AI investment, cognition should be early

In the second half of 2021, Wang Sheng, a partner of Innova Angel Fund, began to seriously study the large model and wanted to find a large model project to invest some money.

At that time, a number of large-scale model-related projects appeared in Silicon Valley in the United States, and the development was very rapid. For example, Jasper, which was founded at the beginning of the year, is very popular in the market, and it can generate marketing copy in various styles with just a simple prompt, and it made $30 million that year. Jasper's underlying system comes from GPT-3, a large language model launched by OpenAI a year ago.

Wang Sheng noticed GPT-3 and felt that there might be a big opportunity hidden in this direction. Although the quality of generative AI was not ideal at the time, he felt that it might be the right time for investment.

However, at that time, there were not many entrepreneurial teams in China that made large models. At the end of 2021, Innova Angel Fund set up a scholarship program at Tsinghua University to support students in the Department of Computer Science to do research, and the 2017 Ph.D. Fanchao applied for this program. Wang Sheng unexpectedly found that the large-scale model project that he was doing was exactly the direction he wanted to invest in.

In April 2022, when graduation was approaching, Di Fanchao revealed to Wang Sheng the idea of starting a business, and Wang Sheng decisively decided to invest, and Shenyan Technology was established.

After the angel round, he began to prepare for the next round of financing, and talked about forty or fifty VCs within three months, and none of them were willing to invest. "People don't understand it, they don't know what a big model is. Wang Sheng recalled to me.

On a Friday in late September, Wang Sheng took Di Fanchao to have a meal with a partner of Sequoia Capital, and three days later he received an invitation to discuss in detail.

Wang Sheng, Partner of Innova Angel Fund

The industry seems to be business as usual, but some investors with a keen sense of smell have already felt the upcoming waves of the large model industry in advance.

Two months later, at the end of November, OpenAI officially launched ChatGPT, and the entire technology circle was shocked because of the amazing effect. For a time, the AI large model exploded, and technology companies followed up, and the entire investment circle rushed up to see the large model project. Shenyan Technology subsequently received investment from Tencent, Good Future, Primavera Capital, Sanqi Mutual Entertainment and other institutions.

In the "100 model war" in 2023, VCs invested heavily in more than 20 start-up companies, and the project valuation was temporarily pulled to a very high level, and many investment institutions were confused. According to Wang Sheng's words, "Everyone can't understand the first, dare not vote second, and can't vote third." For the vast majority of early-stage investment institutions, it is no longer possible to play. ”

Innova Angel Fund got the ticket for this wave of large-scale model competition through the layout in advance. In addition, Innova Angel Fund has also invested in a number of infrastructure and application projects. Wang Sheng told me that after more than a year of investment, the valuation of Shenyan Technology has increased more than ten times.

Today, the craze for large models continues. Not long ago, I had an in-depth chat with Wang Sheng, and he told me about the process, experience and feelings of investing in large models, and also shared some of his observations and insights on large model venture capital. Here is the text:

01

There is a bubble in the valuation of the large model, which is compared to the "market dream rate"

Q: When did Innova identify AIGC as an important investment direction?

Wang Sheng: It's not too late for us to invest in AI, and we have spent a lot of time on the AI track before. The first wave of AI in China almost began in 2013, giving birth to the AI Four Dragons (SenseTime, Megvii, YITU, Yuncong), we began to invest in 2016, and invested in projects such as WeRide, IDRIVERPLUS, and Zhongke Yuandong, and later invested in some image and video AI companies, and then this wave of large models.

Q: Large-scale model projects have a high valuation and can easily cost billions of RMB, is this not very friendly to funds that focus on early-stage investment?

Wang Sheng: Early funds have to be invested early. After OpenAI launched ChatGPT at the end of November 2022, many VCs went to learn, improve their cognition, and began to look for projects. In fact, the vast majority of the projects that you see today came after that, and we started laying out before that.

Q: After the advent of ChatGPT, what other projects are suitable for early-stage investment?

Wang Sheng: Hardly.

Q: Why?

Wang Sheng: I think the market is different from what you see. In 2023, the entire AI investment will converge at both ends, with more than 20 VC investments on the capital side being active, and dozens of projects on the asset side financing a lot of money, without generalization on both sides, and the market is not prosperous enough.

According to the statistics of AMAC, there are 12,000 VCs in the whole market, and there may be less than 200 VCs in the actual shot, and there are only about 20 active ones, and the rest are watching the excitement and shouting with the flag. Because everyone can't understand the first, the second dare not vote, and the third can't. A bunch of entrepreneurs advocate how powerful the project is, but in fact, no one votes at all.

Q: Why is there a pattern of convergence at both ends?

Wang Sheng: There is a cycle of innovation, and when basic scientific research in many aspects reaches a threshold, there will be genius entrepreneurs who will transform scientific research achievements into technological innovation and drive a new technological paradigm.

After the emergence of a new technological paradigm, the first thing to do is to do infra (infrastructure), just like building a shopping mall, you must first have water and electricity and elevators, but doing infra will not be generalized, because it requires a large investment, and there are not many players who can afford to play. If the infra itself is not perfect, there is no point in making an application, and it will be crushed by infrastructure upgrades.

In addition, there is a limit to all the innovation that can be done today, and it will take time for good entrepreneurs and product managers to emerge. In 2007, the iPhone appeared, and the wave of mobile Internet began, but many important applications appeared after several years, Pinduoduo was in 2015, Douyin was in 2016, and ChatGPT has only been launched for a year, and no one knows where the boundaries of its capabilities are.

Q: So do the right thing at the right time.

Wang Sheng: Yes. In the past two years, the industry has been vigorously building infra, and it is not yet the time to do a lot of innovative projects, and this year there will be some applications that can quickly improve productivity, and the current consensus is that 2023 is the first year of large models, and 2024 is the first year of Agent, but it is not yet the first year of AI applications. All investors do things that are in line with the cycle in this regular cycle.

Q: Are the dozens of startups you mentioned getting investment based on large models?

Wang Sheng: Large models may account for nearly half, and some are infra, such as chip computing power, lossless networks, and optical modules. We also invested in end-side computing chips, but we don't believe that a startup can run on a general-purpose GPU like a data center, because the hardware is particularly unsuitable for VC to invest a lot of money.

Q: So the main investors in large models are large institutions and giant companies with strong financial strength?

Wang Sheng: At present, it seems that this is the case, unless you invest early, because it is very important for angel investors to invest early.

Q: Now there are more than 200 large models on the market, do you think there is moisture?

Wang Sheng: I think there are only a dozen or so companies that can belong to the head.

Q: The valuations of several projects in the head have skyrocketed, some as high as 10 billion yuan, do you think there is an overvaluation?

Wang Sheng: It's definitely overestimated. SenseTime is now more than 30 billion Hong Kong dollars, while it used to be 300 billion. Now that the valuation of the large model is so high, it is the market dream rate, and it may be higher in the future, but in the end, it depends on how to land and whether it can avoid the situation that the last wave of AI cannot find a particularly effective business scenario on the landing.

Q: How is the valuation in the industry now?

Wang Sheng: Based on the future income, everyone will give more than ten times or twenty times the PS (price-to-sales ratio), and now the actual PS is dozens or even hundreds of times larger. If you don't have income, pat your head and give a number.

Q: As an early-stage investor, would you be anxious about this situation?

Wang Sheng: There's nothing to be anxious about, because we voted for Shenyan Technology. This is by far the highest return of all large-scale projects. We can invest in these large-scale model companies if we want to, but some of them we choose not to sell.

Q: Why not shoot?

Wang Sheng: Our single investment does not match the valuation of the project well, which does not mean that we can't invest in it. In fact, for the vast majority of early-stage investment institutions, it is no longer possible to play, a large number of VCs first do not know which direction to invest, and secondly, when everyone is swarming, he doesn't know who is powerful, and then he can't find these people, even if he finds someone who may not be interested in letting him vote, he can't vote. We don't have these problems.

Q: Wang Xiaochuan's Baichuan Intelligence was very popular before, have you been in contact?

Wang Sheng: At the beginning of April 2022, I had a meal with Wang Xiaochuan, and he welcomed us to invest, but it was too expensive. At that time, it was a relative-friend round with a valuation of 500 million US dollars, and Innova's typical investment was generally 10 million yuan, which really did not match the plate of 500 million US dollars.

Therefore, I think that early-stage investment institutions need to recognize and make judgments early, and they can't react when VCs rush to the top, which is no longer something that early-stage investors can participate in. If you make a move before the market reaches a consensus, the price will be controllable, and the price will rise very quickly after the consensus is reached.

02

The core of investing in a large model is to look at people

Q: How do you know which project will make it?

Wang Sheng: The core depends on the team's ability to make large models. There is a big difference between China and the United States, investors often urgently require projects to be implemented in specific application scenarios, get orders to obtain benefits, and only talk about business models are not recognized, but on the other hand, Chinese users and enterprises are unwilling to pay, and rarely get a high annual income like OpenAI. However, investors can recognize the ability of a team to do a good job of large models, and we believe that large models may create great value when deeply integrated with certain industries.

Q: So investors are not looking at the current business model, but at a capability and possibility.

Wang Sheng: Yes, you can understand that the large-scale model projects invested in 2022 have no business models, and everyone is not sure how to make money.

Q: What are the areas of competence that are critical to a large model team?

Wang Sheng: The core is people, and this is true for all companies. When you have the best talent, you can attract investment, when you have money, you can buy computing power, and then you can attract better talent, and the flywheel can roll.

Q: What background or type of talent is most desirable?

Wang Sheng: I think it's dynamic. After LLaMA was open-sourced, everyone's cognition of the large model was quickly pulled, and the team with stronger engineering and financing capabilities was easier to win.

Q: What have Chinese entrepreneurs learned from LLaMA open source?

Wang Sheng: If you practice a large model from scratch to one, the experience, scientific research ability, and method accumulated in the past are very important, maybe some aspects are not difficult in terms of technology, but it is a comprehensive thing, involving a lot of skills and methods, once there is a point stuck, it may take a year or two to solve it on your own. Meta's large-scale model technology accumulated over the past few years was originally a secret, but as a result, it was opened source, and a lot of cognition was quickly leveled, which means that many pits do not need to be stepped on anymore, and those who have done huge system engineering in large factories have the advantages to be reflected.

Q: What's likely to evolve next?

Wang Sheng: I think LLaMA should also be open source in the future, probably multimodal, and still maintain a high level. When the capability of open source large models is capped in a few years, the greater challenge comes from the ability to make products. At that time, the biggest challenge for a team was not scientific research or engineering, but the operational ability of the entire system, which may have been the era of Wang Huiwen.

Q: In this case, Wang Huiwen entered the venue too early?

Wang Sheng: I think Wang Huiwen did something wrong, his role was wrong. His stage is to fight technology, scientific research, and engineering, and his advantages are in the later market and operation, which will be more friendly to him in the future.

Q: Some large-scale model companies are innovating on open source models, and some are shells, do you think there is an opportunity for these two?

Wang Sheng: There are three types of enterprises: the first type is the enterprise that really has the ability to make large models, the second type is based on the open source large model for fine-tuning and deployment, and the third type is to call the API interface of the external large model. The latter two belong to large-scale model application companies, and they should focus on application scenarios, focus on discovering needs, and do a good job in products, because its value lies in using the model well and meeting the needs well, and I think the opportunity is not small.

Once Infra is perfected, there will be a lot of applications. But limited by our imagination, these applications are based on the past, and now we use new technologies to do migration, and use new technology paradigms as a means of innovation and transformation, so hundreds of large models, many of which are made by large companies or even listed companies, are actually trying to use large models to empower the original business and optimize the business chain, tool chain, and workflow.

Q: But some companies seem to be just shouting and not acting.

Wang Sheng: Some listed companies shout loud slogans, probably to pull up stock prices. There are very few companies that are really making large models in the secondary market. In the process of technology paradigm migration, there are not many opportunities for startups, because large companies occupy scenarios, customers, and data, and startups can only wait for innovation and need time to emerge.

Q: Everyone said that they wanted to be China's OpenAI, do you think there are any companies that deserve this title?

Wang Sheng: I don't think there is yet.

Q: How big is the gap?

Wang Sheng: At present, almost all the large models in China are at the level of GPT-3.5, and the top one may still be two years away from GPT-4 and Gemini. On the one hand, there is a long gap in language ability, and on the other hand, there is a long gap in multimodality. Whether the gap can be narrowed depends on how high the ceiling of the large model technology paradigm is this time, including whether multimodality can also become generative, and I think it will become a generative multimodal large language model in the future. If LLaMA is not open source, our level today may be even worse than it is now, and it will be greatly left behind by foreign large model companies.

Q: Nowadays, many large models will say that they have surpassed GPT in a certain indicator and won the championship in some rankings, can this represent the real level of the model?

Wang Sheng: There are many and miscellaneous lists now, and the basic principle is to give questions to AI, but we found that some of the questions and answers on the list are wrong, so the higher the score, the worse the model ability. For investment institutions, the score of the list is not important, what kind of team can do something, in fact, everyone knows, the focus is still on people.

03

There is no advantage for large factories to make big models, and the opportunity belongs to innovators

Q: There is a view in the industry that it is good to hand over the basic large model to the big manufacturers, and there is no need to rebuild the wheel, what do you think?

Wang Sheng: It's the butt that decides the head. I believe that large-scale disruptive innovation should be done by start-ups, not by some business monopolies, because as vested interests, first, it is not good, and second, it is not conducive to the entire innovation ecosystem.

Q: On the surface, big manufacturers have a lot of advantages.

Wang Sheng: I don't think there are any advantages.

Q: They will talk about cloud computing, talent, resources.

Wang Sheng: Schumpeter's talk about innovation is not actually disruptive innovation, but creative destruction. Innovation will deal a heavy blow to the old forces, and all enterprises that do not innovate will be destroyed, and the big factories represent the old forces, the old productivity, and the old paradigm, and innovation itself is to subvert them. What big factories are willing to do is jump out and snatch the fruits of victory when something looks ripe. Projects that are progressively innovative and unable to form specific barriers are easy to be hit by large manufacturers, and projects with explosive innovation and flywheel effect are difficult for large manufacturers to kill as long as the flywheel turns.

Q: What do you think is the biggest difference between this wave of AI and the previous round?

Wang Sheng: Large model. In the past, small models did not bring generalized intelligence and could only perform well in specific scenarios, but now we have seen the emergence of large models, and the language task in the past was comprehension, now it is generative, and it can do much more.

Q: How is the difference reflected?

Wang Sheng: We often joke that humans can perceive it, and animals can do it, and they are no worse than us, but animals can do very little. It is only generative that there will be large, more complex, and more diverse economic activities like humans.

Q: Is the vertical model a false proposition? Some people think that if there is too little data, there will be no intelligence.

Wang Sheng: In fact, large language models already have a lot of data, and even a small amount of data can emerge from it, just like a person with a lot of education can touch the bypass, because the base is too powerful. More importantly, large models should be embedded in the workflow, so that AI becomes a working person, not just a chatbot. AI is only valuable in the industry if it's into the workflow, and that's what Agent does.

Q: Is it no longer possible for new teams to emerge and squeeze into the top camp in the competition of large models?

Wang Sheng: It's unlikely at the moment.

Q: Who are you more optimistic about with your existing players?

Wang Sheng: The existing ones have their own characteristics, and they actually represent different directions of exploration, and now winning has become a more complicated matter, and it depends on who really digs up the demand in the long run.

Q: Both Huang and Elon Musk have mentioned that they will see AGI (Artificial General Intelligence) by 2029. Are your forecasts optimistic or conservative?

Wang Sheng: I don't think it's a joke. From the perspective of the big paradigm, some people believe that AGI can be achieved based on transformers, but there is a paradox in transformers, it is easy to do complex things, and sometimes difficult to do simple things. I think there are some other paradigms that have an opportunity as well, such as embodied intelligence based on reinforcement learning, and whole-brain computing.

Q: What areas do you think are likely to be popular apps next?

Sheng Wang: Multimodal large models, generative video, and agents.

Q: In this wave of AI entrepreneurship, if entrepreneurs don't understand technology, will they have little opportunity?

Wang Sheng: Then wait a few years. Now the stage of the transfer of the old and new paradigms belongs to the technical masters, and when the underlying innovation is completed and the innovation is made into a platform, product masters and operation masters will be needed, and that is when a new wave of the public can participate.

Q: How did you learn as an investor and how did you get information?

Wang Sheng: In addition to the regular information channels, I think it is still necessary to communicate with powerful people, so as to improve awareness. In addition, investment is a very complex and comprehensive system, not that I only learn AI when I invest in AI, it also involves a lot of interdisciplinary knowledge, for example, I can read brain science, behavioral things, a broad knowledge system is still very important.

Edit / Dawn

—— end ——