TENCENT "Bulk" Large Model

Under the wave of large models, the Internet industry is surging, and everyone is looking forward to who can take the lead in running China's GPT.

Even if we exclude the national teams of universities and start-up companies, major companies that have already flexed their muscles in the field of large models, such as Baidu, 360, Alibaba, Kunlun, Shangtang, Jinshan, and IFLYTEK, can still make a list. Chinese technology giants have always liked to chase the trend, and the era of large models is no exception. However, there is one giant that is very calm and has not shown any major moves so far.

As early as April 2022, TENCENT officially launched its Huan Yuan large model. It has made progress in many areas such as computer vision, natural language processing, multimodal content understanding, copywriting generation, and text-based video. However, it gives the outside world a feeling of "silence" now.

Since February of this year, TENCENT has only made two major moves in the field of large models that have been perceived by the outside world. One is the establishment of the "Hunyuan Assistant" project team for the ChatGPT-like dialogue product on February 27th, with Zhang Zhengyou, the highest professional rank holder in TENCENT's history, as the project owner. The second is the release of the new generation of HCC (High-Performance Computing Cluster) high-performance computing clusters on April 14th.

While other competitors are showing their strength and voice, striving to gain more recognition in the era of large models, TENCENT's silence may be a step slow. What's more interesting is that TENCENT has been falling since March 29th, and major shareholders led by Naspers have been continuously reducing their holdings and cashing out. This may also be a manifestation of the capital market's dissatisfaction with TENCENT's silence in the large model race.

Is TENCENT still "waiting for the wind"?

"Silence" is just the surface

In the Internet industry, efficiency is everything. Even OpenAI, a company that keeps burning money, occasionally releases opinions and information to the outside world, trying to briefly attract attention in the modern society of information explosion. TENCENT is no exception.

Let's take February 27th as a node and look for clues about TENCENT's large model training progress from the actions announced by TENCENT to the outside world. According to incomplete statistics from Photon Planet, TENCENT's actions since February 27th have focused on three areas: AI commercialization, video accounts, and government-enterprise cooperation.

In fact, TENCENT's actions in various aspects are not small. In addition to reaching cooperation with China Mobile and China Unicom in cloud computing, on March 30th, TENCENT released the AI intelligent creation assistant "TENCENT Zhiying". In April, TENCENT released the smart travel car cloud on the eve of the Shanghai Smart Station, launched the allegedly most powerful large model computing cluster in China, and broke the ice with ByteDance to explore video cooperation. When May arrived, TENCENT's Weishi recruited creators for top platforms such as Douyin, Xiaohongshu, and Kuaishou, and released the "Enlightenment" AI open research platform on May 9th. It also opened QQ channels as carriers. AI drawing tool Midjourney launches official Chinese version beta.

Compared with other companies that have released universal large models and started exploring their applications, TENCENT may not be "quiet" despite its small voice. In fact, whether it is the "Enlightenment" AI open research platform, "TENCENT Smart Shadow" AI intelligent creation assistant, the release of computing clusters, or Midjourney-related actions, they are all closely related to large model training.

Intelligent travel car cloud is a "car-cloud integration" service based on cloud services and AI training. "Enlightenment" itself was launched by AI Labs in conjunction with the King of Glory project team. It is a platform for AI training borrowed from game scenes. The application "TENCENT Smart Shadow" that supports virtual digital people, text dubbing, intelligent AI painting, and article-to-video conversion also belongs to the functional category of large language models. Midjourney's open beta is an exploration of AI's native application in the C-end. Efficient computing clusters are also an indispensable infrastructure in large model training.

Digging deeper, these three projects actually belong to different business groups within TENCENT. "Car-Cloud Integration" belongs to CSIG (Cloud and Smart Industry Business Group), "Enlightenment" belongs to IEG (Interactive Entertainment Business Group), and "TENCENT Smart Shadow" should belong to WXG (WeChat Business Group). Large model computing clusters belong to TEG (Technology Engineering Business Group), and QQ's application exploration belongs to SNG (Social Network Business Group).

This inevitably reminds us of TENCENT's characteristic "internal horse racing" mechanism, which is the process of raising different competitors in the past TENCENT business lines to determine the winner. Photon Planet learned from someone close to TENCENT that CSIG, TEG, and PGC are all large models within the training business scope.

However, considering the high investment in large model training, internal horse racing will cause greater sunk costs. Therefore, the current "horse racing" phenomenon is more likely to be led by TEG, with each BG providing corpora and data set annotations to create exclusive models for each BG's business line, and finally summarized into the mixed AI through knowledge distillation.

The benefits of this strategic layout are obvious. It can not only mobilize the whole situation to provide high-quality data sets but also explore feasible paths for commercial scenarios to land in long-term model training. However, the drawbacks are also significant. It is difficult for various BGs in large organizational structures to cooperate with each other, and the cumbersome organizational structure inevitably makes it difficult to exchange results.

"We initially thought this was an opportunity that the Internet had not seen in ten years, but the more we thought about it, the more we felt that this was an opportunity that had not been seen in hundreds of years, similar to the industrial revolution of inventing electricity. So we think (AI) is very important, but it does require a lot of accumulation." Combined with Tony's speech at TENCENT's first-quarter earnings conference, TENCENT is trying to explain its "not fast or slow" to the outside world, but in the current situation where everyone is making efforts, the response that has not revealed specific progress is always a bit pale.

AI Labs, which is not well received, is now in the spotlight

No matter how many large-scale model applications TENCENT eventually lands through the "Tian Ji horse racing" strategy, the ultimate goal is to ensure that the top "element" can help TENCENT stand firm in the AI era and even reproduce the social glory of the mobile Internet era.

Therefore, AI Labs, which has always been on the edge of the organizational structure, has stepped onto the center stage, enjoying the most concentrated resources while also bearing unprecedented pressure.

This is because large-scale AI research laboratories, represented by TENCENT AI Labs and Alibaba DAMO Academy, usually conduct more advanced and general algorithm research. The research results are usually output to other BGs through API interfaces or SDKs. Specific scenarios with strong business nature generally require customized algorithms to be made on demand. Compared with leading the business BG as the coordinator, it is more likely to help the business BG build products with cutting-edge technology in the position of the middle platform.

According to a person close to TENCENT, AI Labs has always given other BGs the impression of "not doing practical work and publishing papers". The reason why colleagues have such a stereotype is that many cutting-edge research results have not been put into production, "not because researchers are slacking off". One of the more typical examples is the intelligent speaker optimized by AI Labs many times. As an employee gift for TENCENT's 20th anniversary celebration, it was also at risk of being stopped.

It is understood that the intelligent speaker project was launched in April 2018 and belonged to the Intelligent Innovation Business Department under TENCENT's Mobile Internet Group (MIG) at that time. On the first day of its release, it achieved a sales volume of 20,000 units on JD.com. However, eight months after the launch of the intelligent speaker project, CSIG, as TENCENT's business focus at that time, also launched its own intelligent speaker project, Dingdang screen speaker. The problem of project division and the subtle "market competition" relationship between different BGs are intriguing, which may also be the reason why the intelligent speaker project was stopped.

"When a group of people who should specialize in algorithms and brainstorm the next stage of the company's core competitiveness are exhausted to deal with the products of various business BGs, and even end up with the 'Tian Ji horse racing' strategy, can the company still expect them to do academic research well?" said a person close to TENCENT.

Technology monetization is an eternal proposition for the technology department of large companies, and the key is whether they are in a hurry or not. A consultant told Photon Planet that compared with becoming a "algorithm outsourcing worker" in the middle platform of a large company, whether it is entering the business department of the industrial sector or turning to academic research in universities, it is a more cost-effective choice. This may also explain why Zhang Tong, the former director of TENCENT AI Labs and a Stanford PhD, left the company in December 2018 and returned to academia. If there had not been a practical path to AGI that is currently recognized, such as the emergence of large models, AI Labs might not have become the focus of attention. However, if the role of the middle platform remains unchanged and there is no strong support for its position, there is still a question mark as to whether AI Labs can complete its organizational delivery tasks.

After all, Alibaba, which once launched the story of the middle platform, has already deconstructed the myth of the middle platform through its spin-off listing. Just like the business departments and IT departments of small and medium-sized enterprises that urgently need digitization, the subtle relationship between the middle platform and various business groups has already led the organizational structure of Internet companies into a dead end.

Win or lose, it's all about horse racing

People are slaves to experience. If AI Labs is a result of Tencent's development inertia towards the middle platform, it is clear that there is more than one such inertia. One of the more typical ones is Tencent's long-standing position in Nanshan, its investment expansion logic of "handing half of its life to partners," and its well-known internal horse racing.

In the current context where MaaS (model as a service) is increasingly valued, if large models can be regarded as products, they are likely to be the tickets to the new world. Therefore, in the race for large models, Tencent is unlikely to hand over the role of Columbus to others, especially in the model layer. However, in the application layer, it may rely on partners to establish an ecosystem, with rich applications on WeChat and midjourney, which has begun internal testing on QQ.

As for why Tencent chose the QQ ecosystem rather than the WeChat ecosystem, the reason is probably that the QQ user group is relatively younger and QQ is obviously better presented on the PC side. In addition, this choice also has a strong "trial" nature, as midjourney itself has not yet been presented in a mature product form and was only accessed through APIs on Discord before QQ.

It is worth mentioning that the myth of horse racing in the race for large models has actually been shattered in the algorithmic era with the rise of Toutiao, because the horse racing mechanism is essentially based on redundancy, finding points through resource advantages, and repeating the same thing, similar to the exhaustive method in mathematics. The reason why the myth was shattered is that while horse racing was safe, it also made Tencent lose its agility, and the most typical example is that the short video application has only produced the hope of Video Number in the whole village so far.

However, the horse racing in the race for large models may really work.

Because training large models is not a one-night task, nor is it a vertical category that requires strong agility. The high expected investment in heavy assets, trial and error costs, and the diverse applications presented at the application layer largely offset the shortcomings of the horse racing mechanism. On the one hand, the diversification of applications avoids the waste of resources caused by duplicating efforts, and on the other hand, the unique knowledge distillation of large models makes trial and error no longer terrifying. Even though OpenAI has set the standard, people still have a tolerant attitude towards various large models in China, both at the public level and within companies. Fortunately, the new round of horse racing may not need to determine a winner or loser, and the more important result is beyond the victory or defeat. TENCENT should also thank OpenAI for opening a new era, an era that no longer pursues short-term gains and quick profits, and even the direction and path are already clear, allowing TENCENT, which lacks agility and the sense of core grasping, to return to the chessboard and make moves with its competitors.

However, the large model is not omnipotent. At least, the stubbornness of TENCENT's organizational structure and the fuzzy overall strategy still need to be adjusted.

If AI Labs wants to be the Columbus of TENCENT's ship, then dismantling the department walls within the organizational structure and reshaping the attitude towards innovation will be the top priority. For example, according to a person close to TENCENT, AI Labs has lost many top talents, and their innovative achievements within TENCENT have also landed in many vertical fields after leaving, including some leading ones.

If AI Labs wants to truly discover a new continent, it needs to firmly move in the right direction. This is exactly the ability that TENCENT lost after embarking on the horse racing road, because the essence of horse racing is to have power but not know where to use it. Now, can TENCENT break through the path dependence of the horse racing mechanism and find a point to run out a large model based on the databases accumulated by various BGs in the application layer under efficient collaboration? We are still waiting for this answer.

However, the time left for TENCENT may not be much. With the arrival of summer, and this year happens to be an El Niño year, the electricity problem will become more severe, and the training of large models will undoubtedly be affected.

One slow step may lead to slow progress. The failure of the horse racing mechanism in short videos has turned the once "latecomer advantage" from a myth into a fig leaf. The large model will be TENCENT's best opportunity to face its own problems and prove to the world that it still has the ability to fight hard. After all, even in TENCENT's most valued market value management, major shareholders are still reducing their holdings and cashing out.