Wenxin's big model raises new "sails", and the tide of industrial application has arrived

At the 2022 Baidu World Conference, Baidu CTO Wang Haifeng showed such a scene in his speech: as the scroll slowly expanded, a "Fuchun Mountain Residence Map" after a "completion" appeared on the screen behind him. And borrowed the "tradition" of the ancients who liked inscriptions, and titled a small poem on the scroll.

As many people are well known, because of the various reasons in history, the "Fuchun Mountain Residence Map" of the famous famous master of the Yuan Dynasty was two, and the "front paragraph" was named "Left Mountain Picture" by future generations. Known as "Utilized Teacher". The two -stage scroll was exhibited in the Taipei Forbidden City, and the lacking part became a regret of nearly 400 years.

What many people do not know is that those who "make up" Huang Gongwang are not masterpieces who have studied ancient paintings for decades, but have a big model of Chinese traditional literati.

01 regrets for nearly 400 years, which is satisfied in a second

During the 2022 Baidu World Conference, Baidu's H5 "AI Painting Connection Love" launched the screen in many people's circle of friends. You only need to simply outline a few strokes in the blank space in the middle, and AI can automatically "complete" "Fu Chun" "Fu Chun". The contents of the incomplete part of the Mountain Residence Map, less than a second of time before and after.

For such a gameplay, some people may not be unfamiliar: such as AI restoration of old photos, "living" in ancient paintings, "painting masters", etc. It is difficult to distinguish whether a work is made by humans or AI.

However, the "Fuchun Mountain Residence Map", which is "completion" of the big model of Wenxin, is very different from the past. To be precise, there are three strict challenges in three aspects:

The first is high definition. Many AI repair paintings are often not clear. Although they can imitate the style of the original work, the details cannot withstand the scrutiny. Wenxin's big model is a high -definition size of "Fuchun Mountain Residence". Every line of lines that outlines, every tree is clearly visible.

The second is high reduction. Huang Gongwang, who ranked first in the "Four Fields", created the painting method of "shallow landscape", but there were not many works handed down. Under the condition of only one "Fuchun Mountain Residence Map", it is still necessary to highly restore the works of a generation of calligraphy and painting masters, which requires extremely high requirements for the migration learning ability of the model.

The third is the participation of the people. Different users have different lines, and some may be to draw an irregular circle on the screen. The large model of Wenxin can be generated in real time and highly restores it. idea.

Judging from the final result, the creation of the large model of Wenxin has been highly evaluated by Mao Chuanyu, the head of the Huanggong Wang Memorial Museum: "consistent with the existing true style style, so that the two sides of the strait scrolls will go out of the museum. It is in line with the aesthetic characteristics of the original painting, and it also has the value of ornamental. "

What needs to be considered is why Baidu CTO Wang Haifeng shows the "Fuchun Mountain Residence Map" after the annual Baidu World Conference? The answer is by no means purely for "showing skills", even if Wen Xin's big model has such technical strength.

As early as the 2019 Spring Festival Evening of 2019, Baidu AI showed the "looking at the face" to the Spring Festival couplet. Users can automatically generate the exclusive Spring Festival couplet through the way of writing words and even the face; by early 2020, Baidu Research Institute was launched online Intelligent video synthesis platform vidpress, one -click input graphic content can intelligently generate short videos; in May 2022, Wenxin · Laogu Community was officially launched. The official positioning is the creative community of Wenxin big model ...

If these clues are linked slightly, the big model of Wenxin may be aimed at AIGC, that is, the use of artificial intelligence technology to independently produce content. Li Yanhong, the founder, chairman and chief executive officer of Baidu also mentioned in the speech of Baidu World that AIGC is a new content production method after UGC and PGC. Creative content of unique value and independent perspective.

02 can make up for the complete paintings, but also detect the deep space

In the live broadcast of the 2022 Baidu World Conference, many netizens left a message in a barrage, and called the text of the poem painting and painting. However, the big model of Wenxin is not only the gene of the literary and artistic fan, but also a "science and engineering man" with spiritual research.

Back to December 2021, Baidu officially became the "China Lunar Monthly Aerospace Engineering Artificial Intelligence Global Strategic Partner". According to the information disclosed at the time, the two parties will carry out related cooperation between aerospace technology and artificial intelligence technology in the field of deep space detection, including the moon exploration and planet detection, etc. The curiosity of the outside world is also born: What are the space for artificial intelligence and deep space detection?

At the 2022 Baidu World Conference, the answer was officially announced: Baidu and Aerospace cooperated to develop the world's first big model in the world-Aerospace-Baidu · Wenxin big model. The integration of data and knowledge, intelligent collection, analysis and understanding of aerospace data, and help breakthroughs in technical breakthroughs in deep space intelligent perception, planning and control.

To understand the application potential of Wenxin's big model in the field of aerospace, it seems that it is necessary to sort out the importance of artificial intelligence during aerospace exploration:

The first is the data level. Exploring the broad space that is countless times larger than the earth, not only has the characteristics of large amounts of data, many types, and high requirements for general big data, but also requires real -time processing of data. If there is no timely and accurate data analysis, even if it is a decimal error, it may affect the success or failure of the overall situation. Followed by the application level. In the past, the processing of aerospace data was mainly manual analysis. Due to the proportional cycle and distance transmission of spacecraft data, relying on the ground to make data judgment, it can no longer meet the needs of the fast state monitoring and fault diagnosis of spacecraft. Spacers' ability to monitor and self -diagnosis.

Once again, the value level. The application of aerospace big data should not stop data analysis. Calling, storing and learning of previous flight data can deeply simulate the operating status of the spacecraft, provide a basis for correction control and improvement design, and continue to nurture the design and manufacturing of spacecraft for continuously nurturing the spacecraft , Security and other services to improve the overall efficiency of aerospace engineering.

The logic of the entry logic of the large model of Wenxin is that the aerospace flight task plan is a typical knowledge processing process. The large model of 100 billion parameters can not only optimize the flight track and posture control in real time, to avoid the risks during flight, but also help help. Aircraft independently plan and detect routes, and choose to detect targets. Or it is the process of integrating existing massive data and resources to accelerate the process of space intelligent perception, space intelligent planning, and space interaction of space.

According to Liu Jizhong, deputy commander of China ’s lunar exploration engineering, Liu Jizhong, director of the National Space Administration and Director of the Aerospace Engineering Center, the country has developed a plan for the next ten years of deep space detection, involving the lunar exploration project and planet detection project, which will be ten in ten During the year, important tasks such as Mars sampling, asteroid sampling, and Jupiter's surround detection. The large -scale model of the core link of the aerospace industry will also follow the pace of national deep space detection towards the sea of ​​stars.

03 Deep cultivation of the industry just needs, release technical dividends

Whether it is unique in content creation or the separation of the aerospace field, the big model of Wenxin shows the ambition of industrial landing. In the environment of some large models that are blindly chasing model parameters and believed in "strong miracles", it has taken a new path of industrialization.

The underlying logic is inseparable from the characteristics of the "big model" itself. Traditional model training requires a lot of data marking to train sufficient accurate models, but the pre -training large models use self -supervision learning methods, which can refine and learn the laws and knowledge in massive labels. When facing tasks and scene applications, only a small amount of tasks are required to mark the data, and a model that is very useful in the application scenario can be obtained by fine -tuning.

At the same time, it is inseparable from the independent thinking of Baidu AI. The pre -training big model broke the bottleneck of data labeling and is increasingly friendly to the industrial landing. Wu Tian, ​​vice president of Baidu Group, deputy director of deep learning technology and application of the National Engineering Research Center, made a different from other manufacturers in the industry. Judgment: "This year is the key year of the large model. The big model has gone through the exploration and breakthrough period of the previous years, and has reached the promotion period to a certain extent."

Reflected to the specific landing link, the large model of Wenxin aimed at the industry's just need.

For example, the completion of the residual volume of "Fuchun Mountain Residence" corresponds to the gap in content production, and the cross -modal comprehensive technical ability of the large model of Wenxin has opened a new way to reduce content production costs and enhance the diversity of content; then For example, the industry's large models launched on the Wave Summit 2022 excavate industry data on the basis of universal models, and further enhance the adaptability of the industry's application through "enhancement of industry knowledge".

Clarifying the idea of ​​large models also needs to solve another problem, that is, supporting tools and platforms, and then comprehensively reduce the application threshold of large models. Corresponding to the development kit, API, and the EasyDL and BML development platforms built into the capacity of Wenxin big models.

There are many cases that can be found.

During the college entrance examination in 2022, the AI ​​digital man blessed by Wenxin's model, Xiaoxiao, challenged the college entrance examination composition, wrote a composition entitled "Bettings, Moly Hands," Essence According to the situation over the years, Du Xiaoxiao's composition score more than 75 % of the college entrance examination candidates.

The State Grid Model of the State Grid jointly developed by the State Grid and Baidu-Baidu · Wenxin Model, digging the power industry data in massive data, introducing sample data and unique knowledge accumulated by power business, and combining both parties in the training in the pre-training algorithm and power field The experience of business and algorithms has significantly improved the application effect in the State Grid scene.

According to similar ideas, Baidu and SPD Bank also jointly developed a big model of the financial industry. Pudong Development · Baidu-Wenxin Model combined with the accumulated industry data and knowledge accumulated by Pudong Development scenes. The technical and business experts of the two parties designed a targeted financial report field judgment and financial customer service Q & A matching tasks to allow Wenxin's big model to learn finance The knowledge of the industry has improved the effectiveness of the typical tasks of SPD.

It is a case of industrial application that allows the outside world to be reconstructed by the outside world: it is no longer an unattainable hall -level technology, but a creative tool that the general public can use it; At the point, solve the core issues of the industry; the mission of the big model is to reduce the technical threshold and let AI go to universal technology ...

The industrial application scene of Wenxin's big model has also become clearer. The gap between cutting -edge technology and application scenarios is now allowing technology to generate real value in landing applications and release technical dividends to Qianxing Baiye.04 Write at the end

It can be seen that Baidu's construction ideas for the large model of Wenxin are the same as flying paddle. They are to accelerate the landing of artificial intelligence and practice the mission of industrial intelligence.

Up to now, the Weixin Model has released more than 20 large models, covering the basic models, task models and industry models. It has built a complete large model ecosystem.Enter the scenario of industrial, energy, finance, communication, media, education, etc.

New skills such as completing famous paintings and detection of deep space are not only a milestone in the intelligent process of Baidu's AI industry, but also further interpreted the original intention and perseverance of industrialization.

- END -