🎉 The #CandyDrop Futures Challenge is live — join now to share a 6 BTC prize pool!
📢 Post your futures trading experience on Gate Square with the event hashtag — $25 × 20 rewards are waiting!
🎁 $500 in futures trial vouchers up for grabs — 20 standout posts will win!
📅 Event Period: August 1, 2025, 15:00 – August 15, 2025, 19:00 (UTC+8)
👉 Event Link: https://www.gate.com/candy-drop/detail/BTC-98
Dare to trade. Dare to win.
Can generative AI generate the future of Baidu?
Original source: Light Cone Intelligence
Author: Liu Yuqi
The wind begins at the end of Qingping, and the accidents and inevitabilities of fate are often intertwined.
2019 is the darkest year in the entire history of artificial intelligence. Following AlphaGo's defeat of Lee Sedol in 2016, the emergence of the Tansformer architecture in 2017, which led to a large-scale explosion of technology, and the entrepreneurial boom in 2018. In 2019, the ebb of capital, the technical bottleneck period, and the difficulty of landing scenes opened a "chaotic era" in the history of artificial intelligence.
No one knows when general AI will arrive, just as a trisolaran doesn't know when the sun will rise. **
That year, a large number of AI companies fell into layoffs, broken capital chains, and difficult production of products, and there were few left to persist.
But it was also in 2019 that a new hope for AI began to be conceived: OpenAI accepted Microsoft's investment and deep cooperation with it in July, and it was from that year that GPT-3, which is now shocking to the world, began to be developed; Baidu, the first to establish an AI strategy in China, has carried out a half-year personnel adjustment since the beginning of the year, and now it seems that the rearrangement of troops seems to have opened a four-year period of perseverance.
In 2019, Wang Haifeng was promoted to Group CTO and continued to serve as the general head of AI Technology Platform System (AIG) and Basic Technology System (TG), which are Baidu's most important technical foundations; Shen was promoted to Senior Vice President, responsible for the mobile ecosystem business group, and then transferred to Baidu Intelligent Cloud in 2022, becoming the pioneer of Baidu's second growth curve. Among the new forces introduced in 2019, He Junjie is the only post-80s Baidu vice president, first responsible for investment mergers and acquisitions and strategic investment, and then reused, responsible for Shen Jing's vacant mobile ecological business group, with more real power.
The time has come to 2023, and the ability of Transformer has finally broken the ceiling by OpenAI in this year, advancing to a new level, and the breakthrough of technology has made OpenAI the most watched company in the world. Microsoft overshadowed Google's technological brilliance.
**In the past four years, Baidu's core business has formed a triangular structure led by Shen Tian intelligent cloud, He Junjie is responsible for the mobile ecology, Wang Haifeng pressing array technology, and Baidu has finally ended the quiet period and began to turn defense into attack, and the action is continuous. **
"Do every application again" is the first bright sword after half a year of accumulating power, and now Baidu is like a planet, about to encounter a huge "technological gravitational field" and completely change its orbit.
Eve of the emergence of large models
In 2017, Google proposed the Transformer architecture in a paper called "Attention Is All You Need" to replace the traditional RNN and CNN loop models, which showed that the training accuracy of the Transformer is higher than all previous models, the training time is significantly lower than the previous model, and the training effect is also good when the training set content is small.
Since then, the Transformer architecture has been quickly accepted and applied in the field of NLP and CV, especially in the field of NLP, where the attention mechanism allows machines to understand semantics more accurately and generate them, as well as reduce information duplication.
As one of the first companies to discover and follow up the Transformer technology route, Baidu has been researching in the NLP field for more than ten years, and has formed a lean NLP team formed by top experts such as Wang Haifeng, Wu Tian, Wu Hua and so on.
At that time, the first important task of the NLP team was to build Baidu translation. Wang Haifeng boldly applies deep learning and neural networks to translation to enhance the machine's understanding of context and form a smoother translation. This bold attempt also made Baidu Translate the first translation system to support more than 200 languages, even a year and 3 months before Google.
**This is also the prototype of the understanding ability in the four core capabilities of the big model "understanding, generation, logic, and memory". **
But Wang Haifeng feels that NLP is not enough. He specially went to Li Yanhong's office to report and proposed the next "landed Normandy" - speech recognition. His reasoning is that speech recognition technology is about to reach the critical point of industrialization, and once it breaks through, it will soon be commercialized on a large scale.
With his own judgment on technology, Li Yanhong decided to support Wang Haifeng again, and successively established the "speech recognition department", "image recognition department" and "knowledge graph department" parallel to the NLP department.
Wang Haifeng's operation made many Baidu students puzzled, "These technologies have nothing to do with Baidu's current products, is it to save up for the New Year to create so many departments in one brain?" ”
At that time, Wang Haifeng saw that search data has a very strong support for the logic of large model formation. "Baidu has the world's largest search engine, which not only has a strong timeliness of information, but also has a high accuracy rate, which can build the most complete knowledge graph," Wang explained in a public interview. **
Before creating Wenxin, Baidu precipitated a multi-heterogeneous super-large-scale knowledge graph with more than 5 billion entities and 55 billion facts, and was able to obtain a unified understanding of the world through language, hearing, vision, etc. Its knowledge-enhancing model released in 2021, ERNIE (Wenxin) 3.0, is the predecessor of Wenxin Yiyan, and the project is mainly responsible for Wu Tian.
In 2019 again, Baidu Wenxin 1.0 was released, and 3 versions were iterated in 4 years. In November 2022, Wu Tian simultaneously announced 11 industry models that Wenxin has accumulated at the public summit, covering electricity, gas, finance, aerospace and other fields, and the industrial ecology has initially formed.
** These have all foreshadowed and paved the way for Baidu to preemptively release Wen Xin's words. On March 16, after Wen Xin's words were released, Baidu returned to the spotlight again, but more than affirmed, it was doubtful. **
"In the case of such strong market demand, it is still very meaningful who makes it first", Li Yanhong once said in an interview, even if the product is not fully mature, but still have to be released: "After the release of Wenxin's words, countless people, people who have not been in contact with me in the past or who are far away from me in the industry are asking, how can we cooperate with Baidu, how to try it as soon as possible."
**From a technical point of view, large models are high-speed iterations of "more and more used". ** "During the internal test, the employee asked, how can Wen Xin hide his head and write a poem badly? I said wait, I guess I can learn it the day after tomorrow, and it will be able to be used the next day, and the progress of the big model is also a continuous learning process," Wang Haifeng said with a smile.
In half a year, Wenxin Yiyan has carried out three iterations from 3.0 to 3.5 and then to 4.0 version, according to Wang Haifeng at the meeting, the scale of Wenxin Yiyan users has reached 45 million, 54,000 developers, 4,300 scenarios, 825 applications, and more than 500 plug-ins.
The water watered in NLP, the fertilizer applied, finally ushered in a bumper harvest in 2023, and as Li Yanhong said, the path of technological development is the process of "two lives, two lives three, three lives and all things".
Internal Strength Behind Big Models
After the outbreak of general artificial intelligence, the attention of cloud computing and the attention of enterprises reached the peak, and it also came to the "iPhone moment".
The emergence of large models, resulting in a huge computing power gap, cloud computing not only for the large model to provide cloud computing power support, but also the best landing point for large model landing enterprises, whether it is Baidu or any enterprise with a large model, when the large model comes out, the next focus is to promote to the market, let enterprises use.
**For Baidu, such a burden falls on Shen Shu's shoulders. **
Among Baidu's six business group leaders, except for CTO Wang Haifeng, Shen is the only senior vice president. Although they both come from a technical background, unlike Wang Haifeng's "engineer" role, Shen has been responsible for the growth of key businesses since he joined Baidu.
In the 10 years since joining Baidu, Shen has successively integrated the advertising system and improved the system's monetization ability; Combine search and feed streams; It integrates the mobile ecosystem business group upgraded by search to complete the territory of Baidu's mobile ecosystem.
** If Wang Haifeng created a sharp knife, then Shen Shu is a pioneer official who can use this sharp knife to open up territory, in the words of Li Yanhong, "dare to fight a tough battle and can win a battle." **
In May 2022, Shen Jixing was rotated as the president of Baidu Intelligent Cloud (ACG) Business Group. For Baidu, it did not bet fully on cloud computing at the beginning, but it was precisely with the continuous development of artificial intelligence technology that Baidu realized the shortcomings of cloud computing and began to secretly cultivate the second growth curve.
**Due to the lack of first-mover advantage, Baidu Cloud's goal in the past 10 years is very clear, not to compete with the "old guns" in the scale of the IaaS layer, but through the combination of PaaS + SaaS and intelligent capabilities, to play differentiation, and to cut into enterprise digitalization in small battles. **
At the 2023 Baidu World Conference, Shen once again proposed the "cloud-intelligence integration" strategy: "The deep combination of artificial intelligence and cloud computing is the key for enterprises to quickly implement AI native applications. At present, all applications and services of Baidu Group are running on Baidu Intelligent Cloud based on the 'Cloud-Intelligence Integration' technology architecture."
In the five months after taking over ACG, Shen quickly integrated the "big model service super factory" - Wen Xin Qianfan in response to the core needs of enterprises in the era of large models, and divided users into five categories of users according to their needs.
First of all, in view of the demand gap caused by computing power resources, the Qianfan platform provides various types of heterogeneous computing power. For example, in the most expensive training link, through distributed parallel training and microsecond-level interconnection capabilities, Qianfan platform can achieve a training acceleration ratio of 95% and an effective training time ratio of 96%, greatly reducing customer computing power and time costs.
Secondly, at the model level, for customers who want to directly call existing large models, enterprises can quickly call multiple large models, including Wen Xin Yiyan, while Qianfan platform provides tools such as Chinese enhancement, performance enhancement, and context enhancement. According to Shen Ji, the Qianfan platform has served more than 17,000 customers.
**For customers with secondary development needs, Qianfan platform provides a full life cycle tool chain such as retraining, fine-tuning, evaluation, and deployment for large models, with the industry's largest number of 41 high-quality industry datasets, and quickly optimizes them for their own business scenarios.
The conference also carried out a practical demonstration of how to quickly develop knowledge Q&A applications for Sany Heavy Industry based on the retrieval enhancement generation (RAG, Retri Augmented Generation) framework: just select the preset RAG framework in the Qianfan AI native application workbench and configure the corresponding parameters to quickly realize the development and launch of the intelligent customer service application on the official website of Sany Heavy Industry.
**Shen said that building such a "small assistant", even if it needs to process thousands of thousands of words long documents, the cost is only a few hundred yuan; After that, each consultation of the user only costs a few cents. **
For a long time, large-scale industry, manufacturing, and agriculture have been deep-water areas of digitalization, and the core reason is that the complexity of the industry has led to a high threshold for digitalization and is difficult to land.
However, through the large model, not only the threshold of use is reduced, but also the cost of use is reduced. There is no need to build any new system, nor does it require manual participation, it is a more advanced application method of technical components, ** the combination of the two, but also turn the cloud-intelligence flywheel, gradually accumulating. **
Innovation Challenges for Large Models
The past is a foregone conclusion, but the future can change.
All vendors with large models have found opportunities in the application layer. Microsoft began to work on the full line of products including Bing, Office, and Windows systems in March, and Ali Daniel Zhang said: "We must use the big model to redo all the products", but this sentence, simple to say, is the biggest innovation challenge in the era of large models.
How does AI refactor applications? This requires not only business ability, but also imagination, and in the face of a new AI era, Baidu has also put the baton in the hands of young people. **
After May 2022, Shen Jie was succeeded as the head of the MEG business group by He Junjie, the vice president of the post-80s generation introduced in Baidu's talent echelon construction plan. If Shen Zhan is a "hard war faction", then He Junjie is a veritable "young strong faction". As Ren Zhengfei said, it is necessary to "let those who hear the cannon command the battle."
** Refactoring the application is not broken or standing, in contrast, Baidu is indeed "particularly ruthless" to itself. **
At the Baidu World Conference, focusing on "ecology", He Junjie handed over the answers of "mobile ecology", "content ecology" and "business ecology".
Among them, the mobile ecosystem covers AI native applications such as "New Search", "New Wenku", Wenxin Yiyan APP, and Baidu e-commerce "Huibexing"; At the content ecological level, a series of applications such as Baidu APP "AI Editor" empower the creator ecosystem; At the level of business ecology, the AI Native marketing platform "Light" was launched. In addition, He Junjie also announced the Wen Xin Yiyan plug-in ecology - "Spirit Matrix", which is now fully open.
**The new search defined by Baidu is exactly the logic of box calculation proposed by Robin Li in 2010. **It has three characteristics: ultimate satisfaction, recommendation stimulation and multi-round interaction. That is, when users search for questions, "no longer give you a bunch of links", but through the understanding of the content, generate multimodal answers such as text, pictures, and dynamic charts; Recommendation stimulation can recommend the problems that users care about in real time; In response to complex needs, multi-round interaction can meet the personalized search needs of users through prompts, adjustments, etc.
AIGC's capabilities have given new vitality to some of Baidu's old applications, such as Baidu documents upgraded from content retrieval tools to content production tools, and Baidu Editor has become a content generation tool; The other part also explores new scenarios, such as Baidu Diager's one-stop generation of marketing content through AIGC and intelligent delivery, combined with the digital human generation platform "Huaicast Star" to help merchants expand their marketing scope and scenarios.
**At the same time, with the gradual landing of large model applications, Baidu also realized that relying on its own strength is ultimately limited, and the infinite is vertical and horizontal. **
This is the value of the Spirit Matrix, a platform that greatly reduces the cost of large-model plug-in development, allowing ordinary people with creativity and ideas to become plug-in developers. Li Yanhong said that the plug-in is a special AI native application, and it is also the AI native application with the lowest threshold and the easiest to get started.
The feature of the plug-in is the "universal interface", which can connect search, mini programs, content platforms or any entrance, so that the use of "plug and play" allows developers and creators to quickly join the ecosystem.
He Junjie revealed that one month after its launch, Lingjing Matrix has received 27,000 developer registration applications, covering more than 20 vertical fields, including enterprises, institutions and individual developers.
A Baidu insider told Light Cone Intelligence: "The large model plug-in of the application layer takes Lingjing as the main platform and will be placed in Wenxin Yiyan and Baidu App. Qianfan is more at the bottom level, and the spiritual realm is more upper, and it is even possible to replace Qianfan at the level of application plug-ins in the future."
Conclusion
In 2016, Li said Baidu was only 30 days away from bankruptcy. "The dinosaur stepped on a scoop on his foot, and it took hours for his brain to react. So no matter how big dinosaurs grow, they will go extinct."
Baidu doesn't want to be a dinosaur, and at the level of consciousness, it always thinks ten steps away.
Fortunately, Baidu waited for the new era and got through the hardest moments; Unfortunately, at the beginning of this new era, any painstaking snatching will seem insignificant under the long competition.
But getting a new ticket is at least a new beginning.
Reference:
Cross-border experts in various fields of artificial intelligence - Transformer"
The trip is far: they sculpt souls for artificial intelligence