HomeArticle

CEO Tips · Going Global Season | AI is Changing, How Should Entrepreneurs Adapt?

CEO锦囊2024-11-13 10:48
Under the wave of AI, how exactly should entrepreneurs adapt to the trend and change?

In the constantly evolving AI industry of today, if domestic AI enterprises aim to expand globally, how should they make adjustments? As ordinary individuals in the wave of AI globalization, what opportunities can we grasp?

At 13:00 on November 7th (Wednesday), the 36Kr Live Studio invited Yunfei, the founder of Beta Silicon Valley Think Tank, and Huangshan, the founder of Forma Cloud, to join and discuss: AI is transforming, how should entrepreneurs adapt?

This live broadcast primarily focuses on the following questions:

  1. Among the three major Nobel Prizes this year, two were awarded to scientific research related to AI. What are your thoughts on AI sweeping the Nobel Prizes?
  2. It is rumored that at least two of the "Six Little Dragons" of the domestic large-scale models are planning to abandon pre-training. What are your opinions on this? Why does Meta invest a significant amount of money to train its own open-source model?
  3. This year, mobile phone manufacturers have intensively released AI phones. What are your thoughts on the future development of AI phones and Apple's layout in the AI industry?
  4. If domestic AI enterprises go overseas, what are the common needs and unique advantages?
  5. From Clay Camera to Talkie and other social applications, AI products are confronted with the issues of homogeneity and short life cycle. Which type of product do you prefer and how do you view the homogeneity phenomenon?
  6. For AI companies, which is more crucial: achieving PMF more quickly, creating differentiation in the underlying capabilities, or having their own user data?
  7. For Chinese AI enterprises that wish to go global, please offer three pieces of advice from both of you.

 

The following is the conversation between the two guests and 36Kr, and some of the content has been sorted and edited:

36Kr: Among the three major Nobel Prizes this year, two were awarded to scientific research related to AI. What do the two of you think about AI sweeping the Nobel Prizes?

Yunfei:The Nobel Prize this year is a signal that further affirms the inevitable trend of AI development. In Silicon Valley, we had a conversation with Garry Tan, the CEO of YC, and we reached a consensus: In the next ten years, all companies will adopt generative technology, and their products will possess the characteristics of generative technology. This is an irreversible trend.

We often compare AI to a dragon that has been released from its cage. Although there are still concerns about safety and control, the dragon cannot be put back into the cage, and everyone's life has essentially changed as a result. Specifically, whether it is Chinese enterprises going global to the United States or local outstanding talents starting businesses, the directions can be roughly divided into three levels: The first is the chip layer, which is analogous to the power plant in the electrical age; the second is the model layer, which is analogous to the power distribution company; the last is the application layer, which is analogous to the light bulb in the electrical age. In 2024, entrepreneurs began to focus on creating various light bulbs, that is, innovating at the AI application level. Especially in Silicon Valley, they are exploring how to develop the next useful AI application product. In contrast, there are fewer discussions on the model layer and the chip layer because the current pattern has initially formed.

Huangshan:Regarding this year's Nobel Prize, I think this approach is somewhat unreasonable. If AI practitioners can win the Physics Prize, then should the Literature Prize be awarded to ChatGPT, because it is undoubtedly the most discussed "writer" this year. Of course, this is just a joke. In my opinion, the impact of the Nobel Prize on the AI industry might not be significant. Because the two scholars who received the award, John J. Hopfield and Geoffrey E. Hinton, their contributions were concentrated in the 1980s.

I can briefly introduce these two to you. Hinton is more renowned, and the backpropagation algorithm is still utilized in neural network training. Hopfield invented the Hopfield neural network, which is mainly employed for storage and recall functions. Although this type of network is not commonly used now, its core ideas are still adopted by modern AI technologies, such as the attention mechanism and the loss function. Hopfield's idea can be illustrated by a simple example: Suppose the three of us are chatting. If my face is covered, you can still determine whether I am present based on my previous behavior patterns. For instance, if I was often present when Yunfei and Wan Jun were present, then even if you don't see me directly today, you can infer that I am present. This idea can be applied to image completion, such as inferring the appearance of lost pixels through other pixels. Therefore, Hopfield's ideas are still important in the AI field and worthy of everyone's understanding.

36Kr: It is rumored that at least two of the "Six Little Dragons" of the domestic large-scale models are going to give up pre-training. What do the two of you think about this? Why does Meta spend a lot of money to train its own open-source model?

Yunfei:Firstly, Meta's approach is clearly to sacrifice short-term commercial interests in exchange for a long-term strategic advantage in the AI ecosystem. Specifically, Meta intends to promote the establishment of industry standards through open source, attract developers and researchers to innovate around its platform, thereby consolidating its influence in the AI ecosystem. Free is the most expensive customer acquisition method. Even if their model performance is only 70 points, by providing it for free, they can attract more users from the 85-point model in the market. This is its main strategic consideration. Secondly, it is for the construction of the ecosystem. Meta hopes to establish a variety of application products and companies based on Llama3 with itself as the center to promote the construction of the ecosystem. This not only enhances Meta's influence but also encourages small enterprises to develop on its platform. Finally, Meta can optimize the performance and security of its own model through the data and feedback of global users. These three points are the clear goals of Meta in promoting the AI open-source model.

Huangshan:I will discuss from three levels why Meta invests heavily in training open-source models. First, why does Meta train the model itself? Because they are large in scale and need to avoid being "strangled". Meta previously encountered restrictions when providing services on the Apple platform, such as limited access to advertising data, which made Zuckerberg realize that relying on a competitor's system is disadvantageous. For large companies, choosing a partner is a major decision. Secondly, why does Meta provide the model for free? Because free can promote ecological construction. Moreover, the user stickiness of AI large-scale models is low, and users can easily switch. Therefore, in the field of AI large-scale models, it is difficult for enterprises to attract users to spend money to use them unless they are the industry leader. Finally, why does Meta open the model weights? Although Meta did not open the source code and training data, it did disclose some information, such as papers. Initially, Meta hoped that users would sign up to use the model for free, but someone privately released the model and spread the seeds to the anonymous community. They finally decided to fully open the model and also realized the importance of building an ecosystem. Large models not only require deployment but also various manual integration operations, which are not Meta's main tasks. Therefore, Meta focuses on model development, while the community is responsible for applying the model to different scenarios and improving the interface, which is also beneficial to Meta.

36Kr: This year, mobile phone manufacturers have intensively released AI phones. What do the two of you think about the future development of AI phones and Apple's layout in the AI industry?

Yunfei:For Apple, as a hardware company attempting to do software, to be honest, it is a very awkward matter, and they have been making efforts. Currently, we have not witnessed a fundamental change in Apple in the AI field, and its corporate culture does not permit such a transformation. But I am still very optimistic about Apple's ability to do hardware next. They truly have the strength in this area and are also the leader in the industry. In contrast, Google, where the large-scale model Transformer originated. However, due to the internal DNA and culture, the AI "child" was not able to grow up after its birth. Instead, it was raised by others and became extremely strong, to the point of challenging Google. In general, software companies face such significant challenges in doing AI. As a hardware company, Apple faces even greater challenges.

Although industry insiders say that it is slightly better for software companies to do hardware than for hardware companies to do software. But looking at Google's Google Mobile Phone, there is still a gap compared to Apple's hardware. Therefore, if Apple wants to succeed, it needs to refer to Microsoft's approach on this issue. Continue to make its own money, and then invest in a company like OpenAI. You incubate and nourish it, and provide it with your customers and computing power, so that you can cultivate a second growth curve. Satya Nadella has set an example for Apple at Microsoft. Whether Tim Cook will copy the homework depends on his own choice.

Huangshan:First of all, I basically agree with Yunfei's interpretation of Apple's development of AI. Apple is rich in resources and has the ability to acquire small or medium-sized companies. They have been laying out in the AI field for a considerable period of time. The small functions that we used to see in mobile phones, such as intelligent photo selection and video editing, all have AI functions embedded. One of the significant advantages of Apple in the AI industry is the CPU and GPU hardware, especially the M-series chips. Although these chips may have some generation gaps compared to Nvidia's chips, it is only about two or three years. Moreover, Apple's hardware has its unique features, such as small size and low energy consumption. Therefore, in the AI industry, Apple still has the potential to develop excellent products.

Secondly, artificial intelligence in mobile phones is definitely a good thing, but I don't use Apple's AI technology very much. On the one hand, it is for privacy considerations. Although I believe that Apple does a relatively good job in protecting privacy, I still hope that this technology can be more mature. I generally have a conservative attitude towards new technologies. On the other hand, my decision is also based on a direct calculation. For example, the total memory of the iPhone 16 is 8GB, and to run the Llama3 8B model (with 8 billion parameters) with full precision, at least 16GB of video memory is required. In this case, we may need to wait until Apple gradually upgrades the hardware configuration until the memory is sufficient before using it.

36Kr: If domestic AI enterprises go overseas, what are the common needs and unique advantages?

Yunfei:There are many types of enterprises going global, including not only medium and large enterprises but also small companies hatched by large enterprises going global, and even listed companies going overseas to find a second growth curve. In general, the advantages of Chinese enterprises mainly consist of three aspects:

First, Chinese enterprises are known for their rapid execution capabilities worldwide. Whether it is design, software development, or user research, they can quickly implement concepts. Secondly, in terms of service. Chinese enterprises are very outstanding in providing humanized services. Compared with the United States, the service process in the United States is cumbersome and time-consuming, while Chinese enterprises can respond quickly and solve problems. Thirdly, the cost performance of engineers is high. Chinese enterprises can provide high-quality engineer resources at a lower cost. Under the same investment, the operational efficiency of Chinese enterprises may be three times that of American enterprises.

In terms of needs, we can see that there are roughly four ways for domestic enterprises to go global, one of which is not very feasible, and the other three are more effective. We can imagine a quadrant diagram to illustrate these strategies: The horizontal axis represents whether to conduct business in China or the United States in person. The vertical axis represents whether to participate in person or hire others as representatives. The upper right quadrant refers to those who are willing to move to the United States and conduct business in person. These people usually have an overseas study background and are fluent in English, and are suitable for conducting business in the United States. The second category is entrepreneurs who have succeeded in China but have not studied abroad. They may choose to settle in the United States and find a local CEO as a spokesperson. The third category is super giants, such as BAT, who continue to develop in China while hiring American leaders to expand the global market. The last unsuccessful strategy is to try to manage the US business in China at the same time. This method is difficult to succeed. For example, a recent leading domestic enterprise requires employees managing overseas business to move overseas or else they will be dismissed. This shows that there are great difficulties in managing business across regions.

Huangshan:In terms of human resources, China does have an advantage, and it is a recognized fact that the domestic competition is fierce. In addition, I want to pour a basin of cold water on the domestic situation. Although Chinese enterprises have an advantage in human resources, in the AI era, the cost of AI is more crucial. How much is the cost of the large-scale model used in China? This is a very critical question. Domestic enterprises need to establish advantages in the quality and cost of AI so that they will be competitive when going global. Everyone knows that we are "strangled" in the GPU. If time could go back 10 or 20 years, I believe the GPU problem would definitely be solved first. Although the development at the application level is generally feasible now, it is still the most important to solve this technical constraint, and the domestic side should invest more energy in this aspect.

36Kr: From Clay Camera to Talkie and other social applications, AI products are facing the problems of homogeneity and short life cycle. Which type of product do the two of you prefer and how do you view the homogeneity phenomenon?

Huangshan:I think the better field for AI applications is still the so-called monetizable field. No matter what application it is, you cannot expect to always be ahead of others, but before others surpass you, you can monetize. Therefore, I think this is a direction worth paying attention to. Specifically, which fields can be monetized?

First, it is content generation. The current cost of AI content generation is relatively high because it requires a lot of AI resources. But there are two trends that can be judged: One is that the cost of AI will definitely become lower and lower; the second is the tolerance of users for content errors. Large-scale models often make mistakes, but if the direction of the AI application product is well chosen, users will have a higher tolerance for errors. And if the user's tolerance for errors is zero, then this field is not suitable. Secondly, it is the data aspect of AI. AI is mainly generated through training data, and these data can be used as the core assets of the enterprise. You do not need to open it to others, just like Llama, if you do not share the data with others, others cannot copy its products. Therefore, AI training data is a track with good intellectual property protection. Thirdly, I think AI companionship is a good track because it is a rigid demand. If in the future, there is a device that can provide more real companionship, then this field will have a lot of room for development. Peers have not done well enough for the time being, but there will definitely be improvements in the future. Therefore, AI companionship is also a track worth paying attention to.

As for homogeneity, I think the problem of product replication is not caused by AI. Before the emergence of AI, any popular product was difficult to avoid being replicated. If a product is easy to be replicated, then it should be replicated and will inevitably be replicated, and this should be taken into account when developing the product. You can visit the website of Y Combinator (YC) and check out the companies they incubate in each batch. With this question in mind, you may have a new perception when looking at these companies. At least you can learn how to protect your product from being easily replicated, or at least increase the difficulty of replication.

Yunfei:We often discuss this issue internally, and the final conclusion is that "we should not focus on the track but should focus on the founders". Each founder has a deep background in their professional field, such as news or supply chain. Entrepreneurs have accumulated ten years of experience in these fields and know what the real scenarios and problems in the industry are. Then, by embedding AI technology into it, it will be easier to receive orders and monetize. In Silicon Valley, early-stage investors and entrepreneurs are clear that truly successful entrepreneurs can usually achieve sales in the first stage, rather than relying on a grand story.

For most people, a successful business model is often those seemingly unexciting or even a bit boring business strategies. This goes back to what was mentioned earlier. The key to the competitiveness of an enterprise lies in understanding the real needs of the other party, rather than just promoting a universal AI tool. The market has become numb to this kind of promotion and no longer believes that a tool can solve all problems. According to this logic, the more promising AI track is still in the SARS field, especially ToB SARS. I usually encourage AI entrepreneurs to go deep into reality, master seemingly boring business skills solidly, complete transactions, and achieve order closures. Many successful unicorn companies also start from a single order, gradually accumulate, and finally achieve platformization and productization. For entrepreneurs without specific field experience, it is recommended to consider those time-consuming but simple tasks, which are very suitable for AI processing.

Finally, the reality that entrepreneurs are currently facing is that if the AI tool only solves a single-scenario problem, it will be easily replicated and replaced. At the ToC level, users may choose to use a product because of its being free or the activity of the CEO, but eventually, they will turn to a new and cooler product. Therefore, the ToB field provides more opportunities because ToB is not a single-point link but establishes a connection with customers through multiple links. Replacing a ToB product requires a huge cost,