2025, The "Age of Awakening" for the Large-scale Application of Large Models?
The story of large-scale models transitioning from myth to reality teaches us: The true imagination of technology stems from "desensitization". After becoming desensitized to excessive expectations and bubbles, we also need to become desensitized to the technology itself.
Otherwise, the key value of application will become an elephant in the room.
In 2025, the domestic large-scale models that have emerged from the "Hundred-Model Battle" will undoubtedly continue to delve into the deep-water zone. However, who will remain at the table after one year has become an unknown mystery.
Baidu, Alibaba, and ByteDance are the obvious contenders. Their computing power, technology, and data, three hard indicators, are at the global forefront, and they have the potential to benchmark against GPT5 in the future. The "Six Tigers" and DeepSeek are coming on strong, but the failure of the Scaling Law and the disillusionment with the title of "AI Leader" seem to imply that the energy of these startups may be liberated into a commercial path that can generate real returns, avoiding direct competition with large companies.
On the first day of the new year, Li Yanhong wrote a letter to all Baidu employees. Between the lines of reviewing Baidu's technological genes, it points directly to AI applications.
Where will AI applications go in 2025?
Today, when the updates provided by OpenAI are becoming more and more "plain", the penetration rate of AI applications is getting higher and higher, but the market is worried that the investment in AI is not sustainable. As early as the WAIC conference in 2024, Li Yanhong, who emphasized the focus on "rolling out applications" and "no applications, no value", answered this question again.
01
Scalable applications are the "seeds" of disruptive innovation
From transportation and communication infrastructure to e-commerce supply chains, industries in the domestic market that eventually export capacity advantages to the global market often possess a characteristic: They have undergone the verification of high-density applications.
This verification depends not only on whether the technology can withstand the extreme values at the moment of traffic explosion, but also on the daily dependence and perception of ordinary users, thereby determining the direction of human technological innovation and implementation.
In the large-scale model industry in 2024, some causal inversions occurred.
In the world of large-scale models, the technological investment and cutting-edge nature of technology companies far exceed those of university laboratories. Foreign companies such as Google, Microsoft, and OpenAI, as well as domestic companies Baidu and Alibaba, entered the game earliest. The time difference between Wenxin Yiyan and ChatGPT is almost negligible, mainly due to Baidu's heavy investment in AI since the mobile Internet dividend period ten years ago, with more than 20% of its funds used for research and development each year.
This has led to a result: Writing papers, making innovations, finding scenarios, and incubating super applications. The entire chain of AI innovation has been integrated into the hands of these technology giants. Thus, the "innovative competition" has begun.
This kind of competition is not difficult to understand. Taking the scale, cost, and evaluation performance of large-scale models as selling points to attract more users and developers to join the ecosystem, and making a loss before making a profit. This is the classic Internet land-grabbing model.
But the problem is that the large-scale model itself is only a feasible entry point for innovation, and the real innovation must be rooted in the practical scenarios of various industries.
Just like 5G is only an entry point for innovation, 5G-native application scenarios such as short videos and live-streaming e-commerce are the key innovative elements that support a technological era to achieve a qualitative change.
In the field of large-scale models, the current commercialization is weak, precisely because the application scenarios of AI and the corresponding user scale are not yet mature.
This is what Li Yanhong said: Major technological breakthroughs and disruptive innovations are often the results of scalable applications, not the causes.
In the past year, the computing power bottleneck of large-scale models seems to have become a consensus in the AI industry. OpenAI has used Sora text-to-video to tease users' nerves several times, but the effect is mixed. This also proves that technology can no longer rely on "working in isolation" for value iteration. Only the verification of scalable applications can determine the next step of technological innovation.
On the other hand, Sora's decline from the altar and the decrease in OpenAI's market share also confirm Li Yanhong's judgment. In 2024, we will find that the penetration rate of AI has significantly increased in every sub-segment track, but the benefits generated at the model end can no longer be called "disruptive". The innovations that have accumulated for a long time have not naturally brought about a super application.
Instead, the persistence of technology companies in the penetration rate of AI has derived the jungle law of investing money. What is being compared is not only the gene of technological innovation, but also the determination of capital tug-of-war. Whether this path will work in the end is another matter, but in terms of the logical relationship between the underlying technology and AI applications, those who are building the wheels need to reflect seriously.
If it is not the cause, it can only be the result.
Therefore, the AI exploration and innovation emphasized by Li Yanhong are the retrieval-augmented generation technology iRAG that reduces the hallucination rate of image generation, the AI Coding tool Miaoda, and the growth of Baidu Wenku, rather than the increase in the scale of the large-scale model itself.
In 2025, large-scale models will naturally optimize and develop along with application scenarios, and there will no longer be so many artificial gods. The lost way of commercialization and the explosive growth of AI applications will speak for themselves.
In this process, resource allocation and a long-term mindset may widen the gap between top players:
First, it is still the construction of the AI infrastructure layer. In addition to the Qianfan Large-scale Model Platform, which has already helped customers fine-tune 33,000 models and develop 770,000 enterprise applications, Baidu has also chosen a path to popularize programming to a wider population - Miaoda, which also aligns with Li Yanhong's thinking on scalability.
Second, the courage to predict and fail.
Li Yanhong's statement that "we must bear a higher failure probability than our peers" is a summary of Baidu's past lessons and a "preview" of its continued exploration and trial and error in 2025.
02
How will AI applications "boom"? 40 million paying users find a way
Another judgment of Li Yanhong is that AI applications will show an explosive growth in 2025. While the user volume increases, the market may also usher in a blooming of diverse forms.
This trend is not difficult to imagine. First of all, the bullets of AI-enabled scenarios in various industries have already been fired. 2024 can be said to be the first year of AI applications. Various industries and sub-segment tracks of various AI capabilities have been integrated to create tens of thousands of intelligent applications. It took one year to achieve a quantitative change, and it is only a matter of time before the commercial fruits are harvested.
Then, if we are to find a benchmark track that takes the lead among these applications, who will it be?
Intelligent search, intelligent productivity, and intelligent education, which are most closely linked to AIGC, are highly anticipated.
In the domestic market, an enlightening perspective is that in the AI application era, the enterprise service field, which had an unsatisfactory business performance in the SaaS era, is very likely to rise again and become the "bridge" for AI to achieve scalable applications.
The growth in the invocation of large-scale models also reflects this trend. Data shows that the B-end invocation volume of Baidu's large-scale model has increased 30 times in the past year, reaching a magnitude of 1.5 billion. Currently, more than 60% of central and state-owned enterprises are using large-scale models, and Baidu is also the best performer in the large-scale model-related winning projects last year: According to public data, among 728 projects with a total amount of 1.71 billion yuan, Baidu ranked first with 40 winning projects and an amount of 274 million yuan.
It is only a matter of time for enterprises and technology service providers to jointly explore new value-added scenarios to generate long-term commercial potential for more and better invocation effects.
In contrast, in the C-end scenario, although the C-end is the best future solution for super applications, the application scenarios with the best commercial performance in the C-end currently are actually being driven by B-end needs.
Also in this all-staff letter, Li Yanhong disclosed for the first time that the number of paying users of Baidu Wenku has exceeded 40 million, with a 60% increase, and these users basically purchased the AI functions of Wenku.
Baidu Wenku's own positioning is a productivity platform for office documents, knowledge integration and other scenarios. Compared with products based on entertainment generation, Baidu Wenku is closer to B-end and C-end users based on B-end office needs. It is not surprising that it can achieve a miraculous natural growth.
Baidu, based on the productivity logic of the AI era, has prioritized the reconstruction of the Wenku product and reaped the first wave of AI commercialization dividends. This demonstrates the importance of making predictions about where AI applications will boom.
Looking at some less vertical general AI applications, even if they have a good user scale, popularity, and reputation, the lack of productivity value in the paying scenarios will lead to unclear commercialization paths and make it difficult for C-end consumers to buy into them.
With the AI MAU of Baidu Wenku reaching 70 million, the scalable space of the intelligent productivity track has been demonstrated. In Baidu's view, transforming and reconstructing existing star products and converting the user base of hundreds of millions into AI potential seems to be much more convincing than investing money in a Chatbot form that is difficult to penetrate into the low-end market.
03
In a protracted war, super applications are not accidental
This is an era where openness and contingency outweigh everything, and AI applications follow the same principle.
In the overseas market, ProductHunt showcases shiny AI applications every day. Entrepreneurs are imaginative, while large companies are more pragmatic and focused: OpenAI is optimistic about the physical world and embodied intelligence, Apple is constantly looking for ecological cooperation, and Google is also targeting AI applications.
In the domestic market, the primary market has significantly reduced its investment. Data from IT Juzi shows that the total financing amount in the AI field in 2024 is only about 80% of that in 2023. As a track with a not-so-high technical threshold, although new vertical tools enter the public eye every month in the AI application era, it is difficult for them to prove their long-term value.
The principle is similar for the unpopularity of large-scale model advertising. Billions of real money are invested in attracting new users and increasing monthly active users, which is a risky bet. The bet is on a super application that can rival WeChat, Baidu, Taobao, and Douyin.
The poor user retention and conversion rate in reality is not due to inadequate technology, but rather the lack of addictive points in the application scenarios like other super applications.
Even with the same search value, in the past, Baidu was able to become a household name and allow "more than half of Chinese people to obtain information using Baidu every month". Of course, it did not rely on a simple search box. Nowadays, as a new traffic entrance for AI, it has not yet found the balance between "authoritative answers" and commercial forms such as advertising.
Therefore, it is simple to make a useful AI application because large-scale models are inherently useful. But making a super application is a very patience-testing task, and in a protracted war with multiple layouts by large companies, it is difficult to become an accidental event.
Li Yanhong said in the all-staff letter, "The competition is more intense than ever."
Therefore, doing better than competitors in any track, and breaking down professional capabilities into application forms that the general public can use, has become Baidu's consistent strategy. Therefore, the audience of Miaoda is ordinary people, rather than assisting programmers in writing code like the mainstream products in the AI Coding track.
Where will the AI-native super applications appear in the future?
Still, it will appear in the scenarios where ordinary people have the strongest daily dependence and perception.
Baidu Wenku, which now has an AI MAU of over 70 million, will hand in a new answer this year on how far it is from becoming a super application in the intelligent productivity field. In the second half of the year, Baidu will also release the 5.0 version of the Wenxin Large-scale Model, making the model capability an infinite "0", and the application scenario the foremost "1".
"Survive the cold winter and become a new hero."
The long-distance running ability in the field of artificial intelligence has always been the gene for this enterprise to emerge with innovations.
