Big model of Hexia Dream, Baidu will do the test field by yourself

He has a cool dream, which is an extremely familiar word today. It comes from Academician Yuan Longping's ideal pursuit of high -yield hybrid rice. It is Yuan Longping's Chinese dream. He mentioned many times during his lifetime that he dreamed that he had a coolness, and the rice was as high as the sorghum and the grains were as large as peanuts.

Perhaps, Qianxing Baiye has such a "Hexia Dream". Science and technology exploration will bring an inestimable value for production and creation, and greatly changes people's lives. In the field of AI, there is a technology that often reminds me of "Hexia Dream of Liang Liang", which is a pre -training big model. In the dream of Academician Yuan Longping, the hybrid rice is very tall and brings extremely high acres of output; the big model is also very large in training data and model scale, forming a huge advantage in model capabilities.

Since 2018, large models have been widely concerned as an important technical direction in the AI ​​field. Over the years, our biggest expectations for the big model may be that it can truly become a piece of paddy fields that can really become an industry and inclusive economy from laboratory and data concentration.

The development of hybrid rice is inseparable from the test field of careful cultivation. To achieve a big model, "Hexia Dream of Liang Liang" may also need such a "test field" to come forward to the direction of the advancement of the big model of the times -at this juncture, Baidu stood out.

On September 23, the 20.22 million Baidu Mobile Ecological Conference was held in Zhuhai, Guangdong. At the meeting, He Junjie, senior vice president of Baidu Group and the general manager of Baidu Mobile Ecological Division (MEG), for the first time to comprehensively interpret Baidu's mobile ecological strategy. One of the very noteworthy points is that the Baidu mobile ecosystem we are familiar with will be displayed in detail. How is the deep combination of the large model of the Wenxin series created by Baidu, bringing the growth of Hemu in the fields of search, information flow, AIGC. In Baidu's mobile ecology, the ability of Wenxin's big model can be described as full, and it should be done.

From another perspective, if we want to know how to cool down the big model in the future, then today's Baidu mobile ecological scene coordinates may be pointing to the big model of smart China tomorrow.

In the big model 3.0 stage, the industry transforms into the core

Before discussing the importance of Baidu's mobile ecological integration, we must first understand the technical and industrial background behind this. The integration of Baidu Mobile Ecology and Large Models is important because the technology of large models itself is at a very important development node.

From the perspective of development, from the 2018 major modeling models, the development of the industry is valued, and the development of large models can be summarized into three stages. The first stage is the so -called "big refining model". The traditional depth model is already the main theme of the industry, and the pre -training model is a new technical idea, which has begun to enter the bud stage. In the second stage, the value of large models has been widely recognized by all sectors of industry and academics. With the widespread injection of capital and social attention, talking about AI's must -have models to become popular, all circles have also entered the transformation from the "big refinement model" to "big model". At this stage, various large models have emerged endlessly. The expansion of training data parameters and the development of technical ideas have brought the prosperity of large model technology.

When the technology research and development reserves of large models reached a certain stage, the new challenges also followed: how should so many models combine with industrial applications? If this problem cannot be properly resolved, the prosperity and development of the big model is ultimately the air tower. Therefore, the 3.0 stage of the large model is the transformation from the "big model" to "using the big model". How to achieve effective industrial transformation has become the core task in the current major model field.

If you want to make good use of the big model, from the perspective of technology to the industry's transformation, you need to have three prerequisites: the technology is too good, the platform and the tools are all available, and there are clear and clear scene value as a demonstration. Baidu, which has deeply cultivated Wenxin series, apparently has the leading advantage in the previous two capabilities.

Baidu has been developing the pre -training model in 2019, and has released the knowledge enhanced ERNIE series model. In December 2021, Baidu and Pengcheng Laboratory jointly released the world's first knowledge enhanced 100 billion model "Pengcheng-Baidu · Wenxin", with a parameter scale of 260 billion yuan, and achieved the best results in more than 60 NLP tasks. At the Wave Summit Deep Learning Developer Summit in May 2022, Baidu released 10 major models, including the knowledge of integrated learning task knowledge to enhance the 100 billion -dollar model, the visual model of the multi -tasking unified learning, cross -model big model, biological creature Calculate large models, industry models, etc., and propose three key paths that support the implementation of the large model industry: build a large model system that is more adaptable to the needs of scenarios, provide a full process of tools, platforms, and methods to support application landing. Construction stimulates innovative innovation Open ecology, etc.

At this point, the large model of Wenxin has not only included the basic universal models, as well as the industry's large models facing the key task field, and also has a wealth of tool platforms. From the perspective of high technology, the large model of Wenxin has the differentiated advantages of knowledge enhancement. It can complete learning from large -scale knowledge maps and massive structure data, thereby achieving higher model efficiency, better effect, good explanatory explanatoryness Essence From the perspective of platform construction and industrial transformation, the large -scale model has industrial -level characteristics, provides all -process tools and methods to support application landing, and create an open ecosystem that stimulates innovation.

In other words, the clear scene demonstration value is the "last mile" that promotes the large model of the big model to the big model. At this time, Baidu's rich industrial ecological layout shows key value. Baidu Mobile Ecology, the "test field" in the era of big models

There is a saying in the field of technology, called "your own parachute jump first". For Baidu, which hopes to explore the application of large models and industrial transformations, it is also necessary to "use its own big models first."

Fortunately, Baidu's rich mobile ecological layout is not only the value fertile soil of the application of large models, but also can also become a big model of more developers and enterprises " Test field.

At the 2022 million Baidu Mobile Ecological Conference, we saw the integration of Baidu Mobile Ecology and Wenxin Model. Baidu Mobile Ecology has always been the "first stop" of Baidu AI technology. It has long been adhering to the principles that AI technology can use and can be used.

In combination with the big model, we can first see Baidu's latest release of the cross -model big model "Knowing" for searching scenes. It can understand the variety of resources such as the text, pictures, videos, and structured information of the entire network, thereby breaking the boundaries of the resource form, presenting the results that meets the needs of the user the most to users, and truly realizing the big model technology and search, information flow, etc. In -depth integration of Baidu's core business.

For more developers and creators in Baidu Mobile Ecology, the "Creator AI Assistant Group" based on a large model of Wenxin can bring epoch -making AIGC experience. In this AI team, there are copywriting, illustrator, and video producers. With the uniform efforts of these "assistants", creators can achieve "one person becomes a team."

Based on the big model of Baidu Wenxin, the "AI copy assistant" can quickly produce different styles of articles and content paragraphs according to the needs of the creator. The paintings allow the creator to achieve "freedom of pictures." Related capabilities can be experienced in Baidu APP, and the effect is very prominent.

"AI Video Production Assistant" is based on the natural language processing and cross -modal AI technology of the Grand Heart Model, allowing the creator's graphic to automatically generate videos, which can reach the "minute level". This technology has shortened the creator from graphic to videos by more than three times, and also brought the effect of "one content, two benefits" effects to the creators.

"Creator AI Assistant Group" can be said to be a panoramic display of the application value of the big model -AIGC (AI Generated Content, artificial intelligence automatically generate content). Based on the AIGC capabilities created by large models, a variety of AI independent creation can be achieved, and a variety of creative auxiliary functions, realizing a series of technological leaps such as automatic production copywriting, automatic production pictures, and graphic videos. At present, the "Creator AI Assistant Group" has begun to invite internal testing and will be opened to more creators from the fourth quarter of this year.

It is not difficult to see that the ability of Wenxin's big model to bring Baidu mobile ecology is not the simple ability upgrade of "icing on the cake", but a real leaping innovation. Core creative threshold troubles allow creators to focus on the core part of creation.

This ability to upgrade is of course essential for the future development of Baidu's mobile ecology. But more importantly, this combination shows the possibility of a large model towards deep industrial applications.

When the AIGC harvests, the industry transformed deep cultivation: the Dream of Hexin Models

不久之前，科技部发布了《关于支持建设新一代人工智能示范应用场景的通知》（以下简称《通知》），指出要充分发挥人工智能赋能经济社会发展的作用，打造形成一批可复制、 Prompened benchmarking demonstration application scenarios.

It is not difficult to see that with the scene as the lead, the industrial landing and large -scale auxiliary of AI technology are becoming the main task of the current development of the AI ​​industry. The same is true for large models. In the industry, the value of various industries to truly understand the value of the big model. The ability to apply large models must first take the scene as the guide to demonstrate the coordinates.

The in -depth application of Baidu Mobile Ecology for large models can just form this kind of demonstration that is urgent for the times. From the perspective of large models to industrial transformation, Baidu Mobile first proves the availability of AIGC. AIGC, as an important station for the industry to move towards the industry, can digest and absorb the value of AIGC from the experience of Baidu, thereby achieving its own large models. Secondly, the integration of Baidu's mobile ecology and large models also proves the integration value of large models and industrial needs, and the real situation of the industry. Baidu Mobile Ecology has huge traffic and diversified businesses. These conditions are the test of large model availability.

Ren Zeping, a well -known economist, also said during the Vientiane Conference: "In the AIGC era, Baidu APP, as the" outpost "and 'base", is the broadest world of AIGC, and will also become a new AI infrastructure that everyone can use. The model of "Baidu to set up a platform, and everyone's singing" will stimulate the greater creativity of the whole people. Many new models and new applications will break out of the soil. "

From the combination of Wenxin's big model and Baidu mobile ecology, to the exploration of AIGC industry practice, to the possibility of more industries to apply large models, we can see that this path brings triple value to the 3.0 phase of the big model. 1. Self -proven self -proven: the technical ability of the big model has tended to be full

The application of Baidu Mobile Ecology's large model of Wenxin first allows developers in various industries to see what the big model can do and what kind of imagination, and this belongs to its technical capabilities that comes from Wenxin's big model. The seeds of large models need to prove their value in the soil, and AIGC and Baidu Mobile Ecology provides such soil. Taking AI intelligent writing as an example, ERNIE 3.0 Zeus is a 100 billion model related knowledge related to the fusion task related to the latest public model. On the basis of learning massive data and knowledge, it further learns more than a hundred different forms of task knowledge. Various tasks show strong creative ability. From financial and sports news, to literary creation such as poetry and novels, the fields of AI writing are becoming wider and wider, and creative forms and content have become more colorful.

During this year's college entrance examination, ERNIE 3.0 Zeus blessed the AI ​​digital person, Xiaoxiao, challenged the college entrance examination composition, and won 25%of the total exams at 40 seconds 40 articles. ERNIE 3.0 Zeus's intelligent creative capabilities have now been opened to the public. Developers from all walks of life can experience and customize ERNIE 3.0 Zeus text creative capabilities online. Let's talk about AI painting. AIGC applications such as AI paintings want to be guaranteed. Behind it is the strong support of the cross-model big model of Wenxin. As the world's largest Chinese cross-modal generation model, the number of ERNIE-VILG parameters has reached 10 billion. It integrates the mission of the "Wen generating" and "Graph Genesis" to the same model for end -to -end learning, thereby enhancing the cross -modal semantics of text and images. Recently, the ERNIE-VILG text chart generation algorithm ushered in upgrading. Through the gradual diffusion model, the generating space from small and large, generated contours from thick to fine, and automatically select the optimal generation network according to the generation stage. The effect is further improved. It can be said that behind each AI painting, there is a proof of full -scale technology capabilities.

2. Wo Tongtong: The industrial tools and development platforms are gradually complete

Developers in various industries and fields, if you want to really use large models, need targeted platforms and tools, thereby reducing the technical acquisition threshold and improving the efficiency of industrial development. For this reason, the large model of Wenxin has a wealth of exploration in tooling and platformization. Taking AI painting as an example, Wenxin Yige is the AI ​​art and creative auxiliary platform launched by Baidu. It is a product released by Baidu's continuous technological innovation relying on flying paddle and Wenxin big models, and released in the "AI drawing". Facing people with design needs and creativity, Wenxin can intelligently generate a variety of AI creative pictures based on the large model of Wenxin, assist creative design, and break creative bottlenecks. For developers with higher convenience demand, you can also use Wenxin Wentu to generate large-model ERNIE-VILG services, quickly obtain the technical experience of immersive text map generation to generate large models, and achieve product integration flexible and efficiently. In this way, we can see that developers who have different needs and different considerations for AI paintings have platforms and tools to obtain the AIGC ability they want. This is the fertile soil of the big model to the industry and the application.

3. Hexia Liang Liang: The beginning of the large model of large models

Facing the wider industry application prospects and industry imagination, the large model of Wenxin not only promotes the combination of Baidu business, but also actively explores the forms and methods of external empowerment. The knowledge enhancement characteristics of Wenxin's big model can achieve better results in the fields of physical Q & A, knowledge prediction and other fields, and enhance the availability of large models in the industry scene. At the same time, Wenxin · Industry's big model adopts industry knowledge enhancement technology to learn from industry characteristics and industry characteristics, thereby enhancing the adaptability of large models to industry applications. At present, Wenxin has released the NLP big model "State Grid-Baidu · Wenxin", which has enhanced the knowledge of the Power Industry of the Power Industry and the financial industry NLP Model "Pudong Development-Baidu · Wenxin". Complete the landing application. In the field of aerospace, Wenxin Model and China Aerospace released the world's first aerospace model- "Aerospace-Baidu · Wenxin Big Model" to promote the application of AI technology in the aerospace field.

Wenxin Grand Model has already launched a development platform for development kits, APIs, and built -in Wenxin model capabilities, which can allow industrial developers of different fields and different needs to accurately obtain large models to empower large models at lower costs and lower thresholds to promote labor The widespread application of intelligence.

Wenxin's big model has accelerated its landing in the industry, becoming a new engine to promote the development of AIGC, and at the same time as a new base that promotes the intelligence of the industry. The large models of large models and landing and hundreds of industries may be the Dream of Hexia Dream of the Smart China.

- END -