Dialogue with Ji Xinhua, CEO of Youkede: Competitive big models don’t just lie down and win if they have a card

Author| He Sisi

Edit | Zhang Jin

Image source: Generated by Unbounded AI

"To a certain extent, the prohibition of computing power in the United States will limit the development of domestic large-scale models, because there is no way to train models without core computing power." Regarding the importance of computing power for large models, Ji Xinhua, Chairman and CEO of UKED Say so.

Since the large model became popular in China in February this year, computing power has instantly become the most concerned topic in the industry, and it has also become a hurdle that is difficult to overcome in domestic large model training. In fact, from another perspective, the explosive growth of computing power has also brought unprecedented development opportunities to domestic cloud computing vendors.

In this regard, Ji Xinhua also agreed. He said that the disabling of computing power is a bad thing and a good thing. On the one hand, if you want to solve the bottleneck problem within one or two years, you need to speed up the research and development of domestic chips, which is not an easy task; on the other hand, enterprises are paying more and more attention to Computing power, this will instead promote the development of domestic chips.

As a neutral cloud computing manufacturer, Youkede has long felt that behind the competition of large models, the computing power market will usher in huge changes.

"I came into contact with Stable Diffusion for the first time in November last year. At that time, I thought its effect was amazing, so I started to pay attention to the large model." Ji Xinhua explained to Leifeng.com why he paid attention to the large model. At the same time, he also revealed that because At that time, Chinese companies hadn't started to pay attention to large-scale models, and they didn't know how much value it could create, so they didn't make up their minds to do large-scale model-related businesses at that time.

The real application of large-scale models and the service of large-scale model enterprises by UK will start in February this year, which is also the time point for large-scale models to go out of the circle in China.

Leifeng.com learned that UKED first developed four AI products for internal use based on the large-scale model, the knowledge question-and-answer platform "Shiwen", the UCoder code assistant platform, the AI painting assistant platform, and the large-scale model security management platform. For a low-cost, high-value-added self-built data center, UKED has created an AIGC computing power base, which flexibly provides a variety of GPU computing power resources for large model training, reasoning, and data processing.

Immediately afterwards, UKED launched a privatized large-scale model all-in-one machine, which built-in the UCloudStack full-stack private cloud platform independently developed by UKED, providing privatization solutions for virtualization, storage, network and MaaS models, combined with industry vertical large-scale models. Enterprises can deploy large model applications with one click.

Regarding the positioning of UKED in the era of large models, Ji Xinhua once again emphasized the principle of "neutrality". He explained that neutrality means that UK not only has no competition with customers, but also helps large model companies find customers. In this regard, Ji Xinhua also publicly teased that in the era of large models, what Youke has to do is "matchmaker".

From Ji Xinhua's answer, we can also indirectly get the reasons for UKED's internal R&D and application of large-scale model products. One is for the company's employees to understand large-scale models and learn to use large-scale models. On this basis, they can deeply understand large-scale models. What are the pain points, and in which scenarios can it be implemented, so as to better serve customers.

When it comes to the future of large models, Ji Xinhua said that there are too many uncertainties in the future, but in any case, the Internet industry or artificial intelligence industry needs computing power. The interconnection capabilities are limited, so UKEDE's future work will focus on building a large-scale computing power base.

The following is the conversation between Leifeng.com and Ji Xinhua:

Competitive large models do not just lie down and win if they have cards. Engineering capabilities are crucial to model training.

**Leifeng.com:**Many people are saying that only a few cloud vendors can finally participate in the competition of large-scale models. Not all of them have high-speed networks. What do you think are the barriers to competition?

Ji Xinhua: That’s right, technologies like high-speed networks are not the main threshold for competing big models. Including Youkede and most cloud vendors are capable of doing it. Now the network has two structures: the first is the RoCE network, which UKED already has this capability in 2019.

The second is the IB network solution, which is recommended by Nvidia. This solution is relatively simple and only needs to be deployed and maintained. Therefore, for cloud vendors or particularly large companies, technology is not the main threshold.

**Leifeng.com:**Actually, it is said that many large-scale enterprises have started to stock up on cards?

Ji Xinhua: Yes, especially some major manufacturers are hoarding cards, including A800, H800 and so on. On the one hand, its own AI business needs to use cards, and with more capital investment, it will purchase a large number of cards before; on the other hand, domestic large-scale attention to large models began in February this year, and various manufacturers Because of the importance, it will also speed up the speed of hoarding cards.

Leifeng.com: Does it mean that the more cards you have, the more you will win? How do other cloud vendors compete with big players?

Ji Xinhua: I didn’t lie down and win. Many AI large-scale model companies are now using our cards. This phenomenon is very common.

There are several reasons: first, it is not enough to use only one card, and second, why is the big model company willing to cooperate with Youkede? It is because they feel that big factories have computing power, algorithms, data and scenarios, and they are worried about business competition in the end when they cooperate with big factories. As a neutral and secure cloud vendor, Ukerd has no competition with large-scale enterprises. At the same time, UKDE is able to achieve computing power platforms, model libraries, etc. from data centers and underlying architectures, with profound technical accumulation and one-stop system engineering service capabilities.

Do a good job in the computing power service, and do a good job in the role of the big model "matchmaker"

**Leifeng.com: **When did Youkede start paying attention to large models?

**Ji Xinhua: **The earliest contact was during the National Day last year. The first thing I saw was Stable Diffusion. At that time, I thought its effect was amazing, so I started to pay attention to the large model, so we followed up on ChatGPT earlier.

Leifeng.com: So you made up your mind to do this during the National Day last year?

Ji Xinhua: After the National Day last year, UKED set AIGC as the goal for this year, and released the AI painting platform image in November last year. In February of this year, after communicating with leading large-scale model companies in China, I realized that the domestic computing power market will have an explosive development, so I made up my mind to do this at the end of March.

Leifeng.com: How many potential customers are there in these large-scale model enterprises?

Ji Xinhua: We concluded that there are 130 large-scale model companies in China, 78 general-purpose companies, and 52 vertical companies. And it is still increasing, more than 30 of them are already our customers.

Leifeng.com: In what form will the service be exported?

Ji Xinhua: One is our computing power, and the other is computer room services, because some companies buy their own equipment and put it in our computer room.

Leifeng.com: How do you understand the computer room service? Is it because the customer itself has no operational capabilities?

Ji Xinhua: The enterprise itself needs a computer room no matter where it is. The computer room required by a large model has two characteristics: one is that it consumes a lot of power, and the other is that the power consumption of a H100 machine exceeds 10kW. The computer room cannot meet this demand, and the Ulanqab data center of Youkede is especially suitable. In addition, many companies now have the problem of having servers but cannot use them. Ukerde can help them build a computing power platform and carry out subsequent maintenance work.

Leifeng.com: In addition to providing computing power, Youkede also released "Shiwen" some time ago?

Ji Xinhua: UKED has built four internal platforms: knowledge question and answer platform "Shiwen", UCoder code assistant platform, AI painting assistant platform, and large model security management platform. In fact, before the "knowledge", the first thing we did was the large-scale model application management system. This is our first product, and we hope that everyone can use the large-scale model in the future;

Second, in order to solve commercial security issues, we have made some restrictions, including your questions and uploaded files, we will record and filter them to prevent the company's confidential information from being leaked on the Internet;

Thirdly, for user problems, including internal employee problems and external communication and feedback, the system will automatically record, so that the company can continue to carry out its own model training in the later stage.

Leifeng.com: Why did you make these four products? Can it go outside?

Ji Xinhua: First, so that the employees of the company can understand and use large-scale models. On this basis, they can deeply understand what the pain points of large-scale model companies are, and in which scenarios they will be implemented, so as to better serve customers. Provide services.

These four products are currently used internally by the company, but if customers need it, we can also communicate more.

Leifeng.com: Which model are these platforms based on?

**Ji Xinhua: **Train with GPT 4 first, conduct model verification, and then gradually use domestic large models or open source large models to optimize. In this regard, I also mentioned a concept called "sharpening guns abroad and fighting at home".

Leifeng.com: What other scenarios are you optimistic about in the future?

Ji Xinhua: If it is distinguished according to the tolerance for the inaccurate nature of the output content of ChatGPT, we have divided 10 scenarios.

The first is translation and dubbing. The ability in this area is already very high and can completely replace humans; the second is the NPC of the game; the third is social interaction; the fourth is the content output of e-commerce; the fifth is game design; The sixth is customer service; the seventh is document and programming assistance; the eighth is knowledge management within the enterprise; the ninth is education and insurance scenarios; and the last is assisting lawyers and doctors.

Leifeng.com: Will these 10 scenes be done at the same time? Or step by step?

Ji Xinhua: Ucar does not make large models, we just connect customers and partners, which can be understood as the role of "matchmaker". For example, connect game customers to MiniMax, and connect customers in the e-commerce and education industries to Zhipu Huazhang.

It is difficult for domestic large-scale models to catch up with GPT4, and there are more opportunities for start-up companies

Leifeng.com: How many types do you think domestic large model companies can be divided into? Who are the key customers of Youkede?

Ji Xinhua: We are divided into five categories. The first category is giants, including Ali, Baidu, Toutiao, Huawei, JD.com, etc. The second category is started by scientists, such as Zhipu Huazhang; the third category is the original AI company, AI Four Tigers, Daguan, Yunzhisheng, 4Paradigm, etc.; the fourth category is startup companies, such as MiniMax; The fifth category is that the original listed companies switched to large-scale models, such as Kunlun, 360, and also include entrepreneurial leaders such as Wang Xiaochuan and Li Kaifu.

The first type of large manufacturers are not the target customers of Youkede, and the latter types are our key customers.

Leifeng.com: That is to say, big factories have the ability to build themselves, but other companies do not have the ability to build themselves?

Ji Xinhua: Because it involves the field of artificial intelligence, it is not only a network problem, but also a series of problems such as storage and computer rooms. For example, start-up companies can also build their own computing power, but the cycle will be very long. Wait for him to build it himself After that, the competition is over. The competition of large models is all about speed, and whoever has the fastest speed may occupy the commanding heights.

Leifeng.com: For a big factory, how many stages and nodes are there for a big model?

Ji Xinhua: The first is the launch of the large model, and the second is the verification of the effect after the launch. At present, it is known that companies such as MiniMax, Zhipu, Baidu, Ali, HKUST Xunfei, etc. have launched.

**Leifeng.com:**Which one is better, does it mainly depend on how much computing power it uses?

Ji Xinhua: I don’t think so. More computing power for training may not necessarily produce a good model, but if the inference link is used more, it means that there are a large number of users, and more feedback will be obtained. It is conducive to training a good model.

**Leifeng.com:**Regardless of training or reasoning, to build a large model, you must first have computing power?

Ji Xinhua: Yes, the first thing is to have a card. If you don’t have a card, you will definitely be behind. Around 40-50%. If the above two problems are solved, it is a security problem, and the traffic problem is also very important.

Leifeng.com: What is the level of domestic models? Many people say that it will catch up with GPT4 by the end of this year.

Ji Xinhua: At present, there is no model that surpasses GPT3.5 in China. Of course, it is actually very easy to surpass GPT3.5 in a certain aspect. It is difficult to surpass GPT4. The key is that the papers before GPT3.5 are public, but GPT4 has not yet been made public, so everyone does not know what to do.

Leifeng.com: Don’t these big domestic manufacturers have a chance to catch up?

Ji Xinhua: I think the effect of big companies is not as good as that of start-up companies. Because some start-up companies with beliefs and ideals have already started to do it, and Dachang actually started after seeing the popularity of ChatGPT.

Leifeng.com: It stands to reason that large manufacturers have accumulated technology, so don’t they have more opportunities?

Ji Xinhua: There are many directions for AI. In the past, most companies thought that vertical large-scale models had opportunities, but did not think that general-purpose large-scale models had opportunities. Therefore, in the past few years, it was some entrepreneurial teams with conviction and Scientific research teams, such as Zhipu Huazhang, MiniMax, Chinese Academy of Sciences, etc.

For big manufacturers, it is more to see that foreign countries follow up after doing this, and it has not risen to the company's strategy, so they do not have a lot of accumulation in general-purpose large models.

Leifeng.com: The problem of computing power will be resolved sooner or later, so what are the advantages of UKEDe in terms of large models?

Ji Xinhua: It can be summed up in three points: First, maintain neutrality and have no competitive relationship with users. Second, it has the ability to build a complete set of AIGC solutions for data centers, networks, and data storage, which can help small and medium-sized companies quickly build platforms and solve the efficiency problem of using large models; third, it can better help large model companies expand customers.