Produced by |Sanyan Pro Author|doraemon
Today, Zhu Xiaohu, a partner of GSR Ventures, bluntly said in an interview with Tencent News's "Periscope" column that he is not optimistic about the development of domestic large models.
However, to be more precise, Zhu Xiaohu is more optimistic about the prospects of the application of domestic large models, and for the large model itself, Zhu Xiaohu said bluntly: "I don't want to talk about these companies, if you want a scene without a scene, if you want data without data, what is its value?" And the valuation is so expensive. ”
Zhu Xiaohu bluntly said that when they came out to raise funds for large model companies, they were not optimistic about it and would not invest.
Zhu Xiaohu: I believe in application
Zhu Xiaohu pointed out that investing in domestic large-scale model companies may not make money at all. To do GPT-4, it is necessary to invest in scientific research, but at least 40 or 50 million US dollars; And if you want to make GPT-5, it will cost hundreds of millions of dollars.
He believes that what makes large model companies even more "embarrassing" is that even if some companies are willing to spend money, after investing tens of millions of dollars, they encounter others open source, wouldn't it be in vain? Therefore, most companies in China do not dare to spend money on the research and development of GPT-4.
Startups, in particular, can't afford to take this risk.
In addition, in 2023, several start-up companies will be "rolled" out of the domestic large model, but Zhu Xiaohu said bluntly, "Let's see if we look at it in another year, how many of these can still be there?" ”
However, Zhu Xiaohu emphasized that although he does not believe in large models, he believes in applications and believes that they can be commercialized immediately.
Zhu Xiaohu mentioned some cases of AIGC, such as doing AI interviews, using AI to replace people in WeChat private domain marketing, and AIGC advertising, which are all excellent cases at the application level.
What is the significance of more than 200 large models in China? It doesn't make much sense. But there's a lot of innovation at the application layer. In terms of domestic re-data and application scenarios, it far surpasses that of the United States. ”
At the same time, Zhu Xiaohu also pointed out that at present, the commercial application of domestic AIGC is mainly to B, and he believes that it will definitely be done in the short term. Like iPhone, Big Brother, and when the computer first came out, it was all used first. Businesses are willing to spend money when they can improve productivity immediately.
And to C has to wait for the "iPhone 3 moment", like the original mobile Internet to C application outbreak is multi-touch (multi-touch), like the original Angry Birds, watermelon cutting and other popular applications are actually based on multi-touch technology after the launch of the popular.
But when the large model will be able to make the to C explode, Zhu Xiaohu said that he does not know. As for the current large model applied to the so-called "personal assistant", Zhu Xiaohu said that it is a pseudo demand, "Personal assistants are the imagination of technical personnel, how many people need personal assistants?" These are all typical pseudo-needs. ”
How is the commercialization of large models?
To B is the mainstay, and there is rarely a to C mode
Zhu Xiaohu's point of view is simply summarized that he is not optimistic about domestic large-scale model companies, but the development of the application layer will be better. At present, the commercialization of large models is mainly on to B, and to C still needs to wait for the explosion point.
So, what is the commercialization of domestic large models? The author briefly conducted a survey to understand.
Wen Xin said
Let's take a look at Wenxin's words first.
According to the financial report, in the fourth quarter of 2023, the total revenue of intelligent cloud will be 8.4 billion yuan, of which the large model will bring about 6600 million yuan of incremental income.
By the end of December, the number of users of Wenxin Yiyan exceeded 100 million, with a total of 3.7 billion words of text creation, 300 million lines**, helping users process professional contracts with a total of 400 million words and making 5 million travel plans.
Robin Li**, incremental revenue from generative AI and underlying models will increase to billions by 2024, mainly including advertising business and AI cloud service sales. In addition, the daily call volume of the Wenxin model has exceeded 50 million, with a quarter-on-quarter increase of 190%; There are about 260,000 enterprises called the Wenxin model, an increase of 150% quarter-on-quarter. Based on the estimated number of 50 million calls, the number of calls to the Wenxin model by enterprises exceeds 3 per 1 minute40,000 times.
In terms of the commercialization attempt on the C-side, Wenxin Yiyan will launch a paid version in November 2023, which is called "Wenxin Yiyan Professional Edition" and is priced at 599 yuan month.
Compared with the ordinary version, Wenxin Yiyan Professional Edition can provide users with a more advanced Wenxin large model 40 services, while the free version can only use Wenxin model 35。
iFLYTEK Xinghuo cognitive model
At the end of January 2024, iFLYTEK released the Xinghuo cognitive model v35。
At the press conference, iFLYTEK announced that Xinghuo has created a benchmark for large-scale model empowerment with leading enterprises in the fields of insurance, banking, energy, automobile, and communications. Like Pacific Insurance, Bank of Communications, Chery Automobile, China Mobile, etc. are all partners of Xinghuo.
Alibaba Tongyi Qianwen
On October 31, 2023, Alibaba released the large model Tongyi Qianwen 20, at the same time, the Tongyi Qianwen app was also launched on the same day. In addition, Zhou Jingren, CTO of Alibaba Cloud, announced that the Tongyi Qianwen 72B model will soon be open sourced.
In terms of commercialization, in Tongyi Qianwen 1When 0 was released, Ali said that many of its applications would be connected to the large model. Subsequently, Alibaba's own products and business scenarios such as DingTalk, Tmall Genie, Ali E-commerce and Ali Entertainment were successively connected to Tongyi Qianwen.
Huawei Pangu large model
Huawei's Pangu model was initially developed for B-end users.
According to reports, the Pangu large model 3Since the release of 0, Huawei has implemented in industries such as government affairs, mining, meteorology, automotive, medicine, digital human, and computing R&D. For example, HUAWEI CLOUD has signed a cooperation agreement with Payroll and Salary, Shengyu Technology, Hengfang Health, iDesign, and Wandianzhang.com, and announced that it will jointly innovate based on the Ascend AI cloud service and the Pangu model.
At present, in which areas is Huawei's internal positive model can be implemented, and it can also obtain a positive business closed loop.
SenseTime consults sensechat
At the same time, SenseTime has also cooperated with hundreds of customers in multiple vertical industries such as energy, finance, real estate, human resources, media, industrial manufacturing, consulting, etc., to customize polishing products, and the revenue of a single customer has increased by nearly 60%.
A general model for scholars
The general model developed by the Shanghai Artificial Intelligence Laboratory has been authorized to be used by enterprises in communications, finance, high-end manufacturing and other industries, such as cooperating with China Telecom to develop an intelligent customer service system; Jointly released the "CCTV Listening Model" with the first radio and television station to help the best practitioners complete their work.
Baichuan Intelligent Baichuan large model
On December 19, 2023, Baichuan Intelligent announced the opening of the Baichuan2-Turbo series APIs based on search enhancement, including Baichuan2-Turbo-192K and Baichuan2-Turbo. In addition to supporting 192k extra-long context windows, the ability to search for an enhanced knowledge base is also added. Starting today, API (Application Programming Interface) users can upload text materials to create their own knowledge base, so as to create more complete and efficient intelligent solutions based on their business needs.
Wang Xiaochuan, founder of Baichuan Intelligence, said that it is not enough to rely on the model itself to make it bigger, and the large model + search can constitute a complete technology stack.
In terms of commercialization, in February this year, Datapotential Technology, a provider of data intelligence products, announced that it had reached a cooperation with Baichuan Intelligence to jointly release application solutions for the vertical field of large model data analysis.
On the official website of Baichuan Intelligence, you can view the charging standard of Baichuan large model, which is also one of the current commercialization attempts of Baichuan.
In addition, Baichuan Intelligence is currently recruiting many commercialization-related positions, including business managers, industry directors corresponding to customer development in industries such as finance, education, and operators.
Yunzhisheng mountain and sea model
The large model of mountains and seas developed by Yunzhisheng has achieved certain landing results. In the field of smart medicine, the outpatient medical record generation system based on the Shanhai model has been applied to Beijing Friendship Hospital; In terms of government services, Yunzhisheng and Shenzhen Longhua Data have jointly developed a large government affairs model; In addition, the Shanhai model is also stationed in Nanning Railway Station to provide more humanized customer service.
Du Xiaoman Xuanyuan
The Xuanyuan large model developed by Du Xiaoman was selected as the "Advanced Application Case of China's AI Large Model" released by MIT Technology Review. Du Xiaoman said that since the open source, hundreds of financial institutions have applied to use the Xuanyuan model.
Kimi, the dark side of the moon
Here is a special mention of the dark side of the moon, although the company has not yet had a specific commercial project, but the founder of the dark side of the moon, Yang Zhilin, has been affirmed by Zhu Xiaohu.
Zhu Xiaohu said that he invested in Yang Zhilin's previous company, and praised that "others are very powerful, and he can make a large model", however, Zhu Xiaohu said that he did not know how Yang Zhilin was commercialized.
Different from most large models, the Kimi intelligent assistant on the dark side of the moon is currently mainly oriented to the C-end and can support 200,000 words of long text input, which further improves the user experience of the C-end user.
However, at present, the Kimi intelligent assistant has not yet been charged, and it may take the commercial route of charging in the future.
In fact, the author looked at nearly 30 domestic large-scale model products, and the specific commercial landing projects can barely pick out the above ten. Among these ten large models, almost all of the commercialization directions are to B, and only Wenxin Yiyan has launched a paid model on the To C side.
In addition, some large models have also issued relevant charging standards, such as Zhipu AI's standard version of large model pricing and cloud privatization pricing.
On the whole, in terms of the commercialization of domestic large-scale model applications, it is indeed as Zhu Xiaohu said, and most of them are currently mainly to b. At the same time, although the current commercialization direction is not much and is not yet clear, it has great potential. As for the current so-called to B cooperation, which is "pure cooperation without fees" and which is paid cooperation, only they know.
For example, 360 revealed that the 360 intelligent brain developed by the company has been able to generate revenue; Meitu said that the large model drove the company's profit growth by 3 in the first half of the year2 times. Alihe's big model is to empower its own business system to help its businesses generate revenue or improve user experience.
Has large-scale model financing slowed down?
In the past 5 months, there has been an average of 1 financing per month
Let's pay attention to the recent financing situation of large model companies.
The dark side of the moon
According to reports, at the beginning of 2024, the dark side of the moon will complete a new round of financing of more than $1 billion. Investors include Sequoia China, Xiaohongshu, Alibaba, Meituan and existing shareholders to follow-up. After this round of financing, the valuation of the dark side of the moon reached $2.5 billion. This financing is the largest single round of financing obtained by a domestic AI model company.
The dark side of the moon has been established for less than a year, and it completed an angel round of financing on June 12, 2023, with investors such as Sequoia China and Zhen**, with a valuation of $300 million at that time.
minimax
According to the "Science and Technology Innovation Board**", recently, people familiar with the matter said that minimax, a general large-scale model startup, is in the process of raising a new scale, and Alibaba is the core lead investor.
Minimax completed an angel round of financing on January 1, 2022; On July 26, 2022, the Pre-A round of financing was completed, with a valuation of US$500 million, and the investors are Mingshi Capital, miHoYo, Hillhouse and IDG Capital; On June 1, 2023, Minimax completed a Series A financing with a transaction amount of more than 2US$500 million, valued at US$1.2 billion, with Tencent as the investor.
Baichuan Intelligence
On October 17, 2023, Baichuan Intelligent completed Series A financing, with a transaction amount of US$300 million, with investors including Alibaba, Tencent, Xiaomi, Lenovo, Cornerstone Capital, Shenzhen Innovation, Red Dot China, Zhuoyuan Asia, Shunwei Capital, Heart Capital, etc.
linker
According to reports, in February this year, the large model company Linker announced the completion of a new round of strategic financing of hundreds of millions of yuan, with investors China Mobile Capital and Qianhai Mu**.
In addition, the company has now completed four rounds of financing, but did not disclose the exact transaction amount. In January 2022, Lianhui completed Series C financing.
Zhipu AI
According to Tianyancha, on January 17, 2024, Zhipu AI completed the C round of financing, and the investor is Lightspeed Photosynthesis; On September 19, 2023, the company's last round of financing transactions amounted to 1.2 billion yuan, with a valuation of 1 billion US dollars, and the investors were Tencent and Alibaba, as well as Hillhouse, and some old shareholders followed.
It is worth noting that after a total of 9 rounds of financing of Zhipu AI, the company's valuation has increased from 25800 million yuan skyrocketed all the way to $1 billion.
In the past 5 months from October 2023 to March 2024, there have been 5 financings of large model companies, and on the whole, traditional VCs still account for the majority of these financings, but in terms of investment frequency, Alibaba has frequently made moves, and among the 5 companies, Alibaba has participated in a total of 4 companies.
Of these five companies,The Dark Side of the Moon and Zhipu AI are both valued at more than $1 billion; Minimax and Baichuan Intelligent are also valued at more than 100 million US dollars.
In Zhu Xiaohu's view, large factories participate in investment because of FOMO (fear of missing out), afraid of missing out. Most of the money on the dark side of the moon came from Ali.
However, the author has also noticed that the financing events of large model companies have indeed decreased.
According to the "2023 China New Generation Artificial Intelligence Technology Industry Development Report" released by the China New Generation Artificial Intelligence Development Strategy Research Institute, as of October 2023, the total number of domestic large models has reached 238. From January to October 2023, there were 38 investment and financing projects in the field of large models.
Based on this calculation, the first 10 months of 2023 will have an average of 3 large-scale model investment and financing per month8 up. However, in the last five months, the frequency has dropped to an average of one per month.
Have the big models stopped? It is obvious that the movements of the large models are not as frequent as in the first half of last year, and they are even very quiet, what are they doing, have they stopped? The author checked the social platform accounts of some large model companies and found that in fact, most companies still maintain a certain frequency of stable updates.
However, judging from the dynamic reading volume and page views of these companies, the attention is indeed not so high. On the one hand, the current application is mainly to b, and on the other hand, it may be that the popularity has receded, and non-people in the circle do not pay much attention to it. Zhu Xiaohu's views are basically in line with the status quo
The "big model fever" is in decline
All in all,The current development trend of the domestic large model industry is in line with Zhu Xiaohu's point of view, that is, the financing volume of large models is reduced, the development prospects of the application layer are good, and the current commercialization is mainly to b.
In fact, the so-called heroes see the same thing, and some industry bigwigs have the same views as Zhu Xiaohu.
Robin Li once talked about the large modelThe model itself does not produce value, and the application developed based on the basic large model is the meaning of the existence of the model.
In terms of whether the application used to be to B or to C, Zhou Hongyi, the founder of 360, also believes that the opportunity is mainly in the enterprise market. He stressed several timesEnterprises should make full use of the best and most mature capabilities of large models, and focus on the pain points and rigid needs of enterprises. For example, information analysis and decision-making, enterprise knowledge search and management, office collaboration, and intelligent customer service are the most suitable entry points.
In fact, due to the high cost of computing power for large models, it is difficult for C-end users to cover the full cost. This is also as Zhu Xiaohu said, only when there is a "breaking point" on the C-side, there is the possibility of explosive products. Therefore, the vast majority of commercial applications of large models at this stage are to b.
On the "volume" model, perhaps as Robin Li and Zhu Xiaohu said, the final "volume" will not produce results. At present, from the perspective of financing frequency, the financing situation of large model companies has decreased; On the other hand, because the TOC application experience is relatively simple, the "big model fever" at the beginning of last year seems to be declining.
Are you still using large models?