The "National Team" ends, and the large model "rolls" to a new latitude

Original source: One DU Finance

Image source: Generated by Unbounded AI‌

After half a year of blowing the wind of the large model, it finally has a new direction.

On July 7th, at the 2023 World Artificial Intelligence Conference (WAIC 2023), the **General Group of Artificial Intelligence Standardization under the guidance of the National Standards Committee announced that the leader of my country's first large-scale model standardization task force will be led by Shanghai Artificial Intelligence Laboratory and Baidu. , Huawei, Ali and other enterprises are jointly responsible. **

The outside world is not surprised by the first batch of selected "national team" lineups. After all, the development of large models needs to be led by players with exceptionally strong technical strength. After the state clarified their status and tasks, the wind direction of the domestic large-scale model market began to undergo new changes.

01 Hurricane for half a year, the industry ushered in the "national team"

Since the beginning of this year, the large model has been soaring all the way, and the speed has exceeded the development process of any previous technology. If in the first quarter of this year, various players flocked to the entrance of the large model, still discussing the issue of "whether to do it or not", by the second quarter, each player has evolved to the issue of "how to do it".

And such a lively scene ushered in a climax at WAIC 2023. **Over 400 companies participated in this conference, and more than 30 large-scale models focused on the highlights. The exhibition area reached 50,000 square meters, setting a new record. **

In this lively conference, many people were unable to enter the conference site because they did not make an appointment in advance. At the conference, which is called "high standard" by industry insiders, Internet celebrity Musk, Yang Likun, one of the Turing Award giants, Hu Houkun, Huawei's rotating chairman, Tang Xiaoge, a professor at the Chinese University of Hong Kong, and academic and entrepreneurial circles The bigwigs showed up one after another.

In the exhibition hall, the large-scale model era, generative AI, and general artificial intelligence, these words that were unfamiliar half a year ago, have now become symbols that can be seen everywhere in the exhibition hall.

Of course, the players of more than 30 large-scale models at the conference did not disappoint the outside world, and gave their own answers to the large-scale models. Especially the actions of the members of the "national team" have attracted the attention of the outside world.

For example, Baidu, as the first manufacturer in China to announce All in artificial intelligence, its exhibition hall at the conference site is particularly attractive. Of course, on this important occasion, Baidu will naturally display the "treasure of the town hall" that more people can experience. This product, called Wenxinyige, allows the audience who enter the exhibition hall to realize the freedom of P pictures.

Huawei moved its "world's fastest AI training cluster" Atlas 900 PoD A2 to the site. Hu Houkun, vice chairman of Huawei, said that using the Atlas 900, people can complete the training of the typical neural network ResNet-50 on the ImageNet dataset in only 59.8 seconds, which is 15% faster than the second place at the same accuracy. "This is equivalent to hitting the line at the top of the sprint field, and then drinking a bottle of water to see the second runner up to the finish line." Undoubtedly, Huawei's hardware-side fundamental computing power show has made industry professionals and audiences awestruck The attention shifted from the complexity of the large model to the competition on the hardware side.

On the Alibaba Cloud Forum, Alibaba Cloud's "Tongyi Family" added an AI painting model "Tongyi Wanxiang". It is said that this model can assist humans in graphic creation, and it can be applied to art design, e-commerce, games and cultural creation in the future. and other application scenarios. Zhou Jingren, CTO of Alibaba Cloud Intelligence Group, said at the scene that this is a key step for Alibaba Cloud's large model to fully grasp the multi-modal capability, and this capability will be gradually opened to industry customers in the future.

"Money, people, technology, and scenes" are the innate advantages of big manufacturers in making large-scale models. At the same time, it is also a gap that many start-ups are difficult to fill. Some people even directly pointed out on the spot that China's large-scale models have landed. , will only be among the five major manufacturers, namely BAT+Huawei+Tencent.

But what are the giants to do? Where it goes next is a big question.

02 Abandoning concepts and feelings, the giants have focused on the scene landing

This year's artificial intelligence conference, the large model has become a well-deserved top class.

Ali Tongyi, Baidu Wenxin, Huawei Pangu and other national teams have shown their hard power. At the same time, more than 30 vertical large-scale models such as Xunfei Xinghuo, Shangtang Rixin, and Netease Fuxi have not lost their momentum. Work hard in their respective fields.

But judging from the situation on the spot, they seem to have abandoned the big and empty, story-telling, and emotional-speaking practices, and instead began to focus on talking about landing scenarios and cases. This is the only way for large models to move forward, and it is also very likely to become the highlight of the next stage.

At the conference, Huawei Cloud Pangu Large Model 3.0 was officially released, attracting many people from the industry to watch. What impressed the industry even more is that what Zhang Ping'an, executive director of Huawei and CEO of Huawei Cloud, said - the Pangu model is very busy, busy doing things, and has no time to write poems. And writing poems is exactly what the players who released the big model in the previous six months love to do most.

In Zhang Ping'an's view, Huawei hopes that the Pangu model can help various industries, such as finance, government affairs, mining, meteorology, etc., rather than focusing on the language model level. According to his disclosure, as of now, the Pangu large model has been implemented in the fields of meteorology, medical research and development, and electric power, and has delivered multiple large models with hundreds of billions of parameters.

Baidu also put the scene into practice. As an early player, Baidu released the Wenxin large model four years ago, but the industry did not pay enough attention to the large model at that time, so that it did not arouse too much splash. But for Baidu, the Wenxin large model is an advanced layout that is one step ahead of the industry. Today, this forward-looking product has also gained a lot.

At the WAIC 2023 site, Baidu Chief Technology Officer Wang Haifeng said that Baidu has now upgraded to version 3.5 of the Wenxin large model. Compared with the previous version, the effect has been improved by 50%, the training speed has been doubled, and the reasoning speed has been increased by 30 times. In terms of cost, it has dropped to 10% of the past.

"Take promoting the prosperity of China's large-scale model ecology as the primary goal, and provide all-round services to large-scale start-up companies." Alibaba Cloud CTO Zhou Jingren said so. Obviously, this continues the MaaS (Model as a Service) concept proposed by Alibaba Cloud.

Tencent, which was the latest to enter the field of large models, has been making constant moves in the past 20 days. On June 19, Tencent publicly disclosed its thinking on large models for the first time; on June 26, it disclosed its self-developed Xingmai high-performance computing network for the first time; on July 7 at WAIC 2023, the vice president of Tencent Cloud and the person in charge of Tencent Cloud Intelligence Wu Yunsheng disclosed Tencent's innovative achievements in the application of large models, and said that Tencent Cloud's industry large model capabilities have been applied to scenarios such as financial risk control, interactive translation, and digital smart customer service, which has improved the efficiency of intelligent applications.

Of course, the large models in subdivided fields also show strong vitality. Tang Wenbin, co-founder and CTO of Megvii Technology, said in an interview with the media: "Application implementation is the only criterion for measuring the value of large models. Megvii Technology will move from visual large models to general multi-modal large models."

**Focus on the implementation of scenarios, and effectively provide enterprise users with cost-reducing and efficiency-enhancing solutions, which has become the focus of current large-scale model players. **In the future, large-scale models have already moved from "do or not to do" to "how to do it". And that's the next step in the megamodel wars.

03 Participate in the battle for the future, answer these four questions first

Although large models are very popular, there is still a long way to go from the beginning to the market. In the process, many difficulties have been exposed.

However, in the view of 1DU Finance and Economics, the future competition for the largest model will probably be launched in four latitudes. That is: technology, talent, capital and commercialization. **

**First look at the technical level. ** There is no doubt that artificial intelligence is one of the most advanced technologies at present. At the technical level, it is impossible to make up the accumulation it needs in a short period of time. "Big" computing power, "big" data, and "big" models are the basic characteristics of large models at present, and they are also challenges for the industrialization of large models. At present, although the scale of data is large, the quality of data is uneven . Secondly, the size of the model is large, and the training difficulty is higher. The third is that the scale of computing power is large, and the requirements for hardware performance will be higher.

This also means that ** does not have enough funds to support it, so it is difficult to form such a super strong team. **A marketing cloud founder mentioned in communication with 1DU Finance and Economics: "Since investing in the industry's large-scale model in March, the overall capital investment has been very large, even exceeding the sum of the company's establishment to the large-scale model." However, , he also mentioned that if it is done, it will definitely be a reassurance for the company's development in the next ten years.

Prior to this, many people in the industry have proposed that "big models are a game for big manufacturers to burn money." This statement is not without reason.

Although large models are very popular, capital has not kept up with the pace of technological recovery on a global scale. Global venture capital funding nearly halved in the first six months of this year, falling 48% to $173.9 billion, while the number of deals also fell 19%, according to research firm PitchBook.

In China, as of the end of June this year, more than a dozen large-scale model start-up companies have obtained financing. Among the companies that have announced the amount of financing, MiniMax has the largest financing scale. In June this year, it received more than US$250 million in Series A financing from Tencent; Years ago, before being acquired by Meituan, it also received an angel+ round of financing of US$230 million.

Let’s look at the investment of major manufacturers. Previously, Titanium Media’s statistics can explain the problem. In 2022, Huawei’s investment in R&D expenses will be 161.5 billion yuan, becoming the company with the largest R&D investment in China; followed by Tencent, although it is not low. However, it remained at the level of 61.4 billion yuan. Ali ranked third, with R&D expenses of 55.5 billion yuan. According to public information, Baidu, as an early player in artificial intelligence, has invested more than 100 billion yuan in the field of AI in the past ten years. Such investment standards are obviously not comparable to ordinary enterprises.

With technology and capital, big factories are relatively more attractive to talents. At the beginning of this year, each company started a crazy talent competition. Baidu recruits AI large-scale model algorithm engineers with a monthly salary of 25-40K, and the more generous Ali recruits large-scale model training and algorithm engineers with a monthly salary of 40-70K.

After searching for the keyword "big model" on a recruitment platform, you will find that some companies are willing to give 15-25K monthly salary to the 2023 graduates. At the same time, some vertical track companies also participated in this round of competition. For example, a trading company recruited a medical large-scale model product manager with a salary range of 25-50K, and a game company recruited an algorithm engineer for a language large-scale model, and also gave a salary of up to 50K. Even the annual salary of a large model platform product manager recruited by China Telecom can reach 840,000.

Talents, technology, and capital, which are rising with the tide, all urge the players of large models to land and commercialize as soon as possible. After all, according to the laws of business, in the end, these inputs need to be returned in order to be valuable.

However, the landing cost of large models is also a hurdle that major players need to cross. Some people in the industry once estimated that the cost of training a large model is extremely high, reaching 2-120 million US dollars. This also means that the commercialization of AI large models may have to go back to cost accounting.

Conclusion

Looking at the big model from the present moment, the overall situation is very similar to the Internet in 1998. It was just in its infancy, with a lot of bubbles and great opportunities. In this case, a good company with real strength will have better growth and greater value in the future. **