is to continuously refine products, continuously improve the experience, and as long as we win user reputation, we have the opportunity to lead. Some self-media always associate Tencent with "following". We are not only "following", if there is no innovation and highlights, it is impossible to become a market leader.

"Qian Wang": Is the description of "latecomers" accurate? What is the trick to being a "latecomer"?

Tang Daosheng: In history, Tencent has participated in many businesses that directly compete in the market, some of which have been successful and some have been unsuccessful. In successful businesses, some have participated in the market early, and some have invested later. Whether you can "beat" your opponent still depends on follow-up execution. Once you decide to invest in a certain track for a long time, it is more important to find differentiated playing methods, continuous improvement of product experience, rather than "first mover" or "latecomer".

For example, the popularity of smartphones. The iPhone is definitely the most important pusher in the smartphone market, but is Apple the first company to make smartphones? Actually not. At least Dopod did it earlier. The trick of the "latecomer" is, can you meet the pain points of users, provide a better experience, and continue to invest?

In addition, persistence is also very important, sometimes even patiently wait for the opponent's decision-making mistakes. Some companies like to be high-profile at the beginning of a new track, invest heavily, lose a lot, and the extensive management makes it difficult for you to find the key path. Because of long-term huge losses, the results cannot be sustained.

I have had some experience with QQ Music. In the early years, many platforms were competing to capture the online music market in the PC era. QQ Music also faced great challenges--a series of problems such as increasing content copyright fees, severe piracy, lack of a business model, and so on. But we insist on legal music, have gained the long-term trust of copyright owners, continuously improve our music library, continue to combat piracy, avoid low-priced copyright purchases, and attempt to control losses as much as possible. Irrational competition among platforms has led to huge long-term losses. In the mobile era, many people cannot endure it. Product structure, 10-30 billion yuan products operating income of 401/1288/60 million yuan respectively.

Fortunately, we persisted. From 2005 to the listing of TME (Tencent Music Entertainment Group) in 2018, we experienced competition from many people to many people giving up, and the process of finally establishing healthy commercial rules.

Please use your Futubull account to access the feature.

Q: Where is the turning point of QQ Music? And how should the time rhythm be? A: The turning point lies in the country's attention to intellectual property protection, which supports the platform to combat piracy. Of course we need technical means to find out where there is piracy through full network scanning and monitoring, timely discover piracy music, grasp the most important time window, and make piracy music withdraw. If piracy dissemination is not controlled at the critical time, it is difficult to attract users to legitimate music platforms. As for the time rhythm, it is hard to say. In fact, it is waiting--waiting for opponents who disrupt the rules and mess up to abandon the market, and then you can build a healthy and sustainable business model.

Fighting piracy in the long term is definitely right and it is the foundation of the entire creation ecosystem. But it is also hard work, not short-term effectiveness, and must be persistently maintained. Also, if the platform fights piracy, your competitors may benefit in the short term, but you cannot stop just because of this.

There are many things that seem dumb in the short term but are right in the long term.

Q: Overall, how to summarize Tencent's work style? Can you give an example to express this style? A: The work style of different teams may be different. The style of most product teams should be 'small steps, quickly iterate.' More emphasis on product experience, iteration, satisfying user needs on a small scale first, validating agreement when everyone agrees, spreading word of mouth, and then increasing investment.

Q: In general, how do you summarize Tencent's work style? Can you give an example to express this style? A: The work style of different teams may be different. The style of most product teams should be 'small steps, quickly iterate.' More emphasis on product experience, iteration, satisfying user needs on a small scale first, validating agreement when everyone agrees, spreading word of mouth, and then increasing investment.

Q: In general, how do you summarize Tencent's work style? Can you give an example to express this style? A: The work style of different teams may be different. The style of most product teams should be 'small steps, quickly iterate.' More emphasis on product experience, iteration, satisfying user needs on a small scale first, validating agreement when everyone agrees, spreading word of mouth, and then increasing investment.

Q: At what point did you see the explosive growth of the application? A: Our product was launched at the end of December 2019, before the outbreak of the COVID-19 pandemic. The 2020 Spring Festival was the most critical time window, when everyone had to return to school or work. Other companies later discovered this opportunity and began investing heavily to catch up. But we had already been doing it for a year and a half before the epidemic, so the product experience and stability were much better.

Q: Can you talk about how you discover, identify, and foster edge projects with potential as a leader? How do you know that you are not hindering it? A: This is what I want to do. How could I hinder it? (laughs). Many edge projects are actually quite difficult. Since they are on the edge, the company will not allocate key resources to support them, and the initial investment will surely be a loss, which requires other business revenues to subsidize. Then wait for the opportunity and hope to grow step by step. The annual bonus of everyone is strongly related to business income and profits. How to continuously encourage and retain teams, allowing them to see the long-term value and development opportunities of such internal entrepreneurial projects? This is a test of the project leader's ability.

Q: Can you talk about how you discover, identify, and foster edge projects with potential as a leader? How do you know that you are not hindering it? A: This is what I want to do. How could I hinder it? (laughs). Many edge projects are actually quite difficult. Since they are on the edge, the company will not allocate key resources to support them, and the initial investment will surely be a loss, which requires other business revenues to subsidize. Then wait for the opportunity and hope to grow step by step. The annual bonus of everyone is strongly related to business income and profits. How to continuously encourage and retain teams, allowing them to see the long-term value and development opportunities of such internal entrepreneurial projects? This is a test of the project leader's ability.

Q: Can you talk about how you discover, identify, and foster edge projects with potential as a leader? How do you know that you are not hindering it? A: This is what I want to do. How could I hinder it? (laughs). Many edge projects are actually quite difficult. Since they are on the edge, the company will not allocate key resources to support them, and the initial investment will surely be a loss, which requires other business revenues to subsidize. Then wait for the opportunity and hope to grow step by step. The annual bonus of everyone is strongly related to business income and profits. How to continuously encourage and retain teams, allowing them to see the long-term value and development opportunities of such internal entrepreneurial projects? This is a test of the project leader's ability.

Q: Can you talk about how you discover, identify, and foster edge projects with potential as a leader? How do you know that you are not hindering it? A: This is what I want to do. How could I hinder it? (laughs). Many edge projects are actually quite difficult. Since they are on the edge, the company will not allocate key resources to support them, and the initial investment will surely be a loss, which requires other business revenues to subsidize. Then wait for the opportunity and hope to grow step by step. The annual bonus of everyone is strongly related to business income and profits. How to continuously encourage and retain teams, allowing them to see the long-term value and development opportunities of such internal entrepreneurial projects? This is a test of the project leader's ability.

Q: Can you talk about how you discover, identify, and foster edge projects with potential as a leader? How do you know that you are not hindering it? A: This is what I want to do. How could I hinder it? (laughs). Many edge projects are actually quite difficult. Since they are on the edge, the company will not allocate key resources to support them, and the initial investment will surely be a loss, which requires other business revenues to subsidize. Then wait for the opportunity and hope to grow step by step. The annual bonus of everyone is strongly related to business income and profits. How to continuously encourage and retain teams, allowing them to see the long-term value and development opportunities of such internal entrepreneurial projects? This is a test of the project leader's ability.

Therefore, I let Lori (Wu Zulong, head of Tencent Meeting) lead the meeting. At the beginning, he said, 'Hey, Zoom has more than 2,000 people in China. How can I do it with 30 people?' But the small team picked some of the core functions, developed a basic product prototype, and then tested it internally, continuously iterating and grinding the experience.

Q: Can you talk about how you discover, identify, and foster edge projects with potential as a leader? How do you know that you are not hindering it? A: This is what I want to do. How could I hinder it? (laughs). Many edge projects are actually quite difficult. Since they are on the edge, the company will not allocate key resources to support them, and the initial investment will surely be a loss, which requires other business revenues to subsidize. Then wait for the opportunity and hope to grow step by step. The annual bonus of everyone is strongly related to business income and profits. How to continuously encourage and retain teams, allowing them to see the long-term value and development opportunities of such internal entrepreneurial projects? This is a test of the project leader's ability.

Q: At what point did you see the explosive growth of the application? A: Our product was launched at the end of December 2019, before the outbreak of the COVID-19 pandemic. The 2020 Spring Festival was the most critical time window, when everyone had to return to school or work. Other companies later discovered this opportunity and began investing heavily to catch up. But we had already been doing it for a year and a half before the epidemic, so the product experience and stability were much better.

Q: At what point did you see the explosive growth of the application? A: Our product was launched at the end of December 2019, before the outbreak of the COVID-19 pandemic. The 2020 Spring Festival was the most critical time window, when everyone had to return to school or work. Other companies later discovered this opportunity and began investing heavily to catch up. But we had already been doing it for a year and a half before the epidemic, so the product experience and stability were much better.

In the communication market, the network effect is particularly strong, relatively more time-sensitive, and you must seize the time window. Once you build up the network of users, it will be much harder for others to push another app into your network.

"The Lookout": How to motivate the team and quickly respond to fleeting time windows?

Because of the earlier product support, the users were able to use it immediately, thus seizing the key time window as soon as it arrived. At the beginning of the epidemic, various needs emerged, and members of the general office helped everyone overcome various difficulties, and also utilized the WeChat group video interface to promote Tencent Meeting.

I remember that before the eighth day, I had already discussed this during the morning meeting with Pony and Martin. They personally mobilized resources and immediately approved a lot of personnel. R&D personnel from other products in CSIG (Cloud and Smart Industry Group) were temporarily borrowed, TEG (Technical Engineering Group) also actively helped us set up equipment and expand capacity - this team went from dozens of people before going online to hundreds of people during the critical period. On the one hand, it undertakes social responsibility, and on the other hand, it accelerates the development of the product.

When we see clear opportunities, we will increase investment without hesitation, and collaborative operations can be very agile.

"The Lookout": In the face of certain opportunities, how much investment can be made?

It depends on the specific situation and actual needs of the project. Tencent Meeting was offered for free during the epidemic, and the cost of just bandwidth alone was over 1 billion yuan per year, with many resources invested. The investment in AI and large models is even greater, and it has brought together many technical experts from many different teams in the company across different sectors, as well as huge computing resources and external investments. So it depends on the specific context (note: the racecourse and battlefield)

"The Lookout": Under this work style, what kind of organization does Tencent need? Should innovation be more "top-down" or "bottom-up"?

This question cannot be answered by multiple choice. My previous perception was "bottom-up", but later I felt more and more that top or bottom is relative. Many times, innovation is carried out in both directions. In CSIG, I will choose the market and racecourse where to invest. After startup, business managers need to lead the team to continuously explore, and the team also provides a lot of feedback, constantly discussing and optimizing strategies.

For example, although I proposed to develop this product first, the continuous iteration of meeting functions, as well as the details of product strategy and commercialization, such as the idea of pure software services and compatibility with multiple hardware devices, were all proposed by Lori and the team.

In the process of innovation, everyone's roles are different but equally important. Without the approval of the senior leaders, the resources for innovation may be limited (of course, sometimes too many resources may not necessarily be a good thing). Without the wisdom and contributions of the team, constantly solving real-world execution problems, innovation is also difficult to emerge and land continuously and stably.

"The Lookout": Why is it difficult for Chinese to B software companies to grow into billion-dollar companies, regardless of whether they are hardware or software products?

Indeed, in China, it is challenging and difficult to persist in developing general software services that are compatible with multiple hardware devices without exclusively doing hardware products. Although Tencent Meeting is already a market leader in China, it is still in the red.

The development of the software service industry in China is somewhat deformed, with too much emphasis on hardware and too little on software, which is significantly different from the global trend of "Software is eating the world" and "Software Defined This and That." In many bidding projects in China, customers often only reserve a small proportion of the budget for software products, with the majority being allocated to hardware. Some manufacturers often use hardware profits to subsidize software, using vertical integration to capture the market. Therefore, the profit margins of domestic software companies are extremely small, and it is difficult to cultivate software companies of the size of Microsoft, Oracle, and SAP, and the gap between the industrial scale and the global software market is constantly widening.

In the global market, chip, hardware, and software vendors have their respective dominant players, and the industry ecology is booming. Each industry has its own market and profit space, and diversified competition constantly drives technological innovation, which should be the long-term goal of the development of China's technology industry.

"The Lookout": Talking about large models, why does the outside world feel that Tencent is "not in a hurry"? Is it because they don't see any definite opportunities?

The outside world doesn't feel that we're not in a hurry because they may see some enterprise founders also becoming KOLs, constantly sending out strong signals to the market. Of course, this is a very good product promotion, which not only attracts industry attention, but also investments. But we are not good at this, and maybe that's our shortcoming.

But we are not in no rush, we just rarely deliberately talk about the slogan of "All in." For a long-cycle business, the more fiercely Resources are invested in the early stage, the greater the difficulty and pressure of persistence, and the less resilience and patience to cross the cycle. Our investment in AI is long-term, and since the goal is very clear, we run at a rhythm.

In this process, for Tencent's senior management, we will continue to understand the industry from different perspectives, verify ideas by using products, try and iterate products. Many times, our feelings come from our own experience. We use some products on the market and truly feel (Get a feeling) to what extent this technology has developed? What is the problem?

Tencent has always been using this product development strategy in the past. Our judgment of the market and the product is often based on the user experience. So when we develop C-end products, it goes more smoothly because everyone can use them, and you immediately know whether they are done well or not, and it's difficult for others to deceive you. By using, experiencing, and feeling, we get some direction and methods to guide our team.

Games are quite typical. In critical battlefield situations, such as with PUBG and "King of Glory," we all played together every day.

4. Discussion on organizational change and individual development: "I have a '721' method."

"Qianwang": After Tencent's "930 transformation," you transitioned from managing the Social Network Group (SNG) to the Cloud and Smart Industries Group (CSIG). Doing B-to-B work in a company with a strong C-to-C culture requires some exploration. Have you found any good methods?

Tang Daosheng: Definitely, you can't apply the successful model of Tencent's past C-to-C businesses, but you must find a starting point from the accumulated technical advantages and platform resources to create Tencent's unique approach to B-to-B. For example, we take advantage of the WeChat ecosystem and use WeChat scan codes offline to connect mini-programs with WeChat for Enterprises to create an integrated CRM system for sales, services, and marketing, which has been recognized by many C-end enterprises. Another example is Tencent's content advantage, which was an important entry point for early CSIG services to car companies. Once we have identified our entry points, we can gradually expand to more opportunities for cloud product cooperation.

B-to-B businesses are very complex, and each customer and project requires multiple roles and clear division of labor, as well as sophisticated processes and systems to support them. We need to focus on self-developed products, ensure that the assessment rules and goals are consistent, and let partners provide services around our products and develop a broad ecological partnership. These are all methods that we have gradually explored after stepping on and filling in the potholes.

"Qianwang": The CSIG has been roughly divided into two stages since its establishment, with a series of major adjustments made by you around 2022.

Tang Daosheng: The first stage is a period of extensive development. After the establishment of CSIG, we established many industry teams, each integrating internal and external resources, developing industry solutions, finding many partners, and also introducing some integrators. At that time, we set a revenue target and then grew wildly without managing it in detail.

Although the public cloud grew rapidly, many products were sold at low prices or even at a loss in order to win large customers. The supply chain management was also relatively crude, with low resource utilization and high costs, resulting in a very low gross margin and significant losses. As for the private cloud, many salespeople took advantage of large integrated projects to win many regional projects. Although they tried to deliver with many partners, many integrated projects were not delivered well, and many historical issues are still being dealt with today.

During this stage, many teams were able to inflate their revenues, although they had also achieved their revenue targets, the resale of other people's things would not accumulate equity value for our products—salespeople will say that the resale is allowed, and then say they have tried their best to push our products. In the end, some partners even focused on our sales, allowing salespeople to give all the orders to them.

Afterward, we tightened the revenue recognition rules. For example, only gross margins are recognized for resale, and third-party hardware and SMS are also only recognized for gross margins. External whitelists for resale and selected products were basically cut by more than 90%.

Another example is that delivery services do not count as self-developed income. This allows the product team to avoid the consumption of manpower in the delivery and customized development income. This will greatly promote partnerships to provide delivery and customized development services as their business model, making them more willing to promote and recommend our self-developed products. This also forced us to strengthen our product usability, eco-partner training, and certification system.

"Qianwang": How long did you spend preparing for this transformation?

Tang Daosheng: A long time ago. In the early years, I urged salespeople: Hey, it's still meaningful to sell our own products. Don't blindly pursue large orders, or we'll end up with too many things we can't deliver. But talking to them was useless. Although everyone said: Yes, sir! I understand. I'm already pushing hard on this!---But in the end, everyone's behavior is driven by sales rules and assessment rules. So I know I have to adjust the rules.

Adjusting the rules resulted in many setbacks. Many people said this was not enough for B-to-B. Many people doubted you and thought that income would decline. This process was painful. But if you need to make decisions, you still have to do it. So, we are determined to squeeze out unhealthy revenue.

After more than two years of adjustment, a series of management methods gradually began to show results, and we began to do more healthy things.

"Qianwang": Could you list some of the most important rules you adjusted?

Tang Daosheng: For example, only gross margins are recognized for resale, including third-party hardware and SMS. External whitelists for resale and selected products were basically cut by more than 90%.

Another example is that delivery services do not count as self-developed income, allowing the product team to avoid the consumption of manpower in the delivery and customized development income. This way, we can vigorously promote the provision of delivery and customized development services by partners as their business model, making them more willing to promote and recommend our self-developed products. This also forced us to strengthen our product usability, eco-partner training, and certification system.

"Qianwang": Does Tencent Cloud still aim to be number one in China?

Tang Daosheng: I don't care much about our market share ranking in the overall cloud market. I and some heads of research institutions have also said that many of the scoring items in the ranking are not very rigorous, and often compare apples to oranges. Many vendors' provided data have a lot of fraud. For example, most cloud service providers still regard SMS revenue as part of cloud business revenue, and many integrated project revenues are packaged under the concept of "cloud".

Now I am more concerned about the market share and penetration rate of our products in each sub-segment track, such as whether we are the first in audio and video services? Where am I in the database? In the field of big data, is the market share rising? And so on.

"The Lookout": Dowson, you are already China's top professional manager. Do you still have any worries?

Tang Daosheng: There will always be worries. There are many things in life that are not satisfactory or uncontrollable. (Laughs)

"The Lookout": I remember in 1998-1999, you started a venture in the United States, and the Internet was about to bubble burst.

Tang Daosheng: At that time, I had just left Oracle and had once done a multi-source information (crowd sourcing) software-related knowledge Q&A website. It is difficult to charge users for knowledge Q&A. There are too many free information on the Internet, and many enthusiastic people contribute their experiences in community forums. Later, it was determined that this model could not work, so it was abandoned.

Then I joined a start-up company that provides enterprise-level email services: Sendmail. Shortly thereafter, the Internet bubble burst. This company was short of cash and continued to lose money. It needed to shrink the team to reduce costs. I survived in this process-I was originally one of four R&D managers, and the company chose me. In this process, I learned to focus on the most important things, select elite teams to improve efficiency, encourage the team to maintain a positive attitude, and face long-term goals.

"The Lookout": What was your mentality when you entered Tencent at the age of 32? Was there an interview?

Tang Daosheng: I came to Shenzhen from the United States during the May Day holiday for an interview. As soon as I walked into the lobby and took the elevator to the office, it was all black and the lights were not on, which did not feel like a serious office place. In the United States, the lobby and elevator entrances of office buildings are also not turned off during holidays. I was a little worried about the situation of this company at that time. (Laughs) I walked to Tony's (one of Tencent's founders, the original CTO Zhang Zhidong) office in the dark and chatted with him.

Tencent was not a very large company at that time. What impressed me was that I had been doing enterprise software services in the United States, but here I could join a more promising Internet industry. To be honest, I didn't see any instant messaging company in the United States that could go public independently, and instant messaging was a relatively marginal small business in the global market. AIM and Yahoo! Messenger are both free and lose money. There is no business model. At first, I still had many questions about Tencent's development prospects.

But later I understood more and more about Tencent's business model and advantages, and also found many opportunities for learning and growth within the company, so I stayed in Tencent until now.

"The Lookout": In your opinion, what is an outstanding leader?

Tang Daosheng: There are too many different styles of leaders who can succeed, and I think outstanding leaders do not have a fixed portrait. I can only say that my requirement for myself is to think one step further and lead the team in the right direction. in business management, resource allocation needs to be rhythmic. I have a "721" method. 70% of resources are invested in the most important core business in the short term, investing enough energy to ensure success; 20% is invested in new products and opportunities in development, which are relatively clear and may become commercialized in the next two years, which can gradually supplement the growth of mature businesses; the last 10% is invested in frontier technologies that are still in early development, which is oriented towards future opportunities. It may take three to five years or even longer to bring commercial returns.

There are generally many implementation routes in parts 7 and 2, and the strongest and most experienced teams are often invested here. However, outstanding managers need to have judgment and dare to take risks in part 1, where the situation is unknown.

"The Lookout": Are "good leader" and "good person" mutually exclusive?

Tang Daosheng: It's not called exclusion, but you have to choose in some places. Everyone not only wants to come to work happily every day, but also has professional development aspirations and hopes to do something in a winning team. A basic condition for a good leader: you must be able to get things done. If you can't do it, you are definitely not qualified. To achieve the goal of the team, you need to make correct decisions. When encountering some members with insufficient ability who cannot contribute to achieving the team's goals, you need to replace them. From the perspective of the replaced, you are a bad person. From the perspective of others, "Ah, finally replaced the person who can't do it." For them, you are both a good leader and a good person.

You must achieve success as a good leader. If you cannot do it, you are definitely not eligible. A basic condition for a good leader: you must be able to get things done. When you can't do it, you definitely can't. You need to make the right decisions. When encountering some members with insufficient ability who cannot contribute to achieving the team's goals, you need to replace them. From the perspective of the replaced, you are a bad person. From the perspective of others, "Ah, finally replaced the person who can't do it." For them, you are both a good leader and a good person.

"The Lookout": which book has the biggest impact on you during your growth?

Tang Daosheng: "The Seven Habits of Highly Effective People". Be proactive.

(Note: It was called that in the 1990s version, and later versions changed it to take initiative.)

Being proactive doesn't just mean taking initiative. Being proactive means controlling your emotions—you can't let others make you unhappy, only you can make yourself unhappy; no matter how much the environment harms you, you can still avoid being hurt. This has been very helpful for me when facing difficulties in work or setbacks in life, helping me regain positive energy quickly.

"Qian Wang": What do you do in your spare time?

Tang Daosheng: I mainly watch videos and different kinds of content. I like to absorb new things like a sponge. But the things I watch are diverse, I watch everything. I watch a lot of popular science content, such as content related to the semiconductor industry, life sciences, neural networks, and so on.

I also like to watch interview-style content, such as Lex Fridman's podcast. I recently came across an Indian blogger who interviews people from different fields in India, including politicians, businessmen and investors. I found it very interesting. For example, there was a segment discussing the brain drain in India, and how Indian elites think that Indian talent flowing to the United States expands the global influence of Indians and is not a bad thing. I think this viewpoint is also very inspiring.

I also like to watch movies and TV shows in my free time to relax. I just finished watching "Joy of Life 2" at 1.5x speed (laughs).

"Qian Wang": Jim Collins has a viewpoint in "From Good to Great": technology should be used as an accelerator rather than a driver. Excellent companies never use technology as the main means of change, but they are pioneers in applying carefully selected technologies.

Tang Daosheng: Yes and no. Some companies rely on a new technology to build their entire system. For example, do you say it's because of the GPU that there is NVIDIA? Or is it because there is NVIDIA that there is a GPU? It's hard to say.

But as engineering students, we believe that technology is a means, a tool, and the core is to see what problems you are solving. So we first need to figure out what the problem is and what kind of effect we want to achieve, and then find the tool. The more a person understands technology, the more tools they have in their problem-solving toolbox, and they can put them together more effectively. This process also requires a good understanding of the boundaries of each tool, what can and cannot be done, and gradually combining different abilities.

"Qian Wang": Finally, let me ask a bigger question. Companies like Tencent, born in the last technological upheaval, have 20 years of history. What kind of attitude should they have towards the next round of technological change?

Tang Daosheng: Stay open, stay curious. Like the "721 rule" for resource allocation I talked about earlier, often missing out on some big opportunities is due to not focusing on the long-term transformation, or not investing enough or lacking curiosity.

Facing uncertainty, some people do not like it or are not willing to, but you should always treat it with reverence.