Large-scale models have been soaring for half a year, and there are only a few investors who have paid real money

Author｜Liu Yangnan

Source｜Jiazi Light Year

On the 136th day after shouting "to be a Chinese version of OpenAI", Wang Huiwen sold Light Years Away to Meituan.

On June 29, 2023, Meituan announced on the Hong Kong Stock Exchange that it will wholly acquire 100% of the equity of "Light Years Beyond" on June 29, 2023. The total purchase price includes: US$233.673 million in cash + commitment of US$366.924 million Renminbi debt + 1 yuan, totaling about 2.065 billion yuan.

It is hard to imagine that the first large-scale model start-up company founded by a big boss in China will end in this way. This ending has left countless doubts and speculations in the market.

From an overall business perspective, the acquisition of Meituan from light years away is only a common acquisition among enterprises. But for the domestic AI industry, this acquisition seems to indicate that the wave of AI large-scale entrepreneurship that has only been hot for half a year is cooling down.

The capital market has a more intuitive perception of the industry's water temperature. Since June 26, AI concept stocks such as HKUST Xunfei, Kunlun Wanwei, and 360 have collectively plummeted.

Not only in China, people even no longer pursue ChatGPT, which was once flocked to.

According to the data from the website data analysis tool SimilarWeb, the growth rate of ChatGPT’s visits in the early stage was astonishing. The month-on-month growth rate was 131.6% in January, 62.5% in February, and 55.8% in March. It slowed down significantly in April, and the month-on-month growth rate was 12.6%, and in May, this figure was only 2.8%.

When the popularity of large models is gone, it is easy for people to think of a question: Is the large model a good opportunity to start a business?

Obviously, there is no standard answer to this question, and even the opinions of the big guys are quite different. Just a few days ago, Fu Sheng, chairman and CEO of Cheetah Mobile, and Zhu Xiaohu, managing director of GSR Venture Capital Fund, "argued" in the circle of friends about this.

The cognitive inconsistencies of the large model do not focus solely on the individual. When entrepreneurs, investors, and demanders all have cognitive biases, being "cautious and conservative" with large models has become a common state for most players.

At the 2023 Annual Conference of China Film Capital held on June 16, there was a heated discussion between the investors of China Film Capital and the invested hard technology and consumer technology companies on how the industry embraces large-scale models and AIGC .

From the perspective of "Jiazi Guangnian", the large-scale model market urgently needs rational voices, the pace of innovation cannot be stopped, and many issues have yet to be clarified-whether large-scale models can be cast? How to cast? What are the problems with entering the big model? Will the future commercialization of large models repeat the mistakes of the previous AI industry?

No matter how worried people are, it is almost a given that the industry embraces the big model - the question is, in what way.

1. Prudent investors

Baichuan Intelligent, Lianyuan Technology, Lingxin Intelligent, Xihu Xinchen, MiniMax... From the beginning of 2023 to the present, domestic large-scale start-up companies have emerged one after another. The background of each entrepreneur is bright enough, and the strength of each enterprise is highly capitalized. side approved.

During that time, it was not uncommon for a certain star entrepreneur to leave his job to start a business and enter a large model to obtain huge financing.

For example, on June 1, it was reported that MiniMax, a domestic large-scale model start-up company, was about to complete a round of financing of more than 250 million US dollars, and the company’s valuation exceeded 1.2 billion US dollars;

At the time, the financing news of Light Years Away, which had just been acquired by Meituan, was even more confusing. Wang Huiwen once denied that the company had received US$230 million in financing from Source Code, Tencent, Wuyuan, and Suhua, but this financing was eventually rejected. It was confirmed because of the acquisition announcement issued by Meituan.

This wave of investors pursuing entrepreneurs made people think that the big model would activate the entire domestic AI venture capital circle, but this is not the case. In fact, except for star teams with their own halos, investors are more likely to wait and see and examine large-scale entrepreneurship, and only a small number of people actually spend real money.

As early as the beginning of this year, when the ChatGPT wave swept the world, "Jiazi Guangnian" exchanged views with some investors for the first time. During that time, almost everyone was learning quickly and invited experts to conduct popular science within the company.

At that time, there was a hot discussion in the AI field: Is ChatGPT the iPhone moment in the artificial intelligence world? In this regard, the answer given by Xianfeng is not to rush to a conclusion. "We are not yet sure how big this impact is, but we think it (big model) will definitely change something." Li Kang, vice president of Xianfeng, said.

However, some investment institutions have expressed concerns about the large model. A primary market investor told "Jiazi Guangnian" that he was very worried about China's overreaction. After the outbreak of ChatGPT, domestic AI concept stocks were detonated. "Our primary and secondary markets must consider whether the relevant investment injected can generate corresponding returns. If it is for short-term interests, this kind of investment will easily go to waste in the end, because you did not really promote the development of technology, but It's a conceptual investment."

In his view, investors should focus on exploring more basic sciences that have an impact on the future of mankind. This is the real technological power with hidden potential market value. "It is necessary to integrate with market dynamics, as well as market value and real, future social progress. We must not follow blindly. We must clearly understand what can change the future. It shouldn't be a bubble in a wave."

However, an FA practitioner told "Jiazi Guangnian": "Recently, investors have gradually started to invest in large-scale model projects, but the amount is not large."

"The essential problem is still insufficient cognition." Regarding investors' cautious attitude, Zhang Gaonan, managing partner of Huaying Capital, gave his own understanding. He further stated: "Almost no one can clearly define the large model. We need to define the boundaries of the large model before discussing it. The large model you mentioned and the large model I mentioned are probably not the same thing."

In some people's view, investors' caution may be a negative signal for large-scale entrepreneurship, and it is pouring cold water on large-scale models. But from an objective point of view, prudence does not mean rejection, and a rational embrace after deep thinking is more precious.

Whether investors, entrepreneurs, or traditional companies that hope to transform and upgrade their own business with the help of large models, they need to clarify two issues before they really enter the large model market—what is the capability boundary of the large model, and whether they have Need to introduce a large model?

2. Before embracing the big model, clarify two questions

When a new technology emerges, the core question in the business world is: where and how can this technology be used?

This is especially important for large models, and it is also a question that companies that have not really entered the large model should consider carefully.

To answer this question, it is first necessary to delineate the capability boundary of the large model.

The special feature of the large model is that its internal model algorithm is a huge black box, and the generation process of the large model is unexplainable and unpredictable, so it is difficult to define its capability boundary. But what is certain is that the general-purpose large model is not a panacea.

Lin Yonghua, vice president and chief engineer of Zhiyuan Research Institute, once mentioned in a sharing that from the perspective of industrial implementation, "big model + prompt learning" cannot replace everything.

She further mentioned that for many specific tasks or new tasks, hint learning may allow the large model to output the required results through multiple rounds of hints, but the large model "cannot remember" this process, and if the developer adds the entire hint In each call, on the one hand, it may make it longer and longer and exceed the context capability of the large model. On the other hand, it will inevitably lead to an increase in the cost of each reasoning, and the effect will be difficult to control. This instability is even more fatal in the landing stage of products that have invested real money.

Zhang Yitian, chief expert of the National Speech Innovation Center, also said at the annual meeting of Huaying Capital in 2023: "The big model is a generative logic, and what it gives is only an optimal vocabulary clustering, and there is no cause and effect between the answer and the question. What we get is just a 'result', which needs to be identified, not an 'answer'. Therefore, whether the large model can be directly applied in serious fields such as education, medical care, and justice may be a problem. But it is in assisting decision-making. It is meaningful. In terms of directly generating results, if it is to be commercialized and productized, we think there is still a long way to go.”

Therefore, there is a consensus in the industry that in the future each industry will have its own vertical model, and the key point is how to integrate the capabilities of the general model with the company's own industry data.

But before really considering the implementation of the large-scale model project, entrepreneurs need to consider a more important but easily overlooked question-is the large-scale model a "just need" for the enterprise?

In this regard, a domestic multi-modal large-scale model team once told "Jiazi Guangnian" that whether large-scale models are "just needed" by enterprises needs to be understood from multiple perspectives. For some enterprises, not introducing a large-scale model is equivalent to losing a sharp edge in market competition, and they have to use a large-scale model to win the favor of customers-this is also a "just need".

But to some extent, this is more of a market sentiment in the early days of emerging technologies. In the long run, the industrialization of emerging technologies will essentially be driven by business needs. At this time, whether an enterprise needs a large model needs to consider multiple factors.

In addition to specific project implementation issues, companies also need to consider data security issues and the impact of large models on the original market structure.

These two problems have frequently appeared in the information and digital age and cannot be eradicated. In the intelligent age, these problems may appear in a more subtle way.

"Many consumer or platform-based companies, if they embrace the big model without reservation, the big model has a strong ability to backlash the industry, because it means that the industry is easily handed over from the industry. The entry threshold and cognitive key." Zhang Yitian said.

In the context of digital transformation, most industries have achieved digital transformation and upgrading through technologies such as big data and cloud computing. But at the same time, traditional enterprises have also handed over a large amount of industry data to digital technology manufacturers. Digital technology providers have become a main body that cannot be ignored in the industry, and the original market structure has been changed.

However, due to the small amount of data in industries such as industry and construction, and the difficulty of connecting data between business lines, traditional enterprises still maintain high competition barriers.

Zhang Yitian said: "Currently, the construction industry is the industry that is best protected in the context of digitalization. Now the informationization of the construction industry, except for a Glodon who can make a budget, no giant can cut in. Why? Because architecture has design Drawings, construction drawings, maintenance drawings, planning drawings, filing drawings and other eight drawings, all the drawings are not connected to each other, and the government departments do not recognize each other. The cost of opening these eight drawings of the entire building is high enough , so the construction industry has maintained the diversification of this kind of competition. We usually think that the diversification of competition is the source of vitality and power for industrial development.”

Therefore, under the wave of large-scale models, for enterprises whose industry standards and competitive advantages are not perfect, whether to embrace large-scale models unconditionally is a question that every company needs to carefully consider.

3. There is no standard answer for the engineering implementation of large models

For entrants who have invested real money in the large model market, the next thing to do is to solve the actual project implementation problem.

In this regard, the industry has gradually formed a consensus that in the future, large models and small models will complement each other in the process of industrial implementation.

Lin Yonghua once said that narrow-area scenarios that require high precision and low generalization capabilities are more suitable for the "small model + transfer learning" paradigm. Such as industrial inspection, industrial quality inspection, medical image analysis, etc.

In addition, Xuan Xiaohua, the founder of Huayuan Computing, also said that the business model of AI companies in the future is to integrate the general big model driven by big data and the small data model driven by knowledge for vertical industries to achieve two-wheel drive.

Zhang Gaonan also told "Jiazi Guangnian": "When enterprises optimize their own models or train vertical models, they can combine with large models. They don't need high-dimensional data like large models, and they don't need to fully apply large models. However, it is instructive that large model technology can be coupled with other technologies to form an industry vertical model with low computing power requirements, and it is by no means a simple application of large models.”

For example, for the "illusion" problem of large models that has been criticized repeatedly, it may be necessary to combine the previous generation of AI technology to solve it in the short term.

"There are many reasons for hallucinations, and it may be because data is relatively sparse and insufficient in a certain domain. In this case, we need to provide more data to the model for training. Also, when users ask questions, clearly provide More background information is also a way to reduce hallucinations, or to lower the 'temperature'. Sometimes hallucinations occur because the questions are not complete enough, lacking background and premise. Therefore, the question is also very important, and engineering is the key. In addition , if users really want to solve accurate problems 100%, they may still need to use knowledge graphs. Knowledge graphs can ensure the accuracy of logical reasoning, as well as newer technologies including the 'world model' proposed by Yann Lecun, the head of Meta AI ” said Wu Xuening, CTO of Jinyou.com.

In addition to combining with the previous generation of AI technology, it is also an important part to combine the training process of large models with high-quality industry data.

For example, as a hybrid database serving the AI PaaS platform, Tianyun Data has gone through more than ten years, and now it has reached the stage of combining with large models.

Li Congwu, vice president of Tianyun Data, said that the combination of itself and the large model will be considered from two aspects - first, how to combine private domain data with the large model. For example, Tianyun Data has completed a similar policy interpretation project for the China Securities Regulatory Commission. By combining various data such as regulations, precedents and interpretations, Tianyun Data has generated interpretations of violations, similar to the process of court penalties. data to decipher the causes of violations.

Secondly, Tianyun Data has been developing a hybrid database and is one of the earliest companies in China. As early as around 2018, Tianyun Data proposed the concept of AI native database, which is actually similar to the vector database that supports large models today. Tianyun Data has released a self-developed vector database and applied it to its own models.

In general, there is no standard answer to the engineering implementation of large models.

On the road of artificial intelligence, China will definitely blaze a trail that is different from other countries. It is difficult to distinguish the advantages and disadvantages of the two paths, and they are more based on realistic choices under different national conditions.

Li Kang, vice president of Xianfeng, once made an analogy to "Jiazi Guangnian" in an interview, and it still seems to be applicable today: "It is unfair to use the success of OpenAI to describe the many problems of domestic entrepreneurs. It is like two people playing cards. It's different, the other party suddenly played a big hand and got a straight flush, you just say that he played well, I was too cautious, but why didn't you say it when I won?"

Zhang Yitian shared that from the perspective of the central government, large models, including artificial intelligence issues, are an important strategic tool to win the initiative in global competition, and an important strategy to promote the leapfrog development of my country's science and technology, the optimization and upgrading of industrialization, and the overall leap in productivity resource.

"When the report of the 20th National Congress of the Communist Party of China talked about industrial issues, it proposed artificial intelligence separately from the new generation of information technology. Therefore, from a policy point of view, the development of artificial intelligence and large-scale models is not only a technical issue and an industrial issue, but also the core competition of a national economy. The issue of power, from a deeper perspective, is a political issue, and everyone should understand this issue from a higher perspective.”

If we jump out of China and extend the timeline of technological development, the uproar caused by ChatGPT may be just a point in the history of artificial intelligence technology development, and all judgments may be premature.

After all, even the technicians who have been immersed in the front line of artificial intelligence research for a long time have not yet reached a consensus on the future of artificial intelligence, and are deeply in a state of anxiety.

In the recently popular book "Why Greatness Can't Be Planned", the authors Kenneth Stanley and Joel Lehman write: "We have to face the uncomfortable fact that we cannot Determine whether any rules of thumb can be reliable guides in pursuit of AI goals."

At the 2023 Beijing Zhiyuan Conference that ended not long ago, Huang Tiejun, director of Zhiyuan Research Institute, also had very similar anxiety. He directly used the four words "unable to close" as the title of the speech for the closing ceremony. He said: "We are in an uncertain state. Is this Near AGI stronger than us? Is it more intelligent than us? Or when will it surpass us? I don't know. We are in a completely In a state of being out of control."

At the end of the closing speech, he concluded with this sentence: "If we can deal with risks with the same enthusiasm as investing in large models, at least it is possible to grasp the future. But, do you believe that humans can do it? I don't know."

Looking at all technologies and industries, this "sense of loss of control" from front-line core technicians is not common in other fields. Now, almost everyone is crossing the river by feeling the stones. Every company entering the market today has the potential to become a pioneer in the technological virgin land.

And time is the best proof to prove everything.