The difficulty of starting a business with a large model: it is difficult for a small model to make money, and the large model is too expensive, and the amount of investment and financing shrank by 70% last year

Source of this article: Times Finance Author: Xie Silin

Image courtesy of Pixabay

The craze for large models has been surging for more than a year, and looking back at this juncture, it has not led to the influx of countless hot money as people expected. For entrepreneurs who focus on this field, survival is still their first concern.

"Now (large-scale) entrepreneurship is very specific, either to help the company make money within a year, or to help customers increase efficiency within three months, in essence, technology has to serve these two purposes. Fan Ling, founder and CEO of Tezan, said.

As a professor and director of the Design AI Lab at Tongji University, Fan Ling, a technology unicorn company focusing on AIGC, hopes to help enterprises reduce the cost of content production, management, and distribution, and improve efficiency through powerful generative AI technology.

Compared to the dark side of the moon, the stepping stars, and MiniMax, which believe in AGI (general artificial intelligence), pursue AGI, and try to change the world, Fan Ling's vision is not a particularly imaginative story, but it is more practical and can make money faster.

According to a previous report by Tencent Technology, at present, the attitude of China's technology industry towards large models has split into two camps. One group believes in technology, believing that if they do not pursue "bigger and stronger AI capabilities", once other people's model capabilities take a leap, they will quickly reduce the dimensionality and crush the existing base areas and moats, while the other group believes that they can use the huge and unique data of the Chinese market to build barriers by investing "enough AI capabilities" into business scenarios that can be quickly monetized.

The former is certainly exciting, but the latter may be the reality faced by more Chinese large-scale model entrepreneurs.

Small models are difficult to make money, and large models are too expensive

Recently, focusing on the theme of "Entrepreneurial Ecology in the Era of Large Models", Tencent Research Institute, Qianhai Institute of International Affairs and Qingteng jointly launched a high-end seminar on AI &Society artificial intelligence + social development.

"The big model can't afford to invest, and the small model can't see the ability to make money. At the seminar, Wu Shichun, founding partner of Meihua Venture Capital, used such a sentence to describe the current investment problems in the field of large models.

In Wu Shichun's view, there are three main investment opportunities in the field of AI, all of which are related to rapid landing: one is to invest in application scenarios, such as embedding AI into smart watches, the second is to invest in AI supporting services, and the third is to invest in enterprises with hematopoietic capabilities.

In fact, this is the result of a combination of factors.

On the one hand, the "expensive" of large models is well known, and the training cost of tens of millions of yuan is destined to be a game for a few people. Coupled with the end of the golden age of US dollar funds, venture capital has become more cautious. Perhaps it is precisely because of this that it is difficult for the large model boom to ignite the primary market.

According to the "2023 State of Artificial Intelligence (AI) Industry Report" released by CB Insights, a research institution, in 2023, the number of investment and financing in China's AI field will be about 232, a year-on-year decrease of 38%, and the total amount of financing will be about US$2 billion, a year-on-year decrease of 70%. Especially in the first quarter of 2023, both the amount of financing and the amount of financing have hit the lowest level in five years, and the financing fever of China's AI industry has "cooled" significantly.

When it becomes more difficult for entrepreneurs to obtain financing to cover the costs of long-term research, they must quickly obtain commercial results and complete the commercial closed loop to ensure the sustainable development of the project – which has become the biggest difference between China and Silicon Valley's large-scale model entrepreneurial ecosystem.

Yao Xing led the creation of Tencent's first artificial intelligence laboratory, AI Lab, and later left to start a general large-scale model company, Yuanxiang XVERSE. He shared the insights he gained when he visited Silicon Valley at the beginning of the year: "It is generally believed in China that to make a large model, you must first think clearly about how to make money, and only if you have money can you survive. In contrast, Silicon Valley's large-scale model practitioners generally believe that the most important thing is to do general-purpose large-scale models, and they just want to 'run to the moon'. ”

However, in Yao Xing's view, there is no superiority or disadvantage between the two development directions, only cultural differences, and it is equally important to do a good job in the application of large models.

Zheng Yongnian, a professor at Chinese University of Hong Kong (Shenzhen) and dean of the Qianhai Institute of International Affairs, also believes that the development direction of China's large-scale model startups is different from that of OpenA, and it should not be a gap, but a differentiation.

In an interview with Times Finance, he pointed out that there are differences in the development model of artificial intelligence in different countries (regions), and China has stood in the world's leading position in the supervision of large models, and the next step should be to encourage development. "If we can build brakes, we can also build faster cars, so we can do both development and regulatory safety. ”

As for how to encourage innovation and development, Zheng Yongnian believes that the flow of information between government departments, schools and enterprises should be strengthened, and the allocation of resources should be readjusted to put more resources on young scholars and entrepreneurs, and give them room for trial and error.

ToB is still the mainstream of commercialization

However, entrepreneurs are still exploring how to apply the powerful capabilities of large models and realize business closed-loop from them. Some of the once-promising directions seem to prove difficult.

For the highly anticipated AI native application in the past, Robin Li, founder and CEO of Baidu, has publicly stated many times that it is meaningless to roll up large models, and only AI native applications are valuable.

However, a year after the large-scale model craze, Wu Shichun found that only 30% to 40% of the projects on the market are completely AI-based, and most of the projects are still existing product upgrades in the past.

"On the one hand, AI native applications have to bear the cost of the education market, and the current users are either based on curiosity or self-consumption in the industry, which is difficult to sustain. On the other hand, the capabilities of the underlying large models are still evolving and upgrading, and native AI applications are easily replaced by general large models after the upgrade. Wu Shichun analyzed.

Similar stories have already happened. After OpenAI released Sora, a video generation model that amazed the world, a number of Wensheng video startups such as Pika and Runway have been seriously challenged. And Jasper, a star startup that relied on its copywriting generation ability to become one of the unicorns, also fell rapidly after the emergence of ChatGPT.

"ChatGPT destroys you, it's none of your business. In this case, it is difficult for entrepreneurs to form an independent logic and complete a large closed-loop experience service for users. Wu Shichun said.

In contrast, enterprise services are still the most mainstream commercialization path for large models. Compared with C-end users who lack the willingness to pay, B-end customers have a clearer desire for advanced technology.

IDC conducted an AI application survey in the fourth quarter of 2023, and the results showed that among the 100 surveyed companies, only 7% of them have no plan for generative AI, which also means that more than ninety percent of the surveyed companies have deployed AI applications. Twenty-four percent of companies have already invested in generative AI and have a clear budget, another 34 percent have already started working on potential use cases, and 35 percent have started experimenting with pilots without a clear budget.

The top three apps that are getting the most attention and want to derive value from them are: intelligent customer service apps, apps that support financial and operational decision-making, and apps that focus on improving employee productivity. In the actual exploration, it is expected that the first scenarios to be implemented are digital marketing, intelligent customer service, and applications to support financial and operational decision-making.

In Fan Ling's view, the key to breaking the situation of large-scale entrepreneurship is to clearly know what enterprises need. "What enterprises need is growth, short-term cost reduction and long-term efficiency increase. They can also accept using AI to make some scene innovations, but if it is just pure innovation, in fact, the vitality is very short. ”

In addition, AI supporting services such as computing power supply are also another major entrepreneurial direction. For example, the computing power service company Wuwen Xinqiong is committed to creating a more flexible and adaptable middle layer between large models and chips, and using GPU inference and acceleration technology to reduce deployment costs.

In the view of Dai Guohao, co-founder and chief scientist of Wuwen Core Dome, 2024 is likely to become a turning point in the development of large-scale model technology. Dai Guohao found that on models of a specific scale such as 3 billion and 7 billion parameters, the difference between open source and closed-source models is getting smaller, but more and more people are applying it. This means that over time, the model's capabilities will gradually reach its peak.

"The more important thing is to do the landing, how to make the cost low enough when deploying. Dai Guohao said.

After all, price and application scenarios are still the main reasons why it is difficult to implement large models on a large scale. It can be seen that compared with the distant dream of AGI, entrepreneurs in the domestic large-scale model track at this stage must first consider ROI (input-output ratio).