news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Interface News reporter | Wu Yangyu
Jiemian News Editor | Song Jianan

At the annual meeting of Jiuhe Ventures in June this year, the VC firm did something unusual: in addition to convening people in the venture capital circle to give speeches and dialogues, it also released a report. The theme of the report fits the current technology boom, and is called "Immortal Computing."

It regards "computing" as a way to express and understand the world, and makes the following judgments:

Computing is showing infinite vitality and scalability, and its expansion boundaries are spreading rapidly, until the scope of computing exceeds the life boundaries and life experiences of individuals, until the dimension of computing exceeds the knowledge density and knowledge volume of a single workforce, until computing digitizes everything and immortal computing becomes a reality.

In the view of Jiuhe Venture Capital, AI will become the next generation computing platform, and we are currently in the early stages of platform transformation, which will profoundly affect future investment and entrepreneurial opportunities.

Jiuhe Venture Capital was founded in 2011 by Wang Xiao, one of the "Seven Swordsmen of Baidu". In the nearly 12 years since its establishment, the organization has spent a lot of energy focusing on the hard technology track. After the ChatGPT storm hit, the team has made more frequent and concentrated moves in the field of artificial intelligence. In 2023, nearly 50% of the projects invested by the organization were related to AI.

In the AI ​​1.0 era, China formed a market structure dominated by the "Four Little Dragons of AI" (SenseTime, CloudWalk, Megvii, and Yitu). In the AI ​​2.0 era, which is centered on large language models, larger-scale technical challenges and business prospects have begun to emerge.

Despite the unknowns, no investment institution focusing on the technology field can avoid this trend. Their focus is nothing more than whether to stand on the underlying big model or to invest in the AI ​​application layer.

Jiuhe Ventures has not invested in any general-purpose large-scale model companies, but has prepared relatively sufficient ammunition for the AI ​​application layer. Currently, many AI application companies can meet demand, provide value, and make initial attempts at commercialization. Wang Xiao pointed out that this has shown signs in the fields of headhunting, games, sales, etc.

If we have to talk about whether one believes more in the market or technology, Wang Xiao believes that compared to "when AGI (artificial general intelligence) can be realized", he is more concerned about "when AI applications will explode."

"Technology must be useful. Useless technology is misleading," he said.

However, the overall primary market is cooling down, the entrepreneurial environment is becoming increasingly difficult, and more VC institutions prefer companies that have the ability to generate capital at this stage. What will Jiuhe Venture Capital choose?

Wang Xiao believes that the primary market should not only focus on companies with markets and revenues, but also look at them comprehensively. "If we only invest in companies with revenues and buy them at very cheap prices with a price-to-earnings ratio of several times, China's technology will be doomed." He pointed out that sometimes we need appropriate bubbles, time, and a yearning for the future and imagination.

"If there is no imagination, who would invest in Google when there were only two of us?" he said. "You can only invest if you dare to think, and only if you dare to invest, it will be possible."

The following is a slightly edited transcript of an interview with Wang Xiao conducted by Jiemian News: Building a new framework for understanding investment

Interface News: Why release this report at this point in time?

Wang Xiao:In order to prepare the content for the annual meeting, I happened to encounter the emergence of big models. The good thing about this topic is that human intelligence is limited by the length of life. With the end of life, carbon-based intelligence is dead in a sense. Machine learning is relatively not limited by time. As long as the earth is not destroyed, big models are estimated to exist forever, and its information input bandwidth is not too limited.

We often say that humans are immortal, but here we say "computational immortality", which in a sense gives life to intelligence. We think this is a good word.

Jiemian News: What is the significance of emphasizing the value of computing for the primary market?

Wang Xiao:Our intention is to summarize the current changes in "computing" from a more macro perspective, so that we can have a better understanding framework when investing.

Interface News: What is the most important element among these?

Wang Xiao:It is about calculating in which directions greater value can be brought in the entire process of human life evolution.

Artificial intelligence is not just an ordinary technology, but also has long-term significance and disruptive capabilities. Its impact is a bit like the Internet. The Internet connects billions of people around the world, while artificial intelligence provides the vast majority of people with intelligent capabilities, such as driving with the help of autonomous driving technology, children being educated by robots, and part of the work being completed by robots.

Pointing this out helps us strengthen our resolve to invest, because investing is essentially about believing that something has different meanings.

Jiemian News: Jiuhe has been paying attention to the structural changes brought about by "computing evolution". What are the important nodes you observed before the emergence of large language models? What entrepreneurial opportunities have they derived from each?

Wang Xiao:From the perspective of artificial intelligence, the first wave of AI began to rise with AlphaGo as a landmark event. At this time, image processing technology was relatively mature and applied to fields such as autonomous driving and medical imaging. In this wave, we invested in Momenta and Eagle Eye Technology.

The first wave of AI had a relatively small impact, and not many real AI companies emerged. The larger ones included autonomous driving companies and some image detection companies.

Since GPT became popular at the beginning of last year, China has shifted to large models. This second wave of artificial intelligence has better capabilities than the previous one, and the application areas and scope are much larger than the first wave. However, we still need to observe for another three years to see how far it will be implemented and what the applications will look like.

Interface News: What is the reason why this wave of entrepreneurial opportunities is greater than the previous one?

Wang Xiao:The essence of this wave of large models is that the underlying architecture of machine learning is done with Transformer, that is, these contents are tokenized before learning, and the amount of data that can be absorbed is very large, which in a sense generates intelligence. The previous wave is simple image recognition, which can solve problems closely related to images.

Today’s big models are both multimodal and linguistic, and language carries a certain cognitive state of people. Big models have a certain degree of intelligence, and intelligence can be generalized to various fields.

For example, the big model can assist junior headhunters in recording information such as candidates' career plans, and improve matching efficiency when suitable positions are available. This is very limited if it is recorded by humans, but the big model can triple or quadruple the efficiency of headhunters. Similarly, in the fields of programming, art design, and advertising planning, those less advanced skills can have a certain generation capacity.

Jiemian News: Is it true that the threshold for this generation of applications is lower than that of the previous generation? The cost of trust for users seems to have been reduced a lot.

Wang Xiao:If we talk about applications such as autonomous driving, this is for sure. If autonomous driving makes a mistake, it will be a big problem, so the threshold is very high. The experience of this generation of applications is not that good yet, or the accuracy is not that high, but it does not have much impact. We are willing to give AI time to grow.

Without investing in large models, AI applications do not necessarily require AGI

Jiemian News: Do you think it is necessary to achieve the so-called AGI in order to realize the value of the AI ​​application companies you invest in?

Wang Xiao:It is not necessary. Now many scenarios have been implemented in large models, such as customer service, game chat, and robots. They essentially improve labor efficiency or replace people. It’s just that these scenarios and cases have not yet reached the scale of capabilities, and the income level is not high, but these scenarios are established in this experimental process.

Interface News: What are the restrictions now?

Wang Xiao:Each field is a little different. I think it may still be a problem of effectiveness. It can't really do anything particularly intelligent, but it can do basic tasks such as chatting and summarizing articles very well.

Interface News: This effect is directly linked to the intelligence level of the model.

Wang Xiao:There are two parts to the effect. One is the intelligence of the basic model, and the other is to build a set of your own data system based on the basic model. For example, if you want to chat with people in games, you need to train a set of game-related content; if you want to be a headhunter, you need to understand the so-called professional terms. In short, you need to have a unique data closed loop in your own field, and put it together with the basic model to form better services.

Jiemian News: What kind of models do the AI ​​application companies you invest in use? What is the relationship between the optimization of the data system and the improvement of the intelligence level of the base large model?

Wang Xiao:Some use open source big models, some use cloud services similar to big models, and some modify the open source big models directly. Big models are not expensive now, and the price has been lowered in China, so I think China's application may grow rapidly, even faster than the United States, because the means of production are indeed cheaper.

In terms of data systems, they use large models and their own data systems to form services. After having users, they can generate data, which in turn trains their own models. This data closed loop will get better and better. Therefore, AI application companies are valuable and will eventually form a certain data threshold.

Interface News: However, using other people’s big models is still limited at the underlying level by the development of the big model’s intelligence level.

Wang Xiao:This is certain, but there are also continuous breakthroughs. Now the open source large model LIama 3.1 has accumulated hundreds of billions of parameters.

Jiemian News: As an investor, do you accept this kind of restriction? The development of an industry determines the future of this company.

Wang Xiao:Accept it, this is the ecosystem. Generally, AI application companies cannot make large models themselves because it is very expensive. The best way is to use several at a time, so that they can be disassembled at any time in case of any emergency.

Jiemian News: Are you worried that, for example, the performance of a closed-source big model will be better than the best open-source big model? If those companies make the same application, they may quickly surpass the company you invested in?

Wang Xiao:I am not too worried because big models are basic capabilities. There may be differences when using different models, but they are not that big. AI application companies have built a closed loop of data and services in this scenario and industry. This service has its uniqueness, which improves its capabilities. Compared with the data accumulated in the industry and the overall user solutions, the changes in big model capabilities have less impact on the final competitiveness.

Interface News: So when you look for companies in the same application track, you will pay more attention to their understanding of the uniqueness of their own scenarios and how they provide this service?

Wang Xiao:Yes. For example, we invested in a company called Xingzhe AI, which does AI generation related to the gaming industry, including image generation and music generation. This company was born out of a gaming company. The two founders are very familiar with the gaming industry, and most of their customers are also from the gaming field. This company uses big models to solve various problems for the gaming industry. They know where and how to use the technology.

Interface News: Compared with large model companies, will the way of judging AI application companies be clearer?

Wang Xiao:It is relatively easy. The main consideration is whether it is useful in this industry and whether the founder has an understanding of the industry and big model technology. There are not many people who understand both aspects. He can also make products that are well used by users and receive money. It is relatively easy to judge based on these two aspects.

The competition for big models is essentially a competition for resources, including financing capabilities, talent density, graphics cards, computing power, etc. So big models are more like the food of big companies. If OpenAI had not received so much financial support from Microsoft, it might not have been able to do it. Even if a small company raises $1 billion, it is difficult to continue to stand at the table.

Interface News: Is this the main reason why you don’t invest in large models?

Wang Xiao:Yes. To put it bluntly, it is hard to see clearly. There is no first-mover advantage, no commercialization potential, and no absolute technological gap, so it is difficult to invest.

Interface News: But in fact, unlike OpenAI, it is also possible to invest in an advantageous startup company in China.

Wang Xiao:Yes. I didn’t invest because such companies cost a lot of money, and the commercialization prospects are unclear at present. If large-scale model companies do not develop applications, there is actually no clear commercialization path, and it is also difficult to charge through tokens.

Character.AI also raised a lot of money at the beginning, but later found that it could not scale its model, so it took a group of people back to Google, and the rest continued to work on applications, which was equivalent to starting to use other people's models. Small startups will eventually find that the money they raised is simply not enough to iterate on others' models, and then the model capabilities will have problems.

In two years, when some startups have little money, the truth may come to light.

Interface News: Have you looked at big model startups at that time?

Wang Xiao:I have seen some of them, some from the "Six Little Tigers" and some from others, some of which charge very high prices right from the start. Of course, competition is voluntary, and startups logically have few opportunities. If there are any, they should be at the application layer, so that China's AI startups can flourish.

Starting a business requires building an ecosystem. A good ecosystem should essentially have some people working on the underlying layer and some working on the application layer, rather than everyone rushing in to do the same thing.

Interface News: Every company now thinks that they will create a so-called super app in the future, which theoretically will be large in scale and have great business prospects. How do you judge this?

Wang Xiao:Super apps need scenarios. These companies are essentially making models. Now they have all done conversations and chats. In fact, user retention is challenging, so it depends on whether there are sufficient capabilities in that scenario. There must always be an entry point.

Jiemian News: If a company that has the ability to build large models recruits a team that is particularly familiar with scenarios, will they be more likely to successfully create such an application than an application company?

Wang Xiao:Not necessarily. A company that develops applications with a big model is not necessarily more valuable than a company that directly uses other people's big models to develop applications, unless its big model is unique, not open source and cannot be used by others, and can solve problems that others cannot solve.

We are now also starting to look at some big models of other modalities, such as music. Companies like Suno do not open source their big music models, so others cannot use them. I think it is okay to use them for applications.

Interface News: For the companies you invest in now, will you push them to commercialize, or just let nature take its course?

Wang Xiao:It is a natural process. To B (enterprise) will certainly start commercialization from the beginning, and To C (consumer) will maintain user growth first, and make appropriate commercialization attempts in the process, rather than trying after reaching a large scale. Because tokens have costs, there must be some commercialization to match user growth.

Interface News: It will also cost a lot of money to achieve large-scale growth in this field.

Wang Xiao:Yes, you also need to invest money to acquire users.

Interface News: Will they be encouraged to take this action at this stage?

Wang Xiao:To encourage, you still need to get users, you must first run out. But in fact, commercialization is not that difficult. Now that users have developed a certain payment habit, it should not be too difficult to get users to pay a little money.

Interface News: Regarding the topic of market believers and technology believers, are you more concerned about when AGI will be realized, or when the AI ​​application track will explode?

Wang Xiao:At this stage, I am more concerned about when AI applications will explode. Technology must be useful. Useless technology is deceptive. Technology needs to serve people. Talking about technology without considering people is just deceiving people.

Essentially, we cannot just look at the market and only look at companies with revenue. We still have to look at it comprehensively. Now there are indeed VCs that only look at companies with revenue, and use VC prices to invest in companies with price-to-earnings ratios, because PE investments are now more cautious, and this approach can make money. But if everyone only invests in companies with price-to-earnings ratios, and buys them at several times the price-to-earnings ratio at a very cheap price, wouldn't China's technology be doomed?

No super app in sight anytime soon

Jiemian News: When investing in AI applications, is there anything that needs to be changed in the past investment logic? Can we think clearly based on past cognition and consensus, or have many new problems emerged?

Wang Xiao:Many new problems have emerged. Take the Internet for example. It is a connection-based business model that relies on traffic advertising. I think the big model is more service-oriented. The ultimate profit model will not be selling advertisements, but directly providing end-to-end services.

This is what we need to think about. We need to distinguish what can be borrowed and what is definitely different, so that we can establish our own cognitive framework.

Interface News: Are there any important changes that make you feel that you have to think in this way?

Wang Xiao:One change may be the data flywheel. The data barrier of Internet companies is the user scale, while the barrier of large model companies is the improvement of capabilities under the data closed loop. The data flywheel is the core point of this competition, so we will pay more attention to whether the generated data can be fed back for subsequent model training and make the model better.

Another issue is whether there is a unique source of data, which can also produce relatively large benefits.

Interface News: Many investors and entrepreneurs believe that this year is the first year of the explosion of AI applications, and it will explode completely next year. What do you think?

Wang Xiao:There are countless sub-tracks for AI applications, and some tracks may be easier to take off first, but each industry is different. I don’t think there is such a thing as an explosion of AI applications. I am more interested in which field the big model is suitable for application now.

Jiemian News: But after these remarks, do you feel that more people are looking at AI applications this year?

Wang Xiao:Yes, many of the projects we invested in have been looked at by other people, and quite a few people have given them TS (Term Sheet/Investment Terms List). This situation is rare in China now.

Interface News: So overall it’s a good thing?

Wang Xiao:Of course, an active capital market is a good thing, because only when companies can raise money can they grow bigger and be useful to the Chinese economy.

Sometimes we need a moderate bubble, which takes time. If we only look at the profit, we can't do it. It naturally requires some imagination. Without imagination, who would invest in Google when there were only two people? How could those US dollar funds invest in China's Internet giants? Only by daring to think can we dare to invest, and only by daring to invest can it have more possibilities.

Jiemian News: What are your current expectations for the technology and applications of large models?

Wang Xiao:One is whether the intelligence level of the big model can be raised to a higher level, and whether Scaling Law can continue to promote it to open up the situation. For example, the college entrance examination must at least reach the undergraduate level. Now the big model is only good in liberal arts, not in mathematics and physics.

Then there is embodied intelligence, or whether humanoid robots can become popular. This is also one of the few areas that is gaining some popularity.

I don’t expect a so-called killer big model application at the moment. It may be a variety of applications in various fields rather than a single application.

Jiemian News: The report mentioned that "without returning from the first principles of the model to the first principles of the product, it may be difficult to see large-scale use of products in the Internet era." How do you understand this sentence?

Wang Xiao:In the past, everyone was talking about the technical parameters and performance scores of the models, but this is actually not very meaningful. I think the essence is to see what differentiated capabilities can be provided to ordinary users to meet their needs.

Interface News: Will this product be an optimized version of a product that has already been seen, or a product that has never been seen before?

Wang Xiao:Some of them may be upgrades of the original products. For example, search is likely to be upgraded once, and its carrier may not even be a mobile phone, but a product such as VR glasses.

It is possible that one or two relatively good products will appear in each field, such as AI doctors, AI teachers, AI lawyers, AI headhunters, including AI programmers, AI customer service, AI sales, etc. This series of products are likely to exist.

Interface News: Who might be the first to do this?

Wang Xiao:There is no way to make an accurate prediction. Based on the current simplest logic of volume parameters, whoever is willing to invest money will be successful, such as Baidu and ByteDance.

Scaling Law is essentially a money game. It is still early to hit the ceiling. The intelligence level of the big model can be raised to a higher level, and then there will be particularly good super applications.