Let's start with Kimi Chat
If you want to ask what is the most out-of-the-circle product in the domestic AI (Artificial Intelligence) industry in recent times, then Kimi Chat (Kimi intelligent chatbot program) may be the most suitable answer. In the past few months, this large language model has stood out from the crowd of strong competitors with its excellent long text processing capabilities. At the beginning of this year, KimiChat had fewer than one million users, but by March, it had reached more than three million users. With the rapid growth of users, this AI app quickly became the focus of major media reports, and the dark side of the AI unicorn company that developed this app, and the company's genius founder, Yang Zhilin, have become hot search words on the Internet. There are even some brokerages who are not willing to be lonely and hurriedly come up with a series of "Kimi concept stocks" for investors' reference.
With the explosion of KimiChat, the originally relatively unpopular track of "long text" quickly became crowded. Currently, KimiChat can support a text processing length of 200,000 words, and a text length of 2 million words in the beta. In order to be able to overwhelm Kimi, the ability of newly released models released by major AI companies to process text length can be said to be more terrifying than the other. For example, 360 soon announced that it had tested the processing function of 5 million words of long text, Baidu's Wenxin Yiyan announced that it would open the ability to process 2 million to 5 million words of long text for free, and Alibaba's Tongyi Qianwen opened up the long text processing capacity of 10 million words...... This kind of approach of being more than the head of the family is really "rolling" to a new height.
Kimi Chat has raised the limit of text processing to 2 million words, which is completely beyond the actual use needs. After all, 2 million words is almost the length of an entire set of Shakespeare's Complete Works, and the average user has little need to organize or refine such a large text at once. What is the actual value of major AI companies constantly "rolling" the length of text processing?
After reading some of the relevant materials, it dawned on me that fundamentally, AI companies are frantically "rolling" text processing lengths, perhaps not because it is of any practical use, but simply because it is not technically difficult to implement. At present, the technical paths for processing long texts are actually nothing more than external recall, model optimization, attention calculation optimization, etc., and each technical path has a large number of open literature and even open source programs. Therefore, it is actually easy for those AI companies with relatively strong strength to add this function, but it was not provided in time for various reasons. However, when KimiChat unexpectedly exploded with the help of long texts, this unpopular feature became one of the important reference indicators used by users to compare the performance of large models. In this way, in order to prevent their products from falling behind in the minds of users, AI companies have to devote more resources to this function.
Interestingly, an AI company was interviewed by the media and asked why the company was investing a lot of resources in the competition for long texts, rather than investing them in more high-end features, such as text-generated videos like Sora (an AI video production tool developed by OpenAI). The answer from the enterprise side is: the threshold is too high to roll. This answer is really intriguing.
Involution is a lose-lose ending
The influx of major AI companies into the long text track is just a manifestation of the involution of the AI circle today. In fact, since the release of ChatGPT (GPT intelligent chatbot program) at the end of 2022 triggered a boom in AI large models, the entire AI circle has experienced many rounds of involution. First it was the volume model, then the volume parameters, and then the volume multimodal...... In each round of involution, companies are fighting for the blood, but in hindsight, it seems that all involutions are meaningless, and even a little ridiculous.
For the development of the AI industry, the negative impact of unnecessary involution is huge:
For one, it's a waste for everyone. In nature, senseless involution is actually a manifestation of the "prisoner's dilemma". The knowledge of game theory tells us that it could lead everyone to a worse situation.
Second, it can do great harm to true innovators. While involution is wasteful for all AI businesses, the damage caused by it is not symmetrical. In reality, innovation requires a huge investment. If innovators are passively involved in this needless consumption, their already scarce resources will be stretched even harder.
Third, even for the winners, this involuted victory is fragile and unprofitable. Tim O'Reilly, a well-known internet publisher and the originator of the Web 2.0 concept, pointed out in a recent commentary for The Information, that the "Uber Problem" seems to be emerging in the AI community. He pointed out that a few years ago, when online car-hailing (online taxi booking) first appeared in the market, the major online car-hailing platforms had a crazy low-price war. In the end, Uber in the European and American markets, and Didi in the Chinese market relied on huge subsidies and became the "monopolist" of the market. However, unlike the monopolists of the past, Uber and Didi do not have any pricing power. As soon as they try to raise prices, users will switch to other small ride-hailing platforms and cruise taxis, and their market share will immediately lose a lot. As a result, while they control a huge market share, their profitability has been very poor. In O'Reilly's view, it seems that today's AI circle is also following in the footsteps of online ride-hailing, and if this situation continues, then even if an AI company finally wins from the involution, its market position will be very vulnerable.
Fourth, in the long run, this kind of involution will also cause damage to the welfare of consumers. In the long run, in order to compensate for this huge amount of investment, companies will either be able to charge consumers higher fees in the future, or they will need to try to reduce various costs to reduce costs and increase efficiency. But these efforts could ultimately hurt consumers. Regarding this, we can still find a correspondence from the development history of the ride-hailing industry. When the winners of the subsidy war found it difficult to raise prices to make a profit, they turned to increasing the share demanded from drivers. As a result, the driver's motivation to work has been hit, and the quality of their service has declined. Now, the quality of service of some ride-hailing drivers is even inferior to that of traditional cruise taxis, and this is a considerable part of the reason. Although there are certain differences in the nature of the industry between AI and online car-hailing, we can still infer that if the ineffective involution continues, the service quality that consumers can enjoy in the future will be greatly reduced.
Based on the above analysis, we can conclude that the crazy involution of the AI industry will eventually lead to a "lose-lose" outcome, from which neither enterprises nor consumers will benefit.
Why does involution occur?
Why is such a behavior that has a high probability of leading to "losing more" being chosen by many AI companies in the industry? In fact, it simply continues the general development logic of enterprises in the digital age, and the formation of this development logic is the result of the evolution of industry characteristics, changes in strategic thinking, and changes in industrial and investment relations.
Compared with the traditional industrial age, the emerging industries in the digital economy era have many new characteristics. Two of the most critical features are economies of scale and network effects.
The first thing to look at is economies of scale. In various writings on the digital economy, "zero marginal cost" is a buzzword. However, many books that discuss "zero marginal cost" ignore the premise that the premise for achieving it is a large fixed capital investment. For example, although the marginal cost of connecting one more user to an online car-hailing platform is close to zero, the amount of capital required to build the platform in the early stage is huge. This combination of high fixed cost input in the early stage and near-zero marginal cost in the later stage determines the scale economy characteristics of the digital industry: when a company becomes larger, its upfront fixed costs will be more amortized, so their advantage in average cost will become greater.
Let's look at network effects. The so-called network effect refers to the effect of the total number of people who use a certain product on the utility of users who use the product. In reality, many digital products have significant positive network effects. The larger their user base, the more users like them, and the more new users the product can attract.
Both economies of scale and network effects provide a metaphor: in the era of the digital economy, scale will be key to success in the market. Under this metaphor, scale transcends a series of indicators such as quality and differentiation, and has become the most important goal pursued by enterprises.
Along with this change, many well-known theories of corporate strategy have even emerged. Silicon Valley investor Reid Hoffman, for example, famously developed a theory of "Blitzscaling." At the heart of the theory is the idea that the modern business environment is full of uncertainty, and to overcome that uncertainty, companies need to prioritize speed of expansion over efficiency, and then move quickly to build their competitive advantage. Based on this central idea, Hoffmann makes a number of very unconventional assertions. For example, in order to ensure the speed of expansion, companies can tolerate imperfect products, do not care about user feelings, and tolerate chaotic management. In reality, the impact of the "blitzkrieg" theory is enormous. Whether it is a participant in the online car-hailing war back then, or an entrepreneur who is still actively involuting on the AI track, many of them are loyal fans of this theory.
One consequence of the "blitzkrieg" is that it has made the need for capital more intense than ever. In order to quickly achieve scale expansion, enterprises must "burn money" a lot. In reality, few companies can support such a huge investment. This is especially true for those start-ups.
In this case, the dependence of companies on external investors will become stronger. In order to gain the approval of investors, entrepreneurs must listen more to their opinions when making business decisions, and as a result, the management of enterprises will be alienated from "user-oriented" to "investment-oriented". Generally speaking, compared with entrepreneurs in the front line, investors will be much inferior in terms of market perception and information. Compared with entrepreneurs with entrepreneurial passion, investors will be more concerned about visible indicators, such as specific market share, and certain product features that are popular in the market. As a consequence, the direction of technology and market development will largely shift from entrepreneur-led to capital-led.
The story is being repeated in the AI industry. As we all know, the development of AI models is an industry that requires a lot of money. To train a large model, thousands of GPUs are required. This alone could cost hundreds of millions of dollars. As a result, with the exception of a few incumbent giants, most AI companies must seek the support of external investors. From the perspective of investors, in order to ensure the return on their investment, they will screen the capabilities of the enterprise. As non-professionals, their screening indicators are often items that are easy for the outside world to observe, such as whether the model developed by the company is leading in terms of the number of parameters, whether it has some popular functions in the market (such as multimodal capabilities, long text processing capabilities), and whether the market share of the model is higher than that of its competitors. In order to meet these requirements of investors, most AI companies will have to passively join the meaningless involution even if they are unwilling.
Why does involution have no future?
From the perspective of AI business operators, if they want to get out of the involution, they must first rethink their business model. As mentioned earlier, many AI insiders now believe in Hoffman's "lightning expansion" theory, believing that as long as they can seize market opportunities with the help of capital, even if their products are not so good, they can quickly gain a foothold in the market. As for the characteristics of the product, as well as the related profit model, etc., we can leave it for later. However, there is a big problem with this line of thinking. (1) There are no barriers to competition, and it is difficult to maintain expansion
In elaborating on the idea of "lightning expansion", Hoffman also emphasized the importance of building barriers after rapid expansion. This seems to have been overlooked. In the eyes of some, there are already two natural barriers in the AI space – economies of scale and network effects. Since these two characteristics naturally favor larger firms, they can help themselves resist market erosion as long as they focus on scaling the market. Unfortunately, this view may not be correct.
Let's start with economies of scale. As we have already noted, economies of scale are the result of a combination of high fixed cost inputs in the early stage and zero marginal costs (or low marginal costs) in the later stages. So, does the AI industry have such characteristics? It should be said that in the training and development stage of models, there is a requirement for high fixed cost input. However, in the rollout phase of the model, it is difficult to see a situation of zero marginal cost. In current practice, except for star models like ChatGPT, which can achieve a rapid increase in user volume without any publicity at all, most AI models must rely on a lot of promotion to attract users. According to reports, even for a product as phenomenal in China as Ki-miChat, its average customer acquisition cost has reached 10 yuan. More importantly, in order to provide corresponding services and maintain the loss of users, AI companies also need to pay a lot of additional costs. And when the scale of users reaches a certain level, the congestion effect will also lead to an upward trend in marginal costs. Based on these circumstances, it is not difficult to see that while it is true that there may be economies of scale within a certain range of the AI industry, it is by no means unlimited.
Let's look at network effects. Many operators of AI companies believe that AI, as a digital product, should have a network effect. But in fact, this is most likely a misunderstanding.
Essentially, there are two types of network effects: one is direct network effect and the other is indirect network effect. The so-called direct network effect is the direct impact of the number of users of the product on the user evaluation. The so-called indirect network effect refers to the indirect impact of the number of users of a product on user evaluation. In the platform condition, there is a particular indirect network effect known as the intergroup network effect. This network effect means that users on both sides of the platform will evaluate the platform because the number of users on the other side of the platform increases. For the platform, the inter-group network effect is critical, as it creates a "chicken-and-egg, egg-and-egg" echo effect. The existence of this reverberation effect allows the relevant subsidy strategies to often produce considerable multiplier effects. Therefore, in the online car-hailing war that year, the major platforms spared no effort to subsidize users.
With the above understanding of network effects, we can take a look at whether the AI model has the above-mentioned network effects.
Let's start with the direct network effect, even if it exists, the effect will not be very large. At least at this stage, large AI models are still used more as tools. Obviously, the number of users of a model will not have any impact on the performance of the product, and therefore will not affect the evaluation of users. Although in practice, those AI models with more users have some word-of-mouth effect, but not too much. On the contrary, people have a preference for early adopters of products like AI, and when new models with similar performance come along, people will be more inclined to try them, which will lead to the loss of users of old models. Take ChatGPT as an example, it started the earliest among all AI large models. So far, few new models have been able to surpass it across the board. But over the past six months or so, ChatGPT has been losing active users rapidly.
Let's look at indirect network effects. Theoretically, this exists. This is because AI companies usually use the interaction data between users and models for further model training. In this sense, the growth of model users does contribute to the improvement of model performance and the improvement of user evaluation of the model. However, considering that the amount of data required to train the model is huge. From a quantitative perspective, the resulting network effects are likely to be very small.
In summary, since the economies of scale and network effects of AI models are not enough to become natural barriers, it is difficult to easily maintain the market share obtained by relying on "burning money". Even if a company does succeed in capturing the market, it may be just a "monopolist" with no market power, as Uber and Didi did in the past.
(2) It is difficult for the model to make a profit, and it is difficult to return the investment
In addition to the lack of natural barriers, the bigger problem faced by AI companies is the lack of profit models. From a practical point of view, the common monetization models of existing AI models mainly include several types: one is to sell membership services directly to C-end users, the second is to sell API interfaces to developers, and the third is to provide AI as a cloud service to B-end users. However, as we pointed out, in the environment of high involution in the industry, the prices of the first two services have been pushed very low. In contrast, the third profit model can achieve relatively high profits, but it requires the enterprise to have cloud service projects as support. However, in reality, most AI companies do not have such a business. As a result, a large number of AI companies in the current market seem to be doing well, but it is difficult to really make money. Whether it is for the AI company itself or the development of the entire AI track, this business model that only relies on external capital transfusion is quite unfavorable.
How to get out of involution?
In view of the above problems, AI companies need to build barriers and find profit models as the two most important tasks when rebuilding business models.
(1) Construction of barriers
As early as the 30s of the 20th century, economists represented by Joe Bain had an in-depth discussion on the problem of market barriers. Later, a large number of scholars, including Bain's follower disciple Micharl Porter, further expanded on the issue. Both Bain and Porter's research highlight differentiation as a very important market barrier. While this may seem like a common sense, it is very instructive in the AI industry, where products are becoming increasingly homogeneous. At the operational level, AI companies can differentiate themselves from two dimensions:
On the one hand, they can choose some unique and high-barrier technologies to develop relevant functions, so as to ensure that they can stay ahead of their competitors in the market for at least a while. In this way, even if the competitors in question are developing the same features, they can ensure that they have enough time to upgrade the features so that their advantage over their opponents can be maintained. In practice, OpenAI's Sora actually uses this strategy. Because OpenAI's advantages in Wensheng video are very obvious, other competitors naturally retreat from the difficulties and dare not easily burn money to compete with them.
On the other hand, they can consider developing AI models specific to a specific industry. The dedicated model market is easier to defend and more difficult to attack than the general model market, requiring the accumulation of a large amount of specialized knowledge and additional funds, and it is also difficult to be imitated and seized by competitors. Although theoretically, the potential of the dedicated model market is not as good as that of the general model market, considering the current degree of involution of the general model market, it is a feasible option for AI companies.
(2) The choice of profit model
Let's look at the choice of profit model. In my opinion, there are two lines of thought to consider about this issue.
On the one hand, if an AI company decides to adopt a differentiation strategy and choose a certain niche market as a breakthrough point, then it can choose a very direct profit model and obtain benefits by charging fees for services. As mentioned earlier, in a niche market, businesses can easily build their own barriers, allowing for greater pricing power. In this case, even this simple profit model can ensure that the business gets enough revenue.
On the other hand, if AI companies do not adopt a differentiation strategy, it will be difficult to generate revenue by directly providing AI models. In this case, it needs to consider leveraging AI models as a means of engaging users and then generating revenue from other complementary channels. At the operational level, there are many ways to achieve this.
For example, an AI model can be bundled or sold at the same time as a complementary product. For example, although AI tools are now widespread, people don't know much about how to use them. With this in mind, AI companies can follow the example of hardware companies such as IBM (International Business Machines Corporation) that sell product bundles and bundle AI models with courses that teach users how to use the models, among other related services.
In addition, building a platform is also an idea. As mentioned earlier, OpenAI is now trying to build GPTStore, an online marketplace platform based on OpenAI's GPT technology, into a new platform. Although it has not achieved the desired effect at the moment, the overall idea is still correct. It is conceivable that if OpenAI one day straightens out the platform, then it can generate a steady income by charging commissions on transactions that occur on GPTStore, just like Apple charges an "Apple tax".
Of course, in addition to building an app store platform, building a social platform is also a good idea. Most of today's AI models only focus on human-AI interaction, and do not consider human-to-human interaction using AI tools. But in fact, there is a lot of room for imagination in this point. AI companies can build a platform for users to share their AI-generated works, and use this as a starting point to guide them in the direction of social networks. Once this network can form a certain scale, its profit potential will also be considerable.
All in all, if companies in the AI circle can give up their obsession with market size, choose their own niche market according to their own characteristics, and design their own profit models, then it may not be so difficult to get out of the crazy involution.