In 2024, large models will continue to make progress. From Sora at the beginning of the year to the latest o3, newer and better models are constantly being launched. Has "involution" happened?

We must first determine the definition of "involution", which refers to a certain type of industrial model that, after developing to a certain form, falls into a "high-level equilibrium trap" and appears "growth without development". This If this situation cannot be broken, it will lead to stagnation and crisis.

This year, the scaling law of large models has been increasingly challenged. The computing power cluster for training models has grown from 10,000 cards to 100,000 cards, expanding tenfold, while the IQ of the model has not There is no improvement at this rate. There is no killer app (killer app) on the application side, and model manufacturers have begun a price war for bloodshed and volume... These characteristics are consistent with the definition of "involution".

The next question is, has involution put the large model into crisis? And where is the exit from involution?

In the involution cycle, it is difficult for an industry to maintain vitality and innovation power. The intensification of the involution of large models has also brought the industry into a period of adjustment.

The first thing that can be felt is the disappointment of the public and investors. In 2023, people use the phrase "one day for AI, one year for the human world" to describe the development of AI. The "Seven Sisters" of the U.S. stock market (Apple, Microsoft, Google's parent company alphabet, Amazon, Nvidia, Tesla, Meta) are even more so during this period. Stocks hit new highs repeatedly amid the stock boom. Recently, we have clearly seen that this enthusiasm has dropped.

OpenAI’s shareholders and service providers that access model APIs have publicly complained that AI capabilities have not made much progress. The just-concluded 12-day OpenAI press conference was mostly enhancements to existing models, products or technical routes, which were in line with expectations, but lacked highlights and could not provide strong support for AGI. Ilya, the former chief scientist of OpenAI, put forward "Pre-training as we know it will end" at the NeurIPS 2024 conference, which poured cold water on the public.

The skeptical attitude from all walks of life in industry and academia is a relatively dangerous signal, because the AI ​​winter in history stems from a lack of confidence and a decline in investment.

Another crisis signal is the intensification of homogeneous product competition and elimination rounds.

The competition for basic models has also become particularly fierce in 2024. First, the number of models is too dense, and their performance is gradually converging. In particular, the gap between open source models and closed source models is rapidly narrowing, and they are becoming homogeneous. competition.

Second, model families from the same manufacturer are also being phased out at an accelerated pace. For example, GPT-3.5-Turbo has been retired and replaced by GPT-4o mini. The same is true for models from domestic mold factories, and users are willing to use them. No one wants to use the old model when it comes to new models that increase quantity without increasing the price, and are of higher quality and cheaper. GPTAfter the release of -4o Mini, API usage doubled.

The fierce homogeneous competition prevents mold manufacturers from reducing investment in training new models, and they have to lower token prices in order to cope with the price war. The result is an increasingly heavy economic burden. It can be said that the current large model, whether it is the external macro situation or the micro operating status of the company, is not as positive as in 2023.

At the model level, the underlying technical routes, data bottlenecks, etc. cannot be effectively broken through in the short term, so it is inevitable to find a way out from the commercial level.

In 2024, we can see the involution of large models, bringing many challenges to business models.

The first is the cloud + API model. Bloody price cuts and price-for-volume exchanges are not the optimal solution.

Payment for API calls is one of the main monetization models for large models. Using token price reduction to win more large model businesses on the cloud and gain long-term benefits is the basic logic of cloud vendors' price wars. But at present, exchanging price for volume does not seem to work.

The reason is that B-end customers pay more attention to the long-term nature of the mold factory and the quality of the models. Quality comes first, price comes first, and reliability comes first. Therefore, we have seen that some cloud vendors that have succeeded in exchanging price for volume have relatively strong model capabilities. For example, after Wen Xin Yiyan’s two main models were made free of charge, the daily calls of Baidu Smart Cloud increased tenfold in one month. . Based on the Huoshan Cloud of the Doubao large model family, the number of token calls has also increased significantly, and the number of token calls by some customers has even increased by 5,000 times. This shows that new users will tend to prefer head models, while old users will either not consider replacing existing models, or they will put their eggs in the baskets of multiple head manufacturers, taking advantage of the price reduction to access more models, and ultimately leave behind products with high cost performance. of. Cloud vendors that do not engage in price wars, such as Huawei Cloud, which regards the Pangu model as a "sharp product", have also achieved good results in the B-end market. They have collaborated with industry partners to create large models of coal, medicine, and digital intelligence. The solution has been reused in multiple enterprises in vertical fields this year. When many industry users consider Huawei Cloud, they expect that the company has strong risk resistance, can persist in investing in basic models, and maintains stable business operations.

The above-mentioned companies explained that the foundation for the success of the cloud + API model is "quality comes first".

The second is the subscription system. Due to the involution of the large model, user stickiness and loyalty are low, and the membership market shows extremely high dispersion.

Because large models are updated very quickly. On the one hand, new models are often better in terms of quality and cost performance, and the "wait and see party" is more willing to wait and see; in addition, many old models will no longer be updated or retired. This makes members even less willing to be bound to the platform for a long time. This makes it difficult for mold factories to stop new marketing activities in order to continuously attract new users. The cost of acquiring customers remains high and affects the user experience. They need high-frequency pop-up advertisements to disturb users and develop multiple membership levels and paid benefit packages. , increasing the user’s decision-making fatigue. And new customers who are finally attracted often useAfter a period of time, you will switch to the free version, or update to cheaper friend products, and the long-term renewal rate is not high.

It can be seen that the involution of large models has made it difficult for most mold factories to convince customers and developers to establish long-term trusting relationships with them. This poses great challenges to subsequent commercial realization and value mining.

To bid farewell to involution, we must look for a way out. The large number and homogeneity of large models form a barrier lake with high density. To escape the involution, the river must be dredged to relieve congestion. Therefore, 2025 will be a year in which the commercial infrastructure of large models will be increasingly improved. Through more comprehensive "water conservancy facilities", large model users and developers will be able to access it more conveniently.

How to judge whether a large model is "extroverted"? There are several measurement criteria:

First, the openness or compatibility of the model.

As mentioned before, during the involution cycle, users are not willing to put their eggs in one basket or to be bound to a certain mold factory for a long time. This requires a strong openness in the model. degree and compatibility. For example, the free resource package of Tencent Hunyuan large model supports the sharing of multiple models such as hunyuan-pro, hunyuan-standard, hunyuan-turbo, etc., supporting third-party platforms and ISV service providers to provide customers with flexible selection and switching of multiple models. Model arena, etc., to meet the end customer's needs for multiple models.

The second is to develop tools in more detail.

Transforming large model technology into productivity requires more detailed support such as processing tools and workflows. For example, this time OpenAI created three professional tools for Sora: Remix, Blend, and Loop to support more detailed support. For good video generation, there are many Pro users who pay $200 per month for this. In China, the development tools we have tested, such as ByteDance’s Button Development Platform and Baidu Wenxin Intelligent Development Platform, are already very easy to use.

The third is "end-to-end" support for large model applications from development to commercialization.

There will be no national-level third-party AI applications in 2024. On the one hand, the model capabilities themselves need to be improved. Some AI agent platforms are filled with a large number of low-level, easy-to-replicate personal agents. The effects of conversation experience, understanding ability, multi-modal tasks, etc. are average and do not have much commercial value; on the other hand, On the one hand, many developers do not know how to commercialize AI applications, so they have not invested much energy in developing products that are lacking in the market to meet unsolved needs. This requires the platform to increase commercial resource support for developers.

In the final analysis, the technical ceiling will be difficult to break through in the short term, and the situation of saturation and homogeneous competition in the large model market will not be resolved. The commercial success of a large model depends on the success of users and developers' businesses. This is why a sound commercial infrastructure is essential.

Escape from the involution of the barrier lake. The question that all mold factories must answer in 2025 is: If the big model is water and electricity, what will users and developers get when they turn on the switch?