Tencent Cloud Wu Yunsheng: The general model and the industry model are not opposite

Image source: Generated by Unbounded AI

"We would rather solve 100% of a customer's problem than solve 70%-80% of 100 customer problems."

On July 7, at the 2023 World Artificial Intelligence Conference (WAIC) Tencent Forum, when asked why the industry-oriented large-scale model was launched first, Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, replied this way.

Large-scale models are the hottest topic in the AI world this year. Compared with Baidu, Ali and other giants who first launched the underlying general-purpose large-scale model, Tencent focused its attention directly on the industry-before WAIC was held, on June 19, Tencent Cloud officially announced the MaaS (Model as a Service) panorama, which has been The 10 major industries have exported more than 50 solutions, covering many different scenarios such as cultural tourism, finance, media, education, and government affairs.

Naturally, scenario application and landing became the theme throughout the WAIC Tencent Forum.

"General-purpose large models are not the only direction for model applications. Models for vertical industries will become the tipping point of the value of large models." Li Qiang, vice president of Tencent and president of Tencent's government and enterprise business, predicted at the forum.

Wu Yunsheng took OTA (Online Travel Agency) smart customer service as an example, and explained that users often have multiple intentions mixed in the actual communication process. In the process of communication, the intention may also switch at any time.

"(The user) just asked to book the hotel on the 10th, and the machine was about to answer, and suddenly said, let me see the hotel on the 11th." Facing the extremely complicated process in the customer service scene, the general large model cannot perfectly To complete the task, it is necessary to reconstruct some complex models in combination with specific scenarios.

However, the current general-purpose large-scale model is still in the early stage of research and development, and it will face the problem of high cost in industrial applications. Wu Yunsheng said that in some specific scenarios, since the general-purpose large-scale model cannot 100% meet the needs, then "judging the level of the solution cost, there is no Makes too much sense." Although the solutions for different industries and scenarios will vary greatly, there is no problem in improving the efficiency of enterprises by more than 30% in general.

Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence. Source: Tencent

At the forum, Tencent Cloud also introduced the recent important upgrades around the large model. Its two major technical bases - Xingmai Network and Vector Database have all undergone capacity upgrades. The upgraded Xingmai high-performance computing network can increase GPU utilization by 40%, save model training costs by 30%-60%, and improve the communication performance of AI large models by nearly 10 times.

And just on July 4th, Tencent Cloud officially released the AI native vector database. Compared with the traditional method, it is used for classification, deduplication and cleaning of large model pre-training data, and the database can achieve a 10-fold increase in efficiency. Using it as an external knowledge base for model inference can reduce the cost by 2-4 orders of magnitude.

At this year's WAIC, an interesting phenomenon also appeared: more than 30 general-purpose and industry large-scale models were unveiled at the meeting, and they all said to make large-scale models. The issue of homogenization immediately became the focus of discussion: Is entrepreneurship in the field of AI still meaningful? Each big factory has its own big model, if everyone owns it, will the big model still be a Game Changer?

Wu Yunsheng believes that in the initial stage of the development of large models, there is no need to rush to make judgments on these issues. "I would like to see that when the industry is in full bloom, through the combination of technology and industry, explore various possibilities and improve the efficiency of industries and industries." Wu Yunsheng said.

This judgment also comes from his optimism about the commercial potential of AI large models. In the last wave of AI with single-point breakthroughs, many AI companies fell into the dilemma of project and privatization delivery and implementation, and were unable to achieve profitability.

"In the era of large models, the situation may be different from before." Wu Yunsheng believes that with the development of technology, including the development of underlying computing power and GPU chips, the parameters of models that cost hundreds of billions in the past are getting smaller and smaller; Tencent There is also a lot of progress in training reinforcement and reasoning reinforcement, and the cost is dropping rapidly. On the other hand, the possibilities in the application of large models are constantly expanding, and the value is constantly rising.

Today, Tencent is expanding its technology and application ecology around AI large models, and is also strengthening its linkage with the industry. On July 6, the United Nations Industrial Development Organization and Huawei and other partners jointly announced the establishment of the "Global Industrial and Manufacturing Artificial Intelligence Alliance" at WAIC.

Wu Yunsheng also emphasized that the upgrade of the technical base is "practicing internal strength": "No matter whether it is a general-purpose large-scale model or an industrial large-scale model, it needs to have the underlying support capabilities, including huge computing power, data, etc.."

"In the era of large-scale models, openness is very important, and everyone should be open. The underlying technology changes too fast, and the extension capabilities are very wide. When combined with specific industries, there will be a lot of research and development costs." Wu Yunsheng said that only through opening up can more industries Only when experts and various personnel join in can we cultivate a healthier ecosystem and create more possibilities.

The following is the interview record of the media and Wu Yunsheng, edited by 36 Krypton:

**Media:**Tencent Cloud launched an industry-oriented large-scale model at the beginning, rather than a general-purpose large-scale model. Is it a consideration of income?

Wu Yunsheng: This has nothing to do with cost and investment. We have always emphasized that we want to solve customers' problems. We would rather solve one customer's problem 100% than solve 100 customers' problems by 70%-80%. We can make the problem less, but we must solve this problem.

Media: It sounds like the general model and the industry model are in opposition. How do you see the future relationship between the two?

Wu Yunsheng: First of all, I want to make it clear that I personally did not oppose the two. A base mockup is something like a pedestal that solves a need without special customization. The industry model should be based on the general model to effectively improve productivity and serve the public. Only by going deep into the industry can we solve particularly important problems.

The Vincent map function will also have very detailed and specific industry points-for example, generating an advertisement map for a package, and some customers will have special needs, such as some special certification. When faced with practical problems, a different approach is required.

**Media:**What areas will Tencent focus on this year, and what development goals will it have? What are the recent upgrades and iterations?

**Wu Yunsheng:**Our big strategy is to focus on the implementation of practical problems, hoping to solve 100% of customer problems in each specific scenario, instead of finding 100 products to solve 70%-80% of problems. Therefore, we will focus on specific industries and work with customers to solve industry problems.

For example, in the cultural travel industry, customers in the OTA (Online Travel Agency) field will combine their own business scenarios, use large-scale model technology in business processes, and use data-related resources to fine-tune the large-scale model. In terms of technological development, our model and computing power network have been iterated and upgraded, and we will continue to iterate related technologies.

**Media: **Since the last release of the industry model, has there been any significant change in the number of enterprises accessing the model?

Wu Yunsheng: We do have a lot of contacts with companies, and we will honestly study the actual scenarios of customers and how to meet current business needs. "Access" is not specifically defined. There are various needs in the exploration process. Maybe you usually see AIGC more often.

I say something slightly different. We have a company client who does enterprise-level software and needs to make smart forms. For example, in a management meeting, a new form is added. Some items under the form are required, some are not required, and some drop-downs can only select 4 or 5 options. After the option is completed, it needs to be turned into a process, which is approved by A, B, and C, and each person's approval is different.

The original method is to design the form in the system with its own tools and language codes, compile the entire process, and then call the internal organization to realize the process. But the current requirement is to take a photo and put it in the system, and describe it with the system code (own scripting language). Individuals only need simple natural language communication and communication, such as which ones are required and which ones are not required. The first step Where to go, where to go in the second step, use your own system language to design the docking process.

This example is a very specific requirement that cannot be fully addressed by a general model. Therefore, we will have in-depth exchanges with enterprises to see what fields the table is about and what the scripting language is. General-purpose technology may be able to directly solve 60%-70% of the problems, but if customers want to solve 100% of the problems, they need more in-depth communication.

**Media: **How much cost will the company save with the implementation of the large-scale industry model? Compared with general-purpose large-scale models, in which fields will industry large-scale models have advantages?

**Wu Yunsheng: **The cost saved actually varies greatly in different companies and different scenarios. For example, in the customer service scenario, there is a big difference in the size of the customer service itself in the enterprise and how much budget it has. In terms of actual experience, I think it is no problem to increase the efficiency by more than 30%.

When a general large model is faced with a specific industry, it may not be able to completely solve the problems encountered by the industry. For example, what customer service needs is not simple question-and-answer chats, but robots and large models that can understand human intentions, search databases, extract the required information, and then combine them into human-understandable text for replies.

The most important thing is that the user's actual communication process is often mixed with multiple intentions. There are many requirements in one description, and the intention may switch at any time during the communication process. It is very difficult and the process is very complicated, especially when interacting with the customer's system. complex model. This process must not be solved by a general large model, and needs to be combined with specific scenarios.

**Media:**What kind of range can Tencent control the cost of the large-scale model of the enterprise?

Wu Yunsheng: We emphasize that through this technology, companies can reduce costs, increase efficiency, and improve production efficiency, but we will never say what level to control costs. Our product has been released for less than a month, and we have some cooperation in the early stage, but we certainly cannot give overall data.

**36 Krypton:**In the last wave of AI, technology applications headed by CV (image recognition) were more single-point applications, such as calling api billing, but after that, companies started to work on projects and Privatization makes it difficult to make profits. Will AI big models experience the same thing in the future?

Wu Yunsheng: I am still optimistic. Judging from the current time point, there will be relatively big challenges. But whether you look forward half a year or predict half a year in the future, the development of technology is very fast, including the underlying computing power and GPU chips. The large model used to be a model with hundreds of billions of parameters, but with the development of technology, the parameters of the model are getting smaller and smaller, and the capability is still maintained at a very strong level. At the same time, we have made a lot of progress in training reinforcement and reasoning reinforcement, and the cost is falling rapidly.

On the other hand, in terms of application, we see more possibilities, and the trend of application and value that can be generated is constantly rising.

**Media: **What do you think of the balance between underlying capacity building and scenario implementation?

Wu Yunsheng: We have never wanted to look at large models from a single perspective. Regardless of the general-purpose large-scale model or the large-scale industry model, the underlying support capabilities are required, including huge computing power, data, etc., which is the dimension of internal strength. The scene is another dimension. To solve a practical problem, use 50% of the internal strength, the other 30% of the external strength, and add another 20%. In the large model ecology, we look at the problem from different perspectives. But if you only talk about internal strength, there is definitely no problem.

**Media:**Many CEOs have mentioned that the big model is a game changer for the computing industry. Now that all major manufacturers are launching large-scale models, is this judgment untenable? Do we need so many general-purpose large models, or are they already redundant?

Wu Yunsheng: See how to define a game changer. At this stage, the large-scale model industry is in a relatively early stage, and many possibilities have been born. At the same time, we see that large models bring about technological changes and have great potential.

My personal point of view is that there is no need to be too eager to make a conclusion now. I would like to see that in the stage when a hundred flowers are blooming in the industry, through the combination of technology and industry, various possibilities are explored to improve the efficiency of industries and industries.

**Media: **The combination of model and industry is in the early stage. What problems will exist in this stage? There is a view that compared with the general-purpose large-scale model, the cost of the large-scale industry model may not be optimized, but may be higher. What do you think of this point of view?

**Wu Yunsheng: **Large model technology has been around for a short time and is developing rapidly. The industry's understanding of the big model is still at an early stage - I don't know what the big model can do and how deep it can be combined with the industry. There are indeed changes in the relationship between the industry model and the basic model, as well as the cost issue.

There is indeed a view that if a large model solves all problems, the cost will be lower, and there is also a view that for a certain industry, a large model is not needed, and a small model can be used.

This issue cannot be judged from a single dimension, but must be viewed objectively and completely. I have been emphasizing that general large models can solve general problems that are not closely integrated with the industry. But if you want to go deeper, you have to go further in the scene. Many problems seem to be the same, but if you go deeper, you may not solve the same problem. In this case, it does not make much sense to judge the cost of the solution.

Media: From the perspective of the industry, how to judge the market increment and incremental scenario brought about by the large-scale technological change to the entire cloud computing market?

Wu Yunsheng: It can be seen that after the advent of the large-scale model era, the demand for computing power has been greatly promoted, especially for AI-related computing power. However, in terms of the specific quantification of cloud computing, it is difficult to give a number, and it is still in the process of continuous development.

In terms of scenes, all walks of life are now combining large models, and the scenes are very rich, including both general scenes and industry scenes. Common scenarios include smart conference upgrades. Tencent’s Qidian customer service and Qidian analysis released on 619 are also doing smart work. We also do some code assistants on the cloud. In terms of universal application and efficiency upgrade, there are also many applications. In addition, each industry has applications in various industries, which will also bring a lot of demand.

Media: In addition to the large-scale model landing scene, will Tencent provide services for other large-scale model companies? The volcano engine said that 70% of the large-scale model users are in the volcano. What is the data of Tencent?

**Wu Yunsheng:**We provide a series of cloud support or capabilities for unicorn enterprises or other large-scale model solutions. We have released high-performance computer HCC, vector database, and improved acceleration capabilities, which can be provided to manufacturers.

In addition to the underlying capabilities, we also have an integrated fine-tuning solution based on the large model of the TI platform, as well as a series of tools, processes and service support.

**Media:**Many SaaS service providers have accumulated for many years. Is our MaaS (Model as a Service) for them, or for top industry customers?

Wu Yunsheng: These are our customers.

**Media: Combining ** with industry is Tencent’s current approach. Yesterday, Huawei established an alliance. Does it mean that competition is easier to form a contest between giants?

Wu Yunsheng: I don't look at this issue that way. I think openness is very important in the era of large models, and everyone should be open. The underlying technology changes too fast, and the extension capability of the technology is very wide, and it will cost a lot of money to integrate it into specific industries. In this case, only openness can bring out the greatest value. Only by opening up and allowing more industry experts and personnel in various roles to join can the entire ecosystem be healthier and create more possibilities.

Media: Is the opening to each company the same?

Wu Yunsheng: The openness I mentioned refers to capacity building and ecological opening. For example, building a large financial model does not require one person to summarize all the large financial models. Different people have deep experience in different fields, and building together is a kind of openness. Opening directly to industry customers is also a kind of opening.

In addition, based on the capabilities provided by the large model, partners upgrade the application of efficiency tools or industry solutions, and combine different forms of applications. In terms of strengthening the internal strength of the large model, the partners are blooming, and everyone presents different applications, which is also a kind of openness.