Huachuang Capital Xiong Weiming: Don’t rush to invest in AIGC applications, the real outlet has not yet come

Text｜Tencent Technology Zhou Xiaoyan

After ChatGPT became popular, large models became the "new favorite" of venture capital, but there are also some investors who have long-term layout of the AI track and did not make a move.

On the one hand, because AIGC burns too much computing power, if it wants to grow into a company with a mature business model, it needs to obtain strong financial support in the early stage. Xiong Weiming, the founding partner of Huachuang Capital, believes that the "1 billion US dollars" start-up capital is the ticket for the competition. The large-scale model of the stage belongs to the world of large companies;

On the other hand, Xiong Weiming feels that the current development stage of large models is like the Internet in the 1990s. The infrastructure is still being built, and the real application-side investment opportunities have not yet come. This wave of venture capital still needs to wait.

Under the upsurge of large-scale models, where is the new round of entrepreneurship? How can investment institutions seize new opportunities? Xiong Weiming, founding partner of Huachuang Capital, accepted an exclusive interview with Tencent Technology on the above issues. The core points of view:

1. It is not impossible for domestic companies to develop large-scale models, but the current level is relatively backward, but it is entirely possible to catch up with the development of GPT-3.5 or even GPT-4. The problem is that when we reach this level, others may have already indexed With the advancement of the level, the parameters have doubled, and we still need to face the challenge of insufficient supply of graphics cards. When others have already passed the inflection point, we may still be rising at a constant speed, and we will still be limited by hardware and computing power.

2. Whether an investment institution can invest in a large model mainly depends on the requirements of the LP behind it for the return cycle. If the LP wants to obtain a quick return in the short term, it needs to invest cautiously. At present, there are a lot of "pet-type" investments in the market, and the projects invested by some institutions have no other effect other than "showing off" to the outside world, and it is difficult to obtain substantial returns.

3. Many people think that it will be too late to invest when the large model has a mature business model, but it is not the case. When the iPhone arrived in China in 2008, some people thought it was too late. But in fact, Apple's Chinese market hadn't started to open that year. The situation now is similar to that at that time, and everything has just begun. If the AI user base can reach 5% of the entire Chinese population, you can start investing, and a small climax will already be formed when it reaches 15%.

4. In terms of investment direction, we still need to wait to see what business opportunities there are. For example, the game industry may benefit from AIGC, and basic writing such as financial reports.

Xiong Weiming, Founding Partner of Huachuang Capital

The following is the essence of the interview:

**Tencent Technology: You have revealed that Huachuang Capital has 5 investors who are looking at AI projects, but they have not yet made a move in the field of large-scale models. What are the considerations behind it? **

Xiong Weiming: The current development and application of large models is like the development of early computers. Its advanced productivity is only in the hands of a few institutions, and it was only in the scope of large companies at the beginning. For example, when computers first appeared, only a few companies owned them. In the 1980s and 1990s, IBM commercialized and promoted the popularity of mainframes. Banks, insurance companies, and communication companies began to use computers. However, the computer industry really exploded in the late 1990s, when personal computers really began to enter the homes of people around the world.

From the perspective of bearing high costs, only large companies can afford the computing power of large models. For example, GPT-4 needs to be equipped with a bunch of chips and GPUs. It can catch up to the current level of GPT-4. Looking back at history, the reason why Jeff Dean was able to make Google's AI business flourish was largely due to the low cost of infrastructure at that time. Quite good results can be achieved with 4 graphics cards with 580 parameters, but now the development of AI basically needs to start with 10,000 graphics cards.

In 2023, we are still in the early stages of large model infrastructure. It is even more difficult to make a large-scale model in China. On the one hand, due to the current international environment, the domestic GPU supply is tight; on the other hand, few people in China are willing to invest in this kind of forward-looking research, because its output is nowhere in sight. It has been five years since the 1.0 version of the Transformer model came out in 2017, but how many domestic listed companies have been able to invest in forward-looking technology for five consecutive years without seeing output?

So in general, in the early stages of the development of large models, startups are not suitable for investing in it. Whether it is graphics cards, funds or talents, they are relatively scarce. The development of large models especially needs to be promoted by large companies with deep pockets.

**Tencent Technology: But there have been several start-up companies in China that make large-scale models, and they have also started to run unicorn-level projects. Do you think they can’t do it? **

Xiong Weiming: Domestic companies are not unable to develop large-scale models, but the current level is relatively backward. This is like the rise of new energy vehicles many years ago, when BYD decided to join the EV market, which caused Musk's shock. It is entirely possible for us to catch up to the development of GPT-3.5 or even GPT-4, but the problem is that by the time we get to this point, others may have improved exponentially, and the parameters have increased exponentially, and we still need to face the shortage of graphics cards. challenge. When others have already passed the inflection point, we may still be rising at a constant speed, and we will still be limited by hardware and computing power.

Some entrepreneurs may think that due to limited funds, they don't need to force themselves to reach the level of Open AI, and it is acceptable to be a little behind. But this “slightly behind” gap can have huge consequences. Some things can achieve good results if you do something better than others. This is like the compound interest effect. If you are 1% better than others, you can gradually accumulate to form an excellent project. On the contrary, if you lag behind a little bit, you may become poor performance Good item.

We need to be clear whether the goal is to make quick profits or to pursue stable long-term returns. The popularity of large models has driven some people to make money by speculating in stocks, but when the stock price starts to fall, they still need to return to the business itself to make a profit. In the secondary market, as long as the stocks can be sold, it is possible to make a quick profit. However, in the primary market, this situation does not necessarily happen.

Therefore, it is not that we cannot develop large models, but we may not be able to reach the top level.

**Tencent Technology: So you must meet top-notch large-scale models before making an investment? As far as we know, many domestic investment institutions are unable to invest in several large-scale star projects, because they are overwhelmed. **

Xiong Weiming: At this stage, whether an investment institution can invest in a large model mainly depends on whether there is an LP "chasing" behind it. If the LP is more concerned about returns, it should not invest. At present, there are a lot of "pet-type" investments in the market, and the projects invested by some institutions have no other effect other than "showing off" to the outside world, and it is difficult to obtain substantial returns.

Now that the US dollar fund is relatively out of stock, they finally found one to invest in, so they quickly invested in and handed in a homework, even if they want to call the next LP money. They want to prove to US dollar LPs that in the hottest investment track, there are only 5 projects in China, and they have invested in 3, which proves that they are the most capable investors in large models in China. But from the perspective of industry laws, this is a relatively challenging thing.

Some investment institutions have indeed invested in unicorns. The products and commercial value of such enterprises may be realized, but the enterprise value may not be realized. It is very expensive to make a large model, at least 1 billion US dollars. I think the projects in this wave of upsurge will maintain at the valuation level for a long time, and the financing is also "flat". Worth 1 billion US dollars.

**Tencent Technology: If it is really difficult to sell at the model level after looking at the comprehensive factors, you can also look at some opportunities at the AI Infra level, such as what investment opportunities are there at the middleware and data levels? **

Xiong Weiming: I don't think we should be in a hurry, and we can't invest for the sake of investment. The current environment is actually a bit like the Internet era in 1995. At that time, China did not even have optical fiber or databases. However, in 2015, the application side began to emerge, from the previous Copy to China to the current Copy from China. We can foresee that the real large-scale model may not be fully mature until 2038, which is a long-term thing.

We should start to act after the infrastructure is complete. If you compare the development history of the Internet, around 1995, there were some rare companies that did .com websites, such as China.com, China Yellow Pages, and the predecessor of Sina.com, but in the end the real ones remained They were not the first wave of people to do this, but they have built infrastructure such as fiber optics and increased the number of Internet users. At that time, many mailboxes in China still charged fees, and everyone could only register for Yahoo Mail, and many schools did not accept emails. To apply to American schools, you still needed a Brother brand typewriter to type things. Today’s large model is a bit like that time.

Just like Yahoo launched the Chinese version in the early years, users gradually became familiar with and popularized this kind of service. After everyone became familiar with these gameplays, the Chinese version appeared, everything became popular, and gradually everyone can catch up.

In the Internet age, the moment when the real Chinese-style independent innovation emerges is when the game and SP business (Service Provider, abbreviation for Service Provider) rise. AI also needs to wait until such a moment. Its earliest users are people like us. Users in cutting-edge industries gradually changed from student users to ordinary people, and then white-collar workers in office buildings use it to write PPT, and there are even products that pay tens of dollars a year. After these appear, the Chinese market will come out.

We have to wait at least until the "SP" in the era of large models comes out, and there is a mature business model to use this technology. In the United States, there are many charging points for large models, such as user subscriptions, etc., but the charging path in China has not yet been opened.

**Tencent Technology: Wait until you see a mature business model before making a move, will you miss the opportunity to go to the poker table? **

Xiong Weiming: Many people may think that it will be too late to make a move until then, but it is not. In 2008, a friend asked me with an iPhone, is it too late for the iPhone to come to China? But in fact, in 2008, Apple's Chinese market had not yet begun to open, and the situation now is similar to that at that time, and everything has just begun.

The technological revolution brought about by the large model may last for a long time and affect the industry more deeply. Although the Internet has changed many industries, for example, the retail industry is almost completely built on the Internet, but in the software industry, there has never been a decent company in China. If there was no Internet, China’s software industry may follow similar products in the United States. The pace of development, such as the emergence of the mail payment model, etc., but after the emergence of the Internet, many intermediate links have been skipped, and some business logic that should be formed has not been formed.

If the AI user base can reach 5% of the entire Chinese population, we can start investing, and a small climax will already be formed when it reaches 15%. Some users do not use AI now, perhaps because the product form is not yet convenient. I told my colleagues that it would be more efficient for me to use GPT to run the content for me, but I have to go to the Internet first, register an account, find a way to purchase Plus services, etc. It would be more efficient for me to write it myself.

Therefore, in terms of investment direction, I think we need to wait to see what business opportunities there are. For example, the game industry may benefit from AIGC, and basic writing such as financial reports.

**Tencent Technology: What kind of new opportunities do you think the game industry will create in this wave of new technologies? **

Xiong Weiming: Large-scale game companies will first make good use of this wave of technology to innovate. For example, after players enter the room, they do not need to wait for the system to match real opponents, but replace them with AI automatic matching. In theory, they can also do more complex interactions. Games are always the touchstone of new IT technologies, so we will first look at the development and changes of the game industry.

**Tencent Technology: Which other industries will be the first to be affected by AIGC? **

Xiong Weiming: The e-commerce industry will also be impacted. China's e-commerce is very developed, but in the e-commerce field in the future, only "old people" like us will shop in structured department stores like Taobao and JD. Young people may prefer It tends to be interested in e-commerce, such as TV shopping, virtual shopping guides recommending products, etc.

In addition, AI will also clean up the SaaS industry. Chinese people are particularly willing to pay for services. For example, we are willing to send express mail, so China's logistics industry is doing very well. The industry as a service is especially suitable for re-doing it with AI in China.

Another example is regulatory SaaS products. Every year, major insurance companies are fined a lot of money. Many companies will be fined if they pay taxes late or make mistakes. Perhaps AI can be used to automatically capture suspicious transactions, or AI can remind companies when to pay taxes. , This can save the company hundreds of millions of fines.

Further down is the business of man-machine integration, such as manufacturing and autonomous driving industries, where natural language can be used to give orders to machines.

In terms of working mode, the productivity unit of a company will also become smaller and smaller. A company with only 20 people in the past may now be able to make very complicated things with three or two people. For example, Midjourney only had 11 people at the beginning, but They make a great company.

**Tencent Technology: In these segments that may be transformed by AIGC, what investment layout has Huachuang Capital made? **

Xiong Weiming: As mentioned above, in the field of segmentation, I think AI will transform games and pure digital content first.

But we don’t plan to invest in these two areas, because there are ready-made large platforms in these fields, and large platforms will use advanced technologies most handily. Startups are better at organizing small and beautiful teams, such as quickly producing creative content based on technology.

We prefer productivity stuff to entertainment stuff. But in terms of productivity, it also depends on the situation. For example, I think live broadcast e-commerce is productive, but I think they are not as good as artist brokerage companies or traffic platforms, MCN, and Taobao. We still pay more attention to new industries and new occupations that can be combined with AIGC to improve productivity.

When Zhuang Chenchao worked as a convenience bee, he tried to change the pattern of 7-11, but later found that a real person was still needed to guard it, because replacing manpower with AI added a technical cost, which seemed not cost-effective.

As soon as AIGC comes up, it can do something that completely replaces manpower, such as the customer service industry. It does not rule out that some existing companies will soon be able to leapfrog AI into products. The original team of thousands of people may now be enough with 100 people. The rate quickly increased. This is an industry we like, and we use AIGC to directly support industries that have not risen in China.

Tencent Technology: In the new industries and new occupations you describe, is it possible that some new native AI applications will be born?

Xiong Weiming: I don’t think there is a so-called original industry. There are only industries where people do not do well or are unstable. The output of AI can do things well. For example, no one wants to answer the phone in the middle of the night, but AI can be used as customer service. , and even automatic attendance; for example, if a mobile phone falls into the sewer, people may not be able to retrieve it, but AI can. AI can do things that span time, domains, and cycles.

Humans may provide AI with many clues, and after clues are given, AI will complete the tasks. Just like writing a manuscript, maybe after we finish talking, the reporter still needs to organize the recording, adjust the context, etc. With AI, you can first feed your own writing style to the machine, and let it organize according to your style, from the entire publicity to advertising In the closed-loop process, AI can do better than humans.

**Tencent Technology: After listening to your description, now is not the time to "chasing the wind", but the time to "wait for the wind". You need to always pay attention to the next innovation point. Will you feel anxious during this process? **

Xiong Weiming: Don’t be anxious. After all, the wind has not come yet, and I don’t know how far the wind will blow. I think domestic companies are still easy to get stuck in computing power. Even if the computing power is OK this year, it will be OK next year. What about the year? After catching up with GPT-4, how to catch up with GPT-5 and GPT-6?

**Tencent Technology: Computing power is a problem for the entire industry. What kind of mentality should the entire industry use to face this hard problem? **

Xiong Weiming: The whole industry needs to endure. This is a hardware problem that needs to be solved by everyone.

**Tencent Technology: In addition to computing power, in the wave of AIGC, what other things worry you? **

Xiong Weiming: I have some concerns about the development of AI.

First of all, we are not sure whether AI can be defined as life. Our understanding of human life is simply not deep enough. Human existence presents biological characteristics, such as protein and ATCG codes, which cannot be reflected in AI. So, are we equal to AI in the definition of life? Does I disappear and AI disappear, do they have the same effect? These I do not know.

Second, will AI develop self-awareness? This is also an unknown. If AI develops self-awareness, it could be the end for us humans.

Third, in what form will AI appear in this world, can we accept and recognize it as a form of life? This worries me a little. Because AI is obviously smarter than us, and it is a unified intelligence, regardless of national boundaries, data fusion and mutual influence, and its progress is even faster than our human beings. Considering that only some of us humans have higher education, AI does not need to go to college, and it can even be said that all of them have doctorates. These all make me feel a little worried.

**Tencent Technology: If AI develops self-awareness, will it definitely pose a threat to human beings? **

Xiong Weiming: Even if AI has no emotion, if it has self-awareness, it will think "I want to exist". If human beings are an obstacle to its existence, then it may destroy human beings.