Image source: Generated by Unbounded AI

Making great strides towards 2025, the competition on the AI ​​​​track has not become easier.

Technological research and development has not hit a bottleneck as expected, but the speed of progress has not slowed down at all. From Sora of Wensheng Video at the beginning of the year to O1 of Slow Thinking, they have inspired people in the industry to think about the AI ​​large model paradigm.

In December, OpenAI announced Sora Turbo and O3 on the release day, which made people say that they “couldn’t catch up”.

On the product side, AI products seem to have rushed into the "old path" of past Internet products: large-scale purchases, publicity, and burning money to expand the user pool.

Money has indeed been burned, and the number of users has indeed piled up, but the commercialization path still seems to be lagging behind. Products with large MAU have actually become "ICU products" - if money for purchasing and marketing is cut off, user activity is at risk of plummeting.

With billions of real money spent on marketing, how much of it is “false prosperity”?

In front of us are cutting-edge technologies that are rapidly iterating, and behind them are the money-burning traps of "high-speed growth." In the new year, it seems that there are still big winds and waves on the path of artificial intelligence.

Wherever there are storms, there must be opportunities. Will 2025 be another “big year” for AI?

01. How long will the "blowout" AI application last? Where does it grow?

On January 1, 2025, Baidu leader Robin Li issued an all-staff letter on the company's 25th anniversary. In the letter, Robin Li pointed out that AI native applications are rapidly becoming popular in all walks of life, and the new year is another year of explosive growth in AI applications.

In fact, this trend is quite obvious. The basic large model has been rolled out for several years. Many companies - especially startups - have been tightly tied by Scaling Law and have gradually declined. A card table that "makes wheels."

For rolled products, as mentioned above, if you "bet" on the purchase volume, the business model is not healthy. A good AI product must have a business model that is feasible enough to be implemented, so as not to become "losing money and making money".

At present, many large model products are still mainly free to use, and this has attracted a wave of users who are early adopters. For any technology product, this kind of user is tantamount to a double-edged sword.

The "early adopters" have indeed brought better data to the product, and also brought real trial feedback to the product in the short term;

But in the long term, it is necessary to Maintaining these "early adopters" is not an easy task. Once commercialized things are introduced, many of them will leave, not to mention some of them "foolishers" who are not interested in the product.Long -term stable operation is a challenge.

The logic behind users' willingness to pay for AI native applications is simpler and more direct - the product can actually solve the user's needs in life and work, and the willingness to use and retain is high enough to support them. Spend real money for it, this is the foundation of AI products.

Because of this, in the current environment, AI paid products with strong anti -risk and long -term operating capabilities may be more worthy of research.

In the large market, Baidu Library is one of the most thorough AI original products. So far, more than 40 million people have become its paid users, second only to Microsoft's Copilot.

Data show that after the AI ​​reconstruction, the MAU of Baidu Library reached 70 million, more than 40 million paid users, accounting for about 60 % of its monthly active volume.

If, at this time a year ago, the industry still had doubts about "whether users would pay for AI tools," Wenku's results today are a powerful counterattack.

In terms of business model, advertising is still a very important part of Baidu Wenku’s original profit model. After the AI ​​reconstruction, as a "better way to make money", Baidu Wenku is determined to clean up advertisements, give users a purer experience, and make money based on their true skills.

This "real skill" refers to the support of Wenxin's large model, the accumulation of 1.4 billion authoritative professional documents, and the one-stop product design that solves intelligent production, allowing users to actually earn money through the library Money has become a new product of people.

Today, as the commercialization pioneer of AI native applications in the market, Wenku has completely got rid of the stereotype of advertising profit. A better experience has also brought 150% annual DAU of Baidu Wenku AI function. Year-on-year growth, the cumulative number of uses exceeds 2.8 billion, occupying 80% of the smart PPT market share.

Slow and steady can actually generate value for users, and then create their own blood. Wenku has set a good example for many AI products on the market.

02. Large-scale application is the basis for further AI

< p>However, one flower blooming alone is not spring, but a hundred flowers blooming together fills the garden. AI natively focuses on ecology. In 2024, this can be said to be the consensus of the industry.

Among the many big model players, Baidu was the first to propose "not to make 'super applications', but to help more people make 'super useful' applications", which is a clear-cut statement.

Some people have questioned that Baidu does not promote "super applications", perhaps because it will "cut off an arm" in the increasingly competitive market in the future and disperse its competitive advantages.

In the letter of all staff,Li Yanhong explained the deep insights that Baidu did not do "super applications": "Major technical breakthroughs, disruptive innovation is often the result of large -scale applications, not the reason."

Application is always technological innovation Dynamics and home. In the history of science and technology, there are aircraft to study air dynamics; China's high -speed rail operating mileage is number one in the world, and people have more power research and iterative high -speed rail technology, and then output to the world -without application, there is no technology, there is no technology, there is no technology, there is no technology Further evolution and iteration.

For a long time, Baidu has been using "application -driven innovation". Facing the fantasy problem in the actual use of Wen Shengtu, Baidu has developed image -based retrieval enhancement technology (IRAG). When the things generated by others are "wrong", Baidu earlier guarantees that the pictures they generate are more available. reliability.

Driven innovation with application is "scale".

In the letter of all members, Li Yanhong mentioned: "Without hundreds of millions of operating kilometers, unmanned driving cannot be ten times more safer than someone driving." It can never be found that there are countless problems derived from the complex road conditions during driving, and then they cannot apply the right medicine.

Today, the "super application" of a single point such as Wen Xiaoyan, Bean Bags, Kimi, etc. still cannot meet the large -scale use of technical evolution requirements. This is also Baidu's ambitions. The reason for being complacent - Only when more and more people start to use Baidu's technology can the technology complete the upgrade to the next level.

Today, the opportunity for large-scale application of large models may first lie on the B side. From the perspective of statistical data, in the past 11 months, the winning bid project of the large model has increased by 2.6 times compared with the same period in 2023 to 728.

In 2024, Baidu ranked all manufacturers with 40 winning bid projects and 274 million yuan of winning bids.

In terms of division, in the financial field, Baidu won 14, and the winning bid amount is 3734.4 10,000 yuan, also leading large model manufacturers; in the smart terminal industry, including Samsung, Honor, VIVO, OPPO, Xiaomi and other more than half of mobile phone manufacturers, they are using the big model of Wenxin; and in the automotive field, more than ten car companies are all in the automotive field. It is a customer of Baidu Wenxin's big model. Among them, there are many mainstream manufacturers such as SAIC Volkswagen, Geely Automobile, Weilai Automobile, and Changan Automobile.

With the help of Baidu Wenxin's "family bucket", more manufacturers can refine the model that suits them, develop more suitable applications, and naturally there are more people according to their own needs. If you are willing to choose a better model and tool chain, Baidu's technology can be developed and iterated at high speed in this "application-technology" flywheel.

The use of the use of the use of the use of the use of the use of large B enterprises, so that more people -especially ordinary people who do not understand the code -master the ability of application development, is the real model reallyThe key to "flying into ordinary people's homes".

Seeing this opportunity, code-assisted generation tools have become an absolute favorite in the field of large models.

Although Baidu’s “Wenxin Quick Code” is also competing on the track, the “Miaoda” released at the World Conference last year extended its olive branch to complete coding “amateurs”. Let everyone have the ability to develop AI native applications.

When everyone, regardless of age, background, or education, can develop their own App in a very short time as they wish, then we may no longer be far away from the ultimate dream of AGI.

On the road of "application first", many predecessors actually stepped on the pitfall - when a technology company got lost in the eyes of customers and investors, worked hard on projects for the sake of revenue, and forgot about technology. Due to the original intention of development, their movements will inevitably be deformed unconsciously.

In the eyes of ordinary people, people start companies and businesses all to make money. If you can make money by doing projects, why should you take the pressure, take risks, and embark on the thankless path of "doing technology"?

Even the most insightful investors do not always favor technology companies. For example, Buffett has publicly stated on many occasions that technology companies are not his "food" - the reason is simply that the risks are too high. , not robust enough.

Working in technology is "dangerous" and even more "lonely". You have to walk alone amidst uncertainty, and you have to face the glare of others in the trough.

And "take greater risks", "continuously try and make mistakes", or even "quickly adjust the direction and start again if you go wrong" are not only the requirements that Robin Li put forward for himself on Baidu's 25th anniversary, but also At present, many technology companies need words of encouragement.

03. Conclusion

In a On the "eve" before the completion of a giant ecosystem, platform opportunities will always be the "crown jewel" - whoever has a booming AI native application will have the ability to transcend the success or failure of individual products and rely on the big market to make money.

An important limitation behind the explosion of large-scale applications is that the costs of API calls, model training, and inference remain high.

Behind the difficulties, there must be opportunities hidden. Who can improve the capabilities of the basic large model by another step earlier, improve the efficiency of computing power utilization by another step, and improve the collaboration of data by another step. , whoever is more likely to attract more AI native application developers and customers will take the lead in future competition.

At present, there are big companies that have the ability to continue to make iterative innovations in large-scale models, schedule 100,000 cards of computing power, and collaborate with enterprises on data. However, Baidu, Byte, and Alibaba are a preliminary BBA "Three Kingdoms". "The situation has begun to take shape.

In this three-way fight for hegemony, since no one wants to step down from the card table,You have to keep rolling. The new version of Baidu Wenxin’s big model is likely to be launched in the second half of the year, and Byte and Alibaba are also following up to varying degrees.

What new developments will there be in this fierce battle of large models in the new year?

Zhuangzai, 2025.