Written by | Wu Kun proverb
Edit | Wu Xianzhi
The undercurrent of the Internet industry under the wave of big models is surging, and everyone is waiting for who can take the lead in running out of China's GPT.
Baidu, 360, Ali, Kunlun Wanwei, SenseTime, Kingsoft, iFLYTEK... Even if you throw away college national teams and startups, the big factories that have already flexed their muscles in the field of big models can pull up a list. Chinese technology giants have always liked to chase the wind, and the maritime age of large models is naturally not exempt, but there is one giant that is very calm and has not shown any big moves.
As Tencent, which officially revealed its hybrid model as early as April 2022, has made achievements in computer vision, natural language processing, multimodal content understanding, copywriting generation, literary video and other directions, but now it gives the outside world a feeling of "losing voice".
Since February this year, Tencent has only made two big moves in the field of large models, one is that on February 27, the "HunyuanAide" project team has been established for ChatGPT-like conversational products, and Zhang Zhengyou, the owner of the highest professional rank in Tencent's history, is the project owner; The second is the release of a new generation of HCC (High-Performance Computing Cluster) high-performance computing cluster on April 14.
As friends show their strength and voice, and strive to gain more awareness in the era of big models, Tencent's loss of voice is likely to be a slow, slow step. What's more interesting is that Tencent Holdings has been falling endlessly since March 29, and major shareholders led by Naspers have continued to reduce their holdings to cash out, which may also be a reflection of the capital market's dissatisfaction with Tencent's loss of voice in the big model track.
Could it be that Tencent is still "waiting for the wind to come"?
"Loss of voice" is the appearance
Beishang Guang does not believe in tears, and the Internet giants only believe in efficiency. Even OpenAI, a company that burns money constantly, will release some opinions and information from time to time, trying to briefly gain attention in the modern society of information explosion, and Tencent is the same.
We might as well take February 27 as a node to look for clues of its large model training progress from the actions announced by Tencent. According to incomplete statistics from Photon Planet, Tencent's actions since February 27 have focused on three aspects, one is the commercial application of AI, the other is the video number, and finally the government-enterprise cooperation.
Counting down, in fact, Tencent's actions in all aspects are not small. In addition to the cooperation in cloud computing between the two major operators of China Mobile and Unicom, on March 30, Tencent released the AI intelligent creation assistant "Tencent Smart Shadow". In April, Tencent released the Smart Mobility Car Cloud on the eve of the Shanghai Smart Station, released what is said to be the most powerful large-model computing cluster in China, and also broke the ice with ByteDance to explore video cooperation. When the time came in May, Tencent Weishi recruited creators for Douyin, Xiaohongshu and Kuaishou for the head, and released the "Enlightened" AI open research platform on May 9, and opened the QQ channel as the carrier to let the AI drawing tool Midjourney open the official Chinese version of the internal beta.
Compared with the friends who took the lead in releasing a general large model and exploring applications, Tencent's voice is small but not "quiet". Moreover, whether it is the "enlightened" AI open research platform, the "Tencent Zhiying" AI intelligent creation assistance, the release of computing clusters or Midjourney-related actions, they are actually related to large model training.
Smart Mobility Car Cloud is a "vehicle-cloud integration" service based on cloud services and AI training, and "Enlightenment" itself is launched by AI Labs and the Honor of Kings project team, which is itself a platform for AI training using game scenarios; "Tencent Smart Shadow", an application that supports virtual digital humans, text dubbing, intelligent AI painting, and article to video, also belongs to the functional category of large language models; Midjourney Open Closed Beta is an exploration of the native application of AI to C-end landing; Efficient computing clusters are an indispensable infrastructure for large model training.
Digging deeper, these three projects actually belong to different business groups within Tencent. "Vehicle-Cloud Integration" belongs to CSIG (Cloud and Smart Industry Group), "Enlightenment" belongs to IEG (Interactive Entertainment Group), "Tencent Smart Shadow" belongs to WXG (WeChat Group), the big model computing cluster belongs to TEG (Technical Engineering Group), and QQ's application exploration belongs to SNG (Social Network Group).
This inevitably reminds us of Tencent's characteristic "internal horse racing" mechanism, which is the process of raising tricks between different competitors in the past Tencent's business line and finally confirming the winner. Photon Planet learned from a person close to Tencent that CSIG, TEG and PGC are all training large models within the scope of the business range.
But considering the extremely high investment in large model training, in-house horse racing will cause greater sunk costs. Therefore, the current "horse racing" phenomenon is more likely to be dominated by TEG, with each BG providing corpus and dataset annotation, creating exclusive models for each BG's business line, and finally summarizing into hybrid AI through knowledge distillation.
The benefits of such a strategic layout are obvious, which can not only maximize the overall mobilization and provide high-quality datasets, but also explore feasible paths for business scenarios in long-term model training. However, the disadvantages are also obvious, it is difficult for various BGs in a large organizational structure to cooperate with each other, and the inevitable mountains of complex organizational structures also make it difficult to communicate results.
"At first, we thought this was an opportunity for the Internet to not meet in a decade, but the more we thought about it, the more we felt that this was an opportunity similar to the industrial revolution that invented electricity that was not encountered in hundreds of years. 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 "unhurriedness", but at the moment when each company is working hard to open the book, the response that has not disclosed the specific progress always looks a little pale.
Unpopular AI Labs are on the cusp
No matter how many BG large-model applications Tencent has landed through Tianji horse racing, the ultimate goal must be to ensure that the top "hybrid element" can help Tencent gain a foothold in the AI era and even reproduce the social glory of the mobile Internet era.
As a result, AI Labs, which has always been somewhat marginal in the organizational structure, stepped on the center stage, enjoying the most concentrated resources and at the same time under unprecedented pressure.
This is because the AI research laboratories represented by Tencent AI Labs and Ali Damo Academy usually conduct cutting-edge and high-purpose algorithm research, and the research results are usually output to other BGs through API interfaces or SDKs, and specific scenarios with strong business generally require algorithm customization and on-demand production. Instead of coordinating the business BG as the leader, it is more likely to help the business BG build products equipped with cutting-edge technology with the positioning of the middle office.
Photon Planet learned from a person close to Tencent that AI Labs has always given other BGs the impression of "not doing practical things and pouring paper water", and the reason why colleagues have such a stereotype is that many cutting-edge research has only results and has not been put into production, "not researchers paddling". Among them, the more typical is the smart speaker optimized by AI Labs many times, which was given internally as an employee gift for Tencent's 20th anniversary, and the project also faced the risk of being stopped.
It is understood that the smart speaker project was launched in April 2018, belonging to the intelligent innovation business division of Tencent's mobile Internet Business Group (MIG) at that time, and achieved the first day of sales of 20,000 units in Jingdong Mall on the first day of its release. Only 8 months after the smart speaker project was launched, CSIG, as the focus of Tencent's business at that time, also launched its own smart speaker project Jingle with screen speaker, and the project division problem and the subtle "market competition" relationship between different BGs are intriguing, which may be the reason why the smart speaker project was stopped.
"When a group of organizations that should specialize in algorithms and brainstorm the company's core competencies in the next stage are tired of coping with the products of various business BGs, or even the end of the horse race, even then they still can't fall behind in academic output, how well can [the company] expect it to do?" A person close to Tencent said.
Technology monetization is the eternal proposition of the technical department of large factories, and the key lies in being in a hurry. A consultant told Photon Planet that compared with entering the middle office of a large factory, becoming an "algorithm outsourcer", whether it is entering the business department of the industry or turning to a university to specialize in academia, is a more cost-effective choice. This may also explain why Zhang Tong, a former director of Tencent AI Labs and a Stanford doctor, left in December 2018 and returned to academia.
Without the fact that there is no such currently recognized path to AGI as large models, AI Labs may not be on the cusp. However, if the role of the middle office remains unchanged, without strong voice support, whether AI Labs can complete the tasks delivered by the organization still has to be questioned.
After all, Ali, which once unfolded the story of the middle office, has deconstructed the middle office myth through spin-off and listing, just like the business department and IT department in the small and medium-sized enterprises in urgent need of digitalization do not deal with each other, the subtle relationship between the middle office and the various business groups has made the organizational structure of Internet enterprises enter a dead end.
Success also horse racing, defeat also horse racing?
People are slaves to experience, and if AI Labs is the middle office is due to the inertia of Tencent's development so far, it is clear that there is more than one such inertia. Among them, the more typical ones are Tencent's long-term sitting in Nanshan majestically, the investment expansion logic of "giving half its life to partners", and the well-known internal horse race.
At a time when MaaS (model as a service) is becoming more and more important, if a large model can be regarded as a product, it is most likely a ticket to the new world. Therefore, in the big model track, Tencent will not give up Columbus's role, especially the model layer, but in the application layer may rely on partners to establish an ecosystem, far from the rich applications on WeChat, and recently there is also a midjourney that began internal testing on QQ.
As for why Tencent chose the QQ ecosystem instead of the WeChat ecosystem, it can be imagined that the reason may be that the QQ user group is relatively young, and QQ is obviously better in the presentation of products on the PC side. In addition, this choice also has a strong "testing the waters" nature, after all, midjourney itself has not yet been presented in the form of a mature product, and before QQ, only in the form of API access to discord.
It is worth mentioning that the horse racing of the big model track, the myth of horse racing has actually long been shattered in the era of algorithms for the rise of the headline system, because the horse racing mechanism is essentially built on the premise of redundancy, through resource advantages to find points, repeat the wheel, a bit similar to the exhaustive method in mathematics. The reason why the myth is shattered is that horse racing has also made Tencent lose its former agility while insurance, and the most typical is that the short video application has only run out of the hope of the entire village.
However, horse racing on a large model track may really work.
Because large model training is not an overnight task, let alone a vertical class that strongly requires agility. On the one hand, application diversification avoids the waste of resources caused by repeated wheels, on the other hand, the knowledge distillation unique to large models makes trial and error no longer terrible, even if OpenAI is in front, everyone still has a tolerant attitude towards domestic large models, whether it is at the public level or within the company.
Fortunately, the new round of horse racing may not need to decide the winner or loser, and more importantly, the outcome is more important than the victory or defeat. Tencent should also thank OpenAI for the new era, an era that no longer pursues shortness and fastness, and even the direction path has been clearly revealed, allowing Tencent, which lacks agility and lack of smell for core grippers, to return to the chessboard and fall from the sky with the competition.
At the very least, the stubborn problems of Tencent's organizational structure and the ambiguity of its overall strategy still need to be adjusted.
If AI Labs is to be the Columbus of Tencent's ship, tearing down departmental walls within the organizational structure and reshaping attitudes toward innovation will be a top priority. For example, Photon Planet learned from a person close to Tencent that AI Labs has left a number of top talents, and their innovation achievements within Tencent have also landed in many vertical fields after leaving, and even many of them are leaders.
If you want AI Labs, Columbus, to really discover a new continent, you need to at least move firmly in the right direction. This is precisely the ability that Tencent lost after opening the road of horse racing, because the essence of horse racing is powerful but does not know where to do it. We are still waiting for this answer to whether Tencent can break through the path dependence of the horse racing mechanism, based on the database accumulated by each BG in the application layer, and run a large model under efficient collaboration.
It's just that there may not be much time left for Tencent, with the arrival of summer, this year is El Niño, the problem of electricity consumption will become more and more serious, and the training of large models will be greatly affected.
The failure of the horse racing mechanism in short videos has turned the once "last-come-first-come" from a myth to a fig leaf, and the big model will be the best opportunity for Tencent to face up to its own problems and prove to the world that it still has the ability to fight a tough battle. After all, even in the market value management that Tencent attaches the most importance to today, major shareholders are still reducing their holdings to cash out.