Henry Chesbrough once proposed an innovative "funnel model" in his book "The Imperative to Profit Through Technological Innovation". Open innovation initially encourages a hundred flowers to bloom, but in the end only 10% of technologies can pass through this funnel, successfully reach the target market, and enter the next stage of commercialization and industrialization, while the remaining 90% of technologies gradually fade out of the public eye. vision.

The big model of 2024 has experienced a cruel test of the funnel order.

At the beginning of 2023, the issue that the industry is most concerned about is "whether China can produce top large models." In the following year, the number of domestically produced large models exploded, and the number of large models that had been registered and put online has reached more than 100.

So by the beginning of 2024, the question that everyone is most concerned about has become "How should we digest and utilize so many large models?"

Looking at it now, after the hundreds of models war , the basic large model has been "removed from nine to save one". Only about 10% of the large models with market vitality and high user activity stood out and entered the finals. The business market of large models has also evolved from a hundred schools of thought contending and has been condensed into two forces:

The first is technology giants represented by Internet and cloud computing companies, including Baidu’s Wenxin Model and Alibaba’s Tongyi Big models, Tencent’s Hunyuan big model, ByteDance’s Doubao big model, Huawei’s Pangu big model. The second is the leading startups represented by the “Six Little Tigers of AI”, such as Zhipu Qingyan of Zhipu AI and the Yi large model family of Zero One Thousand Things.

It can be said that the big model will go through a complete "funnel model" in 2024. But in the asset-heavy large model industry, competition is far more brutal than in general technology. We predict that 99% of large models will lose industrial space. Therefore, this model elimination competition is not over yet. Next, the innovation funnel of the basic model will be further narrowed, leaving only three or four products as AI infrastructure.

We still need to spend a little space to look back at the 2024 large model knockout competition and which seed players are left.

The large model patterns at home and abroad in 2024 all show a distinct Matthew effect. Overseas, giants such as OpenAI, Google, and Microsoft are still standing, while many large model startups, such as Stability AI, Adept, Humane, Reka AI, etc., are lining up for sale.

The situation in China is roughly similar. Technology giants represented by Internet and cloud vendors (Baidu, Alibaba, Tencent, Huawei, JD.com, ByteDance), and AI startups with outstanding financing capabilities (AI Six Little Tigers) have become dynamic competition in the large-model business market who.

The tide recedes, exposing the rocks on the beach, and the enthusiasm of industry and academia for making large models subsides. We will be able to see a clearer picture in 2024.Big model business model. Specifically, three motivations are needed for a large model to successfully pass through the funnel:

1. Sustainable resource investment. AI large models are an asset-heavy industry, and the Scaling Law is still in effect in 2024. As models continue to grow larger, the amount of high-quality data and calculations required to train new models is also increasing. It's like climbing a mountain, one step at a time. The advantages of leading companies in terms of capital, technology, data, etc. are increasingly prominent, represented by ByteDance.

ByteDance only began to fully invest in large models in 2024. The large bean bag model launched in May quickly emerged in the industry. The average daily Token usage exceeded 120 billion in September from 120 billion in May. 1.3 trillion. With the previous accumulation of Huoshan Cloud infrastructure and talent team, as well as the large-scale recruitment and investment increase this year, it has established superior barriers within a few months.

3. Model capability for rapid iteration. ByteDance's late comeback and rapid overtaking also shows that the large AI model does not have a particularly safe moat. Model capabilities are constantly depreciating. With new and more advanced model versions, old models are no longer valuable; with open source models, closed source models with similar capabilities will be abandoned by developers. This requires mold factories to continuously develop new and more powerful models and iterate old models.

For example, the Wenxin large model benefited from Baidu’s full-stack layout in chips, frameworks, models and applications, as well as the joint optimization of the Feipiao deep learning platform and Wenxin, and the Wenxin large model’s iteration speed , has always been at the leading level in the industry. In 2024, Baidu launched Wenxin Big Model 4.0 tool version and Wenxin Big Model 4.0 Turbo based on Wenxin Big Model 4.0, further improving the inference speed and effect. The fast iteration of the model helps to enhance the confidence of users and developers, and increases the stickiness of use and willingness to pay.

3. Commercial channels that can be monetized. Competition among mold factories is not only reflected in the research and development of basic models, but also in subsequent commercial promotion.

In 2024, large models will go from a "price war" to a "free war." In May, ByteDance brought the market price of domestic large models into the "centi era", and then Wenxin Large Models announced The two main models, ENIRESpeed ​​and ENIRELite, are completely free. As models enter the free era, mold factories must have other business channels to achieve revenue and recover their initial investment in large models.

Among them, most technology giants directly control user data, application products and channel resources, allowing large AI models to reach end users and pay for value. For example, the Baidu Wenku app has launched a series of AI functions such as smart PPT and smart drawing books based on the Wenxin large model through AI transformation. Its paying users have grown rapidly, and it currently has tens of millions of AI monthly active users.

AI startups are expected to stand out in the commercial market with their cutting-edge technologies and product solutions. Among the "Six Little Tigers", Zero One Wanwu made it clear that it will not give up the pre-trained model. at presentBased on the standardization capabilities of the Yi series base models, 01Wuxing has penetrated into the vertical and refined cutting-edge of business scenarios and launched the digital human solution "Ruyi" and the marketing short video solution "Wanshi".

In general, the large model industry in 2024 will mean that after large models are introduced to the market one after another, they have to face a narrow "funnel" exit and undergo a difficult elimination round. Internet and cloud computing giants and a handful of AI unicorns have successfully passed through the funnel and reached the next stage.

The baptism of the knockout rounds in 2024 will allow nine large models to survive, and the industrial structure will be more reasonable, leaving only about 10% of large models to enter the finals.

From the results, the large model shows the Matthew effect of "the strong always get stronger". So, how did these strong men emerge from the battlefield? If in 2023, the key battle for large models is the battle for infrastructure, and each mold factory is sparing no effort to build the computing clusters and high-end hardware resources needed to train large models, then in 2024, the key battle for large models will be Turned to the battle for the commercial market.

Competing for active users, this year’s large-scale business market has two themes:

Theme one is money-burning marketing.

Generative AI (AIGC) products based on large models can complete commercial transformation by providing services to users, which has also become the most direct and fastest commercialization path for large models. In 2024, AIGC products exploded. According to the "Registered Information on Generative Artificial Intelligence Services", as of November 2024, a total of 309 generative artificial intelligence products in my country have been registered. There are so many AIGC products with a large number of overlapping functions. Therefore, mold manufacturers have to compete for active users and increase their user base through large-scale and high-frequency marketing and marketing activities.

Dark Side of the Moon, Zhipu, etc. have all been reported to have spent heavily on marketing. The average customer acquisition cost for a single user of Kimi Smart Assistant is as high as 30 yuan.

These AIGC products that cost money to market have effectively increased brand awareness and user base, but it must also be admitted that it is not clear how much commercial value they can ultimately activate.

Theme 2 is towards application.

Is it possible for large models to make money without burning money to buy traffic and making money without losing money? Then you need to go to the application. Moving towards the industry, towards the majority of users and developers, and achieving commercialization through value payment, project payment, etc., in 2024, "large models can be put to use" is already a fact.

The first is the intelligent agent, which makes large models more useful. The application of large models has shifted from AI assistants to intelligent agents, such as Doubao, Kimi, Wen Xiaoyan, etc., which can automatically disassemble instructions and perform some simple operations. The level of "autonomous driving" is better, which greatly improves the usability of the technology.

The second is the tool chain to make large models easier to use. Wenxin Intelligent Platform, ByteDance Button, Alibaba Tongyi Qianwen, etc. have all launched support for intelligent agent technology and tool chains.With the support capability, ordinary people can quickly and cheaply create their own smart bodies. Among them, Baidu, which is betting on "AI application", has the most comprehensive layout in the intelligent agent ecosystem. It has launched development platforms such as APP builder and Agent builder, as well as locally deployed all-in-one machines and other hardware to support C-side and industry users in developing exclusive agents. ByteDance’s buttons are also very easy to use. Users can copy the official high-quality templates, quickly complete the development of intelligent agents based on private data, and publish them to products such as ByteDance.

“Spending money to buy volume” and “exchanging use for volume” are two major themes intertwined in the battle for commercialization of large models in 2024. A mold factory may use these two methods to ensure that The user base and market vitality of large models have stabilized its leading position at this stage.

Consumer-level technology has a basic rule: simplify complex technology to unlock breakthrough applications. Just like when we usually send emails, we don’t need to explore the SMTP protocol behind it. When we use mobile phones to pay, we don’t need to understand the encryption technology behind it. This "code-hiding" simplification makes the technology easier to use, so it can be more quickly adopted and expanded.

From this, we can predict what changes may occur in the "finals" of the underlying model:

The number of models will decrease. The large models led by technology giants and AI startups will continue to be shuffled, leaving only 3-4 basic models as infrastructure to support a variety of downstream applications. In this process, the sustainability of investment, iteration speed, and commercialization capabilities will still play a key role. Internet companies and cloud vendors have a greater chance of winning.

Usage is further simplified. At present, there is still room for simplification in the use of large model technology. For example, agent development has not yet achieved low-code or zero-code. Once professional plug-ins, knowledge bases, data processing, etc. for personalized scenarios are involved, the complexity of the development project will become higher, preventing some industry experts from developing more professional solutions. intelligent agents, which limits the explosion of large models on the B side. Therefore, in 2025, agent development and exclusive model training should become simpler and fool-proof. Readers with zero foundation who want to get started with AI development may wish to look forward to it.

The ecology becomes bigger. Everyone can get started with AI development, which involves training and analysis of private sensitive data and a variety of personalized functional requirements. Therefore, the basic model factory cannot only provide a simple encapsulation of an underlying model, but must support local training and deployment. The call and combination of multiple models and more diverse release channels require basic mold factories to incorporate AI hardware, AI terminals, vertical mold factories, channel partners, etc. into their own ecosystem to jointly meet user customization need. How big the "circle of friends" is is also a major model competition point in 2025.

In 2024, the midfield battle of the bottom model comes to an end and enters the finals. As the funnel of large models has been reduced to a minimum, the funnel of AI applications has just begun to explode. Listen, 2025, when “everyone can have AI” is just around the corner from usGetting closer.