AI is moving towards application, and the balance is beginning to be reconstructed｜Annual trends & rankings

2024 is the beginning of AI moving towards scenarios and the general public.

After the hustle and bustle of the "Hundred Model War" in 2023, the big model industry as a whole entered a stage of building composite capabilities and differentiated advantages in 2024. The evolution of big model technology under the existing technical paradigm has slowed down, and the focus has shifted from training to reasoning, from competing in parameters, data, and computing power that are biased towards basic research to competing in long text, multimodality, proxy execution, and cost optimization capabilities that are more meaningful for practical applications.

The change in the direction of technological innovation is also driving the accelerated penetration of big models into the industry and the exploration of applications aimed at commercialization : After Kuaishou's Keling made a stunning success in the field of video generation, it quickly launched the Keling AI Director Co-creation Plan; with the support of the code generation capabilities of big models, more individual developers like Zhao Chunxiang have emerged; the combination of hardware such as mobile phones, glasses, headphones, and plush toys with AI has once again been put on the agenda, and assistants with more natural interaction and agent execution capabilities have become new competitive targets...

Another change brought about by the accelerated penetration of technology into the industry is that Internet giants have also begun to adjust their own businesses and find ways to adapt to the future AI era. In this process, the combination of model capabilities and scenario accumulation, focusing on the competition for new entrances and new solutions has begun.

Manufacturers and developers are welcoming their own opportunities for innovation. Especially after DeepSeek dropped a bomb at the beginning of the year, innovative explorations oriented to applications, empowering individuals, and reconstructing balance have become possible at a lower cost.

Trend 1: Large models enter a plateau period

From the perspective of the evolution of basic technologies, the evolution of large models has entered a plateau period. At this stage, the technological dividends brought by architectural innovation have been digested. In addition, the scaling law that drives the rapid iteration of this architecture has also begun to lose its momentum due to insufficient data and computing resources.

Although both OpenAI and Anthropic's CEOs denied that large models had hit a development bottleneck, neither OpenAI's GPT5 nor Anthropic's Claude 3.5 Opus was released within the originally expected time frame. The Wall Street Journal and Bloomberg both stated in related articles that the main reason for the failure of both models was that the performance of the models could not match the operating costs.

This also leads to the fact that in the second half of 2024, large model manufacturers will shift more of their focus from pre-training supported by large-scale computing power and data to improving algorithm efficiency and reasoning capabilities that rely more on reinforcement learning technology. The outbreak of DeepSeek is, to some extent, a reflection of this trend shift. In addition, large model manufacturers have begun to launch a richer model matrix around video generation, video understanding, voice interaction, agent execution, etc. in addition to large language models.

This shift in technology research and development marks the beginning of competition in the field of AI applications. Without a fundamental breakthrough in the underlying framework, both cost reduction and capability expansion are essentially creating an environment for the emergence of applications . The former will lower the threshold for technology application, while the latter will enable large models to adapt to more scenarios and have the ability to solve more complex problems.

Changes in big models are bringing about the rapid development of AI applications. First of all, capital in the primary market has begun to shift its focus from big model technology to big model applications. AI application entrepreneurs who had difficulty raising funds in the first half of 2024 have become the darlings of capital in the second half of the year. In the second half of the year, we can see that AI content platform Dream Dimension, Yuedian Technology, which focuses on enterprise-level Agentic AI, Caizhi Technology, which creates dedicated knowledge models, and AI personalized oral learning platform Keli Oral Language have all received financing. Early projects, small investments, and flexibility and low-keyness have become the key features of this round of AI application investment.

Secondly, Internet giants have also begun to look for ways to combine AI with their businesses. We can see that Baidu Wenku is revitalizing itself through AI, Alibaba has included Tongyi App in the Intelligent Information Business Group to explore AI to C, and Kuaishou and ByteDance are also constantly optimizing the product experience of Keling AI and Jimeng around video generation. The exploration of AI for vertical application scenarios in 2025 is also expected to set off a new round of innovation for major companies.

As capital shifts to AI applications, large model manufacturers are beginning to come under pressure. Among the six big model companies, Zero One Everything took the lead in turning to small-parameter, moderate industry model development, and transferred the pre-training team and Infra team to Alibaba. This is a landmark node for the differentiation and integration of large model manufacturers. Manufacturers with more basic technology innovation capabilities and large manufacturers with abundant resources will continue to stay at the table, and more large model manufacturers will turn to application and commercialization before their resources are exhausted.

Trend 2: Competition for new entry points

In the judgment of Bank of America, Agentic AI with stronger autonomous planning and action capabilities is bringing about a super innovation cycle. Zhipu's AutoGLM, Anthropic's computer use, and OpenAI's Operator are all working towards this goal. In the near future, we are likely to get used to using AI assistants to invoke these agent execution capabilities to achieve operations such as ordering takeout and booking tickets.

As we pointed out in "Agentic AI is restarting the entry battle", becoming the AI ​​assistant that is closest to users, connects the most users, and can facilitate more agent collaboration is a platform-level opportunity that has been recognized by everyone in this innovation cycle. Whoever can win the combination of AI assistant + agent is more likely to reap the new entry bonus and occupy a more important position in the new food chain.

The competition for assistants is getting more intense in 2024, and vendors have started a new round of competition for entry points. Behind this, native AI products such as ChatGPT, Gemini, Kimi, Doubao, Wenxiaoyan, and Tongyi are working hard to transition from chatbots to AI assistants with AI search capabilities, connecting to many agents, and even agent execution capabilities.

At the same time, Internet products such as Quark and Zhixiaobao, and hardware manufacturers such as Xiaomi and Ov are also focusing on AI assistants. Quark, which has accumulated search capabilities and network disk tools, attempts to achieve seamless switching of functions by transforming into an AI assistant and further activate its huge user base. Mobile phone manufacturers with hardware accumulation are relying on AI assistants to create core selling points while preparing for the arrival of AI hardware.

Of course, we can also see the possibility of AI assistants growing from native AI hardware. The core capabilities of AI headphones and AI glasses are AI assistants. Without the support of AI assistants, Ray-Ban Meta is just a pair of glasses that can take photos and listen to music, and cannot carry Meta's ambition to connect virtual and reality. Similarly, the ultimate battle of the Hundred Mirrors War is actually the control of AI assistants.

Judging from the current competitive situation, there will be three key points in the battle for new entry points around AI assistants: first, whether the basic model capabilities are sufficient to support large-scale high-frequency applications; second, whether it is possible to establish a rich Agent ecosystem to make users active and help AI assistants get rid of the dilemma of buying volume and retention; third, whether it is possible to find the best combination of software and hardware to connect virtual and reality.

Trend 3: Big companies turn around

Many companies will face the dilemma of "it's hard to turn a big ship around" when they reach a certain stage of development. However, in 2024, we can also see that the big model is bringing new opportunities for adjustment and breakthrough to Internet giants that are stuck in growth bottlenecks . Whether they can seize this opportunity, reshape their business, and achieve a shift towards the AI ​​era will also affect the competitive situation among Internet giants to a certain extent.

ByteDance is one of the representatives of the rapid U-turn of large companies. ByteDance's response in the field of large models is relatively slow, but after a series of adjustments and sorting out in 2023 and 2024, we can already see that ByteDance has established a complete AI business architecture from the underlying large model to vertical field applications with technology investment and traffic increase. Among them, Doubao, which has a huge user scale, Kouzi, which has not yet exploded, and AI hardware have become the link between ByteDance's technology and applications.

Alibaba's U-turn, on the one hand, has been actively exploring multiple dimensions such as basic capability construction, business scenario transformation, and new demand development under the slogan of "AI-driven"; on the other hand, it has been taking advantage of the U-turn opportunity brought by AI to deeply reorganize its existing business to make it better match the requirements of the AI ​​era.

Next, Alibaba may need to revitalize its AI to C business through integration . Wu Jia's refocus on AI to C business, and artificial intelligence scientist Xu Zhuhong's joining Alibaba to be responsible for the multimodal basic model and Agents-related basic research and application solutions for AI to C business are all concrete manifestations of Alibaba's efforts to re-empower its C-end business with the help of AI.

In addition to ByteDance and Alibaba, in 2024, there are also Internet products such as Zhihu, Alipay, Meitu, and What’s Worth Buying, which are seeking to find opportunities to break through themselves and achieve a U-turn by integrating AI capabilities. As a result, products such as Zhihu Direct Answer, Zhixiaobao, What’s Worth Buying GEN2, WHEE, and Wink were born, looking for the possibility of activating the original resource accumulation.

In comparison, the shift of Tencent, Kuaishou and other large companies in 2024 is not very obvious. Especially Tencent, although it launched the Yuanbao model and Yuanbao App, it did not become the tip of the knife to guide Tencent's AI, and WeChat's attempts in AI are conservative, giving people the feeling that Tencent has not formed a systematic AI action. Kuaishou has not come up with more highlights and attempts in the direction of AI except Keling.

Trend 4: One-person company

The continuous improvement of the big model infrastructure, especially the opening of the big model price war, is continuously lowering the threshold and cost for developers to develop products using big models. This enables individuals and small teams who are proficient in big model skills to launch good products and complete low-cost cold starts of products through platforms or communities such as Discord and Xiaohongshu, becoming a group that cannot be underestimated in the AI ​​application stage.

In 2024, we have seen the success of products developed by individual developers such as Stomach Book and Kitten Fill Light. We have also seen that directors participating in the Keling AI Director Co-creation Plan can create good animated short films with a small team of several people. There are also creators like Chen Kun, the director of "Mountain and Sea Wonderland", who are fully committed to the creation of AI film and television works.

The advancement of AI technology has greatly liberated individuals' ability to transform inspiration into products and works. In online jokes, young people who go home for the New Year will tell their relatives that they manage several highly capable employees such as Kimi, ChatGPT, and DeepSeek. In the future, this joke is likely to become a reality. More and more people will start to use DeepSeek to generate code or scripts, and one person can manage and operate an AI company.

The injection of big model capabilities not only empowers individuals and small teams to start businesses, but also brings new thinking about organizational structure to large companies. Many successful AI-native startups are creating greater value with small and capable teams, and large companies are also exploring how to use AI Agents to replace some manual labor. On the one hand, this is for the consideration of reducing costs and increasing efficiency, and on the other hand, it is also an attempt to regain the flexibility and agility of the organization.

As Microsoft CEO Nadella said, infiltrating AI into every layer of the technology stack is helping all business areas gain new profits and improve productivity. As AI continues to optimize workflows, we need to further find the secret to driving a more streamlined, agile, and efficient organization.

This also prompted Silicon Valley's thinking on the founder model in 2024. Terms such as founder model, nanomanager, and micromanager ultimately point to creating shorter communication links, founders' more detailed attention to the business, and the reduction of middle layers in the enterprise.

To a certain extent, it can be simplified to say that large companies are trying to transform themselves into organizations similar to one-person companies by integrating AI and changing their management models. The difference is that one-person companies are driven by one person to work for AI, while large companies are driven by founders to drive small teams composed of AI and employees to work for them.

Although 2024 is the year when the evolution of large models enters the plateau period, we still see the strong breakthrough capabilities of Chinese companies in the field of AI. In addition to the rapid growth of user scale and the rapid optimization of product experience, we have seen more Chinese AI companies focusing on technology research and development and innovation emerge. They represent the vitality and potential of China's AI industry and contribute Chinese intelligence to the field of AI.