"The open source of large models and the open source of the Database era are not the same concept. Huang Yungang, managing partner of Source Code Capital, said.
The biggest problem with open source is that community developers have very limited contributions to the model architecture. "The most important reason why Linux and Android were able to outperform closed-source products at that time was that they both had a 'public construction' attribute, and the innovation of the developer community could promote the user ecosystem at the same time. "But for most of the current open source models, whether it is the just-launched Llama 3 or Grok-1, the community's help for model development is still very vague.
In Huang Yungang's view, if the optimization and training methods for the scene are the same, then the difference between open source and closed source will not be particularly obvious, because everyone is only users and no "ecology", for example, Grok-1 open source may be effective for user growth in the short term, but it is difficult to say how much help it will help future model upgrades;
If the open source model only sees the same code for developers, but cannot allow developers' contributions to circulate and iterate among more people like the Database era, the ecological advantages of open source are not so important, that is, the core threshold of AI technology openness has not been lowered.
"For entrepreneurs, it is still at the stage where the technical capability of the model equals the opportunity of the product. ”
This stage is very similar to the early days of PC Internet and mobile Internet, and the requirements for capabilities are mainly reflected in the technical side. From China's earliest Internet entrepreneurs Zhang Chaoyang, Lei Jun, Li Yanhong, Ding Lei, and Ma Huateng, to Zhang Yiming, Wang Xing, Huang Zheng, and Su Hua in the mobile Internet period, the first batch of pioneers are basically programmers. Foreign Internet early entrepreneurs such as Google's Larry Page and Microsoft's Bill Gates are also from technical backgrounds.
"But the biggest difference now is that the technology, the products, the infrastructure are all moving forward at the same time. Huang Yungang said. The Internet is the process of bringing existing things online, and there is no need to worry too much about the technical route being particularly adjusted, and more attention is paid to business models and product design issues. However, in the AI era, technology itself is developing rapidly, and the speed of product iteration is also accelerating, and entrepreneurs may be too late if they want to wait until all the infrastructure is mature to make products. It took less than 90 days for ChatGPT to surpass 100 million monthly active users, which was not a speed before. The core of entrepreneurs in the AI era is forward-looking technical cognition - building products "halfway up the mountain".
Experienced product ghosts are not yet able to play at this stage. "This wave of AI opportunities has higher requirements for age and technology than before, and people who used to make PC products to turn to the Internet only need cognitive changes, and there are very few people who have gone through; Huang Yungang said.
The ability of the large model is the main task in the next few years, and the side task is to create features.
Huang Yungang believes that on the one hand, it depends on the scaling law, that is, more data, energy, and more computing power to create an exponentially more "smart" AI. The other branch line is the advance investment and layout of technology, such as long text, such as the more user-friendly prompt interactive logic.
In the end, China's large-scale model startups have to walk on "two legs", on the one hand, they are constantly creating new features (functions) to stimulate users and stimulate the market. At the same time, the general ability should continue to lead, proving that the "thickness" of the product is "thick" enough.
"On the other hand, from the market side, scenarios that are more dependent on large models have a better chance of running out of killer applications. ”
For example, the importance of large models to the next generation of search engines, Huang Yungang believes that at 90%, this kind of model is the product of startups should be the first echelon of attention this year, and then for content platforms, the importance of AI technology may be 50%, this judgment is mainly based on the asymmetric relationship between AI production tools and content (better AI tools may not necessarily produce more attractive content); games may be lower at 30%, and the design, theme and game product itself will be more significant to players than AI.
The greater the technological uncertainty, the more opportunities there are for young people. In the past decade, China's young entrepreneurial population has been growing rapidly, and there is no shortage of talent. This year may also become a dividing line between the ages and generations of AI "Dacheng" entrepreneurs. Ten years ago, the release of the iPhone 4 gave birth to the explosion of mobile Internet, and the post-80s generation chose to start a business between the ages of 21 and 30, which is relatively young. Now there are more and more post-90s and post-95 young people with high education, intelligence, and no path dependence.
History does not repeat itself, but there is always a similar rhyme. Some will admit the law, while others selectively deny it.
Huang Yungang will lead the management and independent operation of the Source Code USD Venture Fund from the next fund.
The more complex the market situation, the more important it is to keep it simple. Huang Yungang said, "Everything is to make the investment decisions of early-stage projects more flexible and focused, and after that, the Venture stage of Source Code will raise investment, manage and withdraw from an independent closed loop." ”
Cao Yi, founding partner of Source Code Capital, will continue to support the Venture business that embraces the AI wave, "Yungang is a very scarce young VC veteran who has a good set of investors, good ICs, and good team leaders, and has made great contributions to Source Code in the past seven years. Yun Gang is calm and introverted, and now is more opportune, and I am very much looking forward to Yun Gang leading the team to operate independently and seize the big winner in the new wave steadily!"
What kind of people are needed for AI startups right now?
Forbes China: What are the new requirements for the new generation of entrepreneurs in this wave of AI opportunities?
Huang Yungang:
Nowadays, while the technical requirements are high, there is also a need for young people who are young enough.
It may be that the feeling given to me by the post-95s is very different from the wave of post-80s that we voted for ten years ago, when the post-80s were between the ages of 20 and 30. This wave of AI opportunities is more dependent on technology, especially at this stage, most of them are still highly educated, and then they are also very smart. But just a product geek doesn't necessarily play out today. For example, the training of the model is too high a requirement for them.
That's why AI entrepreneurs are younger than before. The acceptance of new things is always the best for young people, and it is necessary to embrace the young.
If you are older, people who originally made PC products may also turn to the mobile Internet, which is a cognitive change, but there are relatively few successful people. The change of cognition is not simple to say, and sometimes, the invisible "window paper" is difficult to pierce. However, today there is another layer of difficulty, even if you have the knowledge, you don't have the technical ability. Therefore, only with a very strong ability to predict rapidly changing technologies can a product be made.
Forbes China: At this point in time this year, is it an opportunity for "people" or an opportunity for "things"?
Huang Yungang:
Whether it's AI, or any other field, such as robotics, consumer, or medical technology, there are similarities in these fields, and people and things are important.
Technology is constantly changing, and investment opportunities are starting to appear more and more in the angel, Pre-A and A round stages. As this group of companies grows, it may be three years before there are large growth investment opportunities. So, we're definitely focusing on the early Series A companies right now.
In the early days, people and things mattered. The so-called "thing" is the founder's cognition of things and his ability to choose the direction, but in the end, it must be "people", because after all, there is no data, not even a product.
Forbes China: What are the advantages of young people joining AI entrepreneurship?
Huang Yungang:
Products are divided into two categories, one is a process product, the demand is clear, as long as the process is sorted out clearly, many ToB products are like this. The second category is completely innovative products, made from scratch. Especially for the To C-side, such as social media, etc., these are especially in need of young entrepreneurs.
Should I be anxious about investing in AI in China?
Forbes China: At the beginning of 2024, the capital market's attitude towards large models has become very subtle, what are the opportunities for AI investment?
Huang Yungang:
Now, in addition to companies like Open AI that focus on technology and infrastructure, there are two categories: AI-enabled applications and AI-native applications. Native AI applications can be understood as new product forms that are highly coupled and dependent on new technologies.
In the case of ride-hailing, for example, both computers and mobile apps can be used to request a ride, which means that the scenario itself does not depend on the new technology. But only mobile apps can drive ride-hailing. At that time, the "Internet +" depended on how the former used new technologies to improve efficiency in the original scenarios.
Native applications for To C have not yet achieved a real technological breakthrough. Yesterday, I was chatting with a founder who does a data center. He believes that it is difficult to do this business in China, as if it is "selling iron", but "iron" is not made by itself, it is made by NVIDIA, and the development of the American market is relatively good, mainly because the scene is mature, and domestic customers are relatively scarce and scattered in comparison.
At present, the United States has found the biggest scenario, which is the To B side, such as the integration of various SaaS tools and AI. At the same time, the computing power consumption behind AI will also drive the development of infrastructure. In China, it is still difficult to use the To B side, and there are still relatively few large-scale enterprise-level applications.
At present, there are more customers on the training side than on the inference side at home and abroad, and the training side currently consumes a large proportion of AI computing power. Only when more applications explode in the future and better serve users will the inference cost be further greatly reduced. But this is still an engineering problem, and there will always be a way to solve it, and it will not take too long.
Forbes China: It seems that the current batch of AI applications is iterating faster, faster than the mobile Internet era? Why is that?
Huang Yungang:
I also feel like it's iterating very quickly. Because today's technical capabilities are dynamic, many products are rapidly iterated on in the process of technological change. Therefore, to build and build a good product, you need to start thinking about how to get to the top of the mountain and which top of the mountain to build the product on. When everyone sees the mountain of technology, and then sets off to make products, it is likely to be too late.
Forbes China: Reactively adjust and iterate, instead of actively catering to customers?
Huang Yungang:
That's right. The situation is more complicated now than in the era of mobile Internet, because the mobile Internet is an online process, and the establishment of infrastructure such as smartphones and App stores has given birth to a large number of applications. The rest is engineering, product design. However, now new stuff is being released every day, which surprises everyone.
It's also easy and painful, especially for entrepreneurs and app developers. Last week, I met with a serial entrepreneur who also talked about his pain. He was a very good person at making products, but now he sold the company. He said, I'm not an algorithm master, I don't train large models, I am a product maker, I don't have the ability to make a large model, so how to make a product? There is no very good large model in China, but I can't get stuck here for this, let alone sit still.
Waiting for the model to be done before making the product seems to be very efficient, but it may be a false proposition, because the understanding of the technology or the understanding of the model is actually a precondition for application.
Forbes China: Now everyone has different opinions on which open source and closed source are more likely to be, what do you think?
Huang Yungang:
We have to believe in China's closed source. Most people think that overseas open source is relatively strong, and China's closed-source model takes time and energy, wastes money, and finally cannot reach the level of foreign open source. But why do we believe that China is closed?
Because of two reasons, one is that closed-source optimization or training of certain scenarios is better, and can do deeper and more thorough at these points, rather than relying on seemingly powerful open-source models.
The second reason why I believe that China's closed source can defeat open source in the long run, or should defeat open source, is that the model of large model open source is actually different from the open source (open plus crowd source) that we understood before. Now it's more open source for the sake of open source, rather than gathering the power of the ecosystem to do things - so it only has open code but no corresponding ecology.
As long as there are sufficient resources, talents and patience, we have confidence and expectation in China's closed-source model.
Forbes China: What does it mean that there is only open source and no ecology?
Huang Yungang:
The current open source ecosystem is two different things from the open source ecology of the past.
The current open source, such as Meta's release of Llama's open source, is actually just training Llama and then opening it up. The developers in this community contribute very little to the Llama architecture, they just use it. This ecology is more for everyone to use, rather than for everyone to co-create this model.
The model itself still has to rely on Meta to keep moving forward, so his open source ecosystem today is different from the open source ecology of the DATABASE era in the past.
Forbes China: Where is the friendliest track for AI startups?
Huang Yungang:
China is the C-end. Overseas to B is also a good track, and some to B tracks with a lot of space are not so high.
But it's important to have the idea that model-as-a-product startups are likely to be the first to come out.
For example, when it comes to the importance of AI to the next generation of search engines, I think it may be 90%, or even 99%, mainly because the dependence on these two is very high. At that time, search engines were not actually considered a product, but from the perspective of search box and interaction ability, as well as the ranking of page retrieval, there were indeed certain product attributes, but in essence, it was determined by the back-end search ability. Ranking and the corresponding algorithm, the back-end is very important, and the front-end is actually relatively simple.
Forbes China: In which other areas is AI most likely to create disruptive products? In the Internet era, after the maturity of search engines, the opportunity soon came to the content platform.
Huang Yungang:
For content platforms, the importance of AI may be 50%, and even if Sora is made, it doesn't mean anything. Just like the photo software or editing software back then, it doesn't mean that you can make Instagram or Douyin.
Content platforms and tools for producing content are closely linked, but they are not exactly equivalent, and both have to be strong. Similar to today's AI capabilities, it is not easy to say whether AI tools can be transformed into a good content platform.
In addition, the contribution of AI technology to some industries is not overestimated. For example, the contribution rate of the game may be 30%, and the other 70% depends on the game design, the theme and the game product itself. E-commerce may be 20%, because e-commerce is inseparable from the supply chain and must connect goods, merchants, etc., which cannot be replaced. AI can only do a good job of front-end matching and traffic operation.
Forbes China: What about the opportunities for AI nativeness?
Huang Yungang:
Of course, there are certain opportunities. Just like there was no concept of AI robots, now it is called "embodied intelligence", which is a special robot that the original robot put in certain scenes, or just a special equipment, not to mention the robot level. Today, with AI, embodied intelligence has become a completely new opportunity, almost 100% of the new opportunity.
At this stage, there will be people exploring various forms of products. The same is true overseas, there are all kinds of entrepreneurs. For startups like Mita, when they are making legal translation tools, I understand that they still have a certain accumulation of NLP (natural language processing) and AI, so they can discover new product forms faster. It still depends on whether we can make a product that continues to grow, whether it can achieve differentiation, and whether it has universal capabilities.
Opportunity & Preparation
Forbes China: How do you see the commercialization of AI now, and when will it really open up the market?
Huang Yungang:
The premise of commercialization is productization, and the productization quadrant is divided into 2B, 2C, overseas and domestic.
There are opportunities for the application of domestic ToB products, but the volume will be relatively slow. Overseas ToB is relatively prosperous, and more of them are AI+, or AI-supported Enterprise products. Because it is based on software products, the path of overseas commercialization is very clear, and the market is quite large.
Domestic ToCs are actually the same as overseas ToCs, and they are not particularly successful. However, one of the benefits of overseas ToC is that the subscription market is relatively good, so the Pro version of Midjourney and ChatGPT have considerable large-scale revenue.
The domestic Internet and mobile Internet are the same as in the period, both of which are free plus value-added models. Therefore, in the field of AI, there may be a high probability of following, then this is a free plus value-added business model, and the problem of high inference cost must be overcome. Therefore, it is now necessary to see that the challenge of domestic 2C commercialization is indeed relatively large.
While it's certainly not a good time to produce results at this point in time, it's certainly the best time to put in. Both Bill Gates and Google founder Larry Page entered the game at a very uncertain time. At the same time, they are also very technological, and through their understanding of technology, they can know what the technology can be used for, and then make good products and promote commercialization.
So now it is impossible for investors to wait, and entrepreneurs are even more so, they don't understand the model at all, they don't understand the technology, and when they are all mature, the big opportunity does not exist.
Forbes China: What is the most different thing about the large-scale commercialization of AI C-end compared to before?
Huang Yungang:
In the old Internet era, we didn't have the cost of reasoning at the beginning, or I could do 10 million DAUs before I started to think about how to commercialize.
Now, because of the maintenance of 10 million DAU, it will burn money in particular. Therefore, only a significant reduction in the cost of inference is the basis of the free plus value-added service model.
Forbes China: How long do you think the cost of services can be reduced?
Huang Yungang:
Reducing costs in engineering should only be a matter of time, and may be solved within a year. But the most important thing now is the model capability, and everyone is trying to promote the height of the large model, and then reduce the cost. If the model is not high, it can't even support the PMF of the product, and there is no point in reducing costs.
Forbes China: Will multiple models be integrated into one product, or will there be a single general model to support one product application?
Huang Yungang:
In the end, it is certain that this is all multiple models serving one product, and it is unlikely that one product will only call one model. Because many applications may be small models are sufficient, there are some functions that must be called large models.
Forbes China: Judging from the user data of AI applications this year, the data of some vertical applications, such as educational AI applications, has risen very well, but the growth of some PPT tools and applications like Notion has begun to become weak. Why is this happening?
Huang Yungang:
There are also different product types here, and finding out the point of combining such technologies can lead to rapid success. It could be just a small feature that can be integrated into a big product and get good results. For example, I understand that there are many factors that affect Notion, and one of them may include the issue of technical maturity, which may limit the penetration and growth of the product.
At present, Microsoft Copilot and many document products, such as Feishu, have AI functions for users to use. So Notion's features are diluted. For this reason, entrepreneurs need to find a breakthrough point, that is, it is conducive to growth, but in the end, it will accumulate into a relatively "heavy" product, and its "moat" will be relatively high. A single product is still relatively "thin" and easy to be "diluted", which is very dangerous. For example, in February last year, there was a company called Jasper that was very popular, but now no one mentions it because the product is too "thin".
Forbes China: What are you most worried about when investing in AI?
Huang Yungang:
Some people are pessimistic about AI. In fact, there is a certain amount of truth in this, we belong to the category of people who have a more comprehensive view. We will neither say "AI can't vote" nor "All in AI". Because "all in" is a dangerous thing, we have always chosen to look at the problem in a balanced way, and the difficulty is real. Even if it is an AI-native opportunity, this kind of technology has never appeared before, and it will not be easily made. Even if ChatGPT is very successful, Open AI is still a certain distance from the real product at this stage.
The rest is "AI+", which is to integrate AI capabilities into other existing successful products to make their performance and experience better, and it may be more difficult for entrepreneurs to subvert AI-enabled products than for AI native applications. So, in such an environment, everyone is miserable, as if there are not many new opportunities, or not enough for developers to do, including training large models. We don't think we need to be pessimistic, after understanding and mastering the rules, we will find that technology iteration and development need to be constantly repeated, and the core is how to truly understand the essence of technology and business.
Forbes China: What are the changes in the investment strategy of the source code VC stage?
Huang Yungang:
There are two aspects, the first is to respond positively and invest selectively. At the same time, there is more focus on young people communication, and the younger the better. For example, recent graduates, doctoral students, master's students, and secondly, the organizational form, the fund mentioned above also pays more attention to this aspect, because new technologies are constantly changing, especially AI.
Our team also wants to bring in young people. Only peers who are entrepreneurs can better understand new technologies. At the same time, young people will also understand these equally young entrepreneurs better and understand the needs of their peers. So, it's very important that we build a team that is both young and knows how to invest and mingle with entrepreneurs.
For our organization, that is, our manager team, in the face of new technology, the planning is mainly based on investing in the early A round.
Forbes China exclusive manuscript, please do not reprint without permission
Never miss out on exciting information
▽