Domestic large-scale models accelerate the development of the enabling industry

The industrial version of the humanoid robot connected to the large model was "trained" on the spot in the assembly workshop of new energy vehicles. Courtesy of Shenzhen UBTECH Technology Co., Ltd

　　General Secretary Xi Jinping pointed out: "Whoever can grasp the opportunities of new economic development such as big data and artificial intelligence will have the pulse of the times." The "Government Work Report" proposes to deepen the R&D and application of big data and artificial intelligence, carry out the "artificial intelligence +" action, and build a digital industrial cluster with international competitiveness.

　　Since the beginning of this year, the development of the mainland's artificial intelligence model has accelerated, and the implementation of industrial applications has been accelerated: Shengshu Technology and Tsinghua University have released a domestic self-developed video model, which supports one-click generation of high-definition video content up to 16 seconds and a resolution of up to 1080P; At the Beijing International Automobile Exhibition, domestic car companies launched a number of new models equipped with large models, and continuously optimized the user experience in terms of multi-sensory interaction and autonomous driving. The humanoid robot was connected to the large model for task scheduling and application development, and quickly "learned" fine operations such as folding clothes, sorting and classifying......

　　According to incomplete statistics, the number of domestic large models has exceeded 200, covering multiple industries, and the application scenarios are constantly expanding. According to the latest data released by the Cyberspace Administration of China, as of March this year, a total of 117 generative AI services in mainland China have completed filings. How can large models further empower the development of the industry? What else do we need to focus on in the future? The reporter recently conducted an interview.

　　Stick to the application orientation

　　It has established systematic R&D capabilities covering theoretical methods and software and hardware technologies, and a number of large-scale model applications with industry influence have emerged, forming a large-scale model technology group that closely follows the world's cutting-edge

　　When you enter the demand instructions in the dialog box on the computer, not only the required data report will appear on the screen, but also the analysis chart automatically generated in combination with the data report. Not long ago, the large model created by China FAW and Alibaba Cloud Tongyi Qianwen was applied. "We have to do reports almost every day, which takes a lot of time. Now, with the help of large models, the work efficiency is significantly improved. The staff of China FAW introduced.

　　Compared with the traditional "fixed Q&A", the application can further analyze the reasons and trends behind the Q&A, and deepen the application of data. "Based on the initial corpus of 468 indicators, we formed 60,000 pieces of evaluation data, and built the capabilities of this large model in indicator design, indicator dismantling, data sourcing, data modeling and data analysis." Li Qiang, vice president of Alibaba Cloud Intelligence, introduced.

　　For example, when Li Qiang asks "why the output of a certain model does not meet expectations", the application of the large model can first compare the expected output with the actual output to obtain the difference, and then analyze the variables such as the status of production equipment, the quality of model parts, and the deep-seated reasons such as fluctuations in raw material supply, energy consumption and supply stability, and finally find out the reasons with the greatest correlation and generate visual reports through data investigation. In addition to the automobile manufacturing industry, Tongyi Qianwen has been applied in mining, transportation and other fields.

　　At present, the continental large model has entered a period of accelerated development, and various technical branches such as natural language processing, machine vision and multimodal are all following up and developing rapidly. Under the joint promotion of industry, academia and research, the mainland has established systematic R&D capabilities covering theoretical methods and software and hardware technologies, and a number of large-scale model applications with industry influence have emerged, forming a large-scale model technology group that closely follows the world's cutting-edge.

　　There are two main development paths for the industrial application of large models in mainland China: one is to build a cross-industry generalized artificial intelligence capability platform, that is, a general large model, whose application is accelerating from office and life to medical, industrial, education and other fields; One batch is aimed at industry models in vertical fields such as biopharmaceutical, remote sensing, meteorology, etc., giving full play to its advantages in depth to provide high-quality professional solutions for specific business scenarios.

　　Liu Shijin, former deputy director of the Development Research Center of the State Council, believes that the interaction between the mainland's super-large-scale market advantages and the application of new technologies will bring more development opportunities in the digital economy and other fields. More innovative entities promote innovation through market competition, and provide more application scenarios for technology and engineering. According to Shan Zhiguang, director of the Department of Informatization and Industrial Development of the State Information Center, the development of large models in mainland China should adhere to the application orientation, attach importance to the combination of artificial intelligence basic large models and industry data knowledge, and deepen the research and development and application of artificial intelligence based on application development and market demand.

　　Focus on user needs

　　Consumer electronics and smart terminal products equipped with large models have been launched one after another, and large models are expected to come to more users and become a "good helper" for work and life

　　At present, artificial intelligence large models are divided according to the deployment mode, which is mainly divided into cloud-side large models and device-side large models. Unlike the cloud-side model, which is mainly oriented to the needs of industrial applications, the device-side model mainly serves individual users. Since the beginning of this year, domestic manufacturers have successively launched consumer electronics and intelligent terminal products equipped with large models, which are expected to come to more users and become a "good helper" in work and life.

　　Not long ago, Lenovo Group released a personal computer product equipped with an artificial intelligence assistant. With a large personal model deployed locally, the AI assistant can not only interact with the user, but also learn from these interactions and improve themselves. For example, it can better reason through personal travel records and shopping preferences, and even predict tasks based on the user's thinking patterns and behavior frequency, and take the initiative to make suggestions and find solutions.

　　"From public intelligence to personal intelligence, from technical tools to 'assistants' who can understand individual needs, in the process of artificial intelligence industrialization, by distributing the workload of large models in public clouds, private clouds, edges and terminals more evenly, it will help promote the intelligent transformation of all walks of life." Yang Yuanqing, chairman and CEO of Lenovo Group, said.

　　At the beginning of this year, Honor mobile phones released a new generation of full-scene operating system, with a monthly usage of 15 million times for its large model intelligent Q&A, and a maximum daily usage of 850,000 times. The introduction of large models is expected to bring a functional leap to mobile phones. The use of large models to refine the key points of call content, "dialogue" with mobile phones, automatic selection of materials to generate video content, etc., these new functions have been successfully transformed into a "new engine" in the mobile phone market.

　　"AI mobile phones mean using technologies such as large models to reconstruct and enable mobile phone applications, operating systems, consumer experiences, and business logic, and make full use of terminal manufacturers' understanding of user needs to better serve users." Zhao Ming, CEO of Honor Terminal Co., Ltd., said.

　　As a new generation of intelligent terminals, large models accelerate the deployment and application of intelligent networked vehicles. In addition to interacting with people more naturally in the intelligent cockpit and more accurately identifying people and objects inside and outside the car, large models can also improve the efficiency and safety of autonomous driving systems. "The large cloud model has large parameters and strong computing power, and can complete a large number of tasks such as data annotation and data fusion. The large model of the vehicle side has a small number of parameters, which can save the inference time of the vehicle calculation and help the development and application of autonomous driving technology. The interviewed experts said.

　　The software industry is also "embracing" the big model. Wei Daisen, Vice President of Inspur Group, said: "At present, with the empowerment of large models, the software industry is accelerating its evolution. First, it has triggered a change in the way of software production, with large models promoting automation and intelligence in software production, low-code becoming a hot spot, and artificial intelligence-enhanced development as a trend; Second, the evolution of software form is accelerating, moving towards platformization, assembly, and cloud native; Third, software products are highly intelligent, from embedded intelligence to native intelligence, enabling business process automation and intelligence. ”

　　Speed up the completion of shortcomings

　　Further strengthen the overall planning of resources and R&D forces, strengthen the role of large models in the development of scenarios, and achieve high-quality application breakthroughs in large model technology

　　Mobile applications are developed, data resources are abundant, application scenarios are diversified, and the AI industry chain is highly complete...... Industrial application has become one of the important advantages of the mainland's artificial intelligence industry, and it is also the key to promoting the accelerated development of artificial intelligence.

　　An Xiaopeng, member of the executive committee of the China Informatization 100 Association and vice president of Alibaba Cloud, believes that the application of large models in different industries can promote the upgrading of product competitiveness and create new added value. On the other hand, it can optimize processes and decision-making in a wider range of complex scenarios, and improve the efficiency of enterprise innovation.

　　The technical potential and market space of large models are huge, but the technical threshold is high, the development is complex, and the cost is high, and the deep integration with industrial scenarios is inseparable from the support of platform tools. Peng Zhen, chairman of Inspur Information, said that "artificial intelligence +" is both a goal and an opportunity. Application innovation requires the coordinated development of algorithms, computing power, and data, and it is necessary to carry out global innovation with the system as the core, realize equal emphasis on innovation and application, and enable intelligent applications to emerge at a faster speed and meet differentiated needs.

　　At present, there are certain challenges in the development of large models in mainland China, such as the bottleneck of computing power, the need to expand high-quality training datasets, and the emergence of landmark applications. "For example, deploying a large model on the public cloud can allow users to obtain the latest functions and applications of the large model in a more timely manner. At this time, the public cloud is not only a simple computing or storage facility, but also an application facility for large models. Shan Zhiguang said that the construction of computing infrastructure in mainland China should take the path of sharing and efficiency, attaching importance to both hardware construction and demand and application development, efficient and intensive use of computing resources, and further improving the service level and ability of algorithms, computing power and data resources.

　　Zhang Xiaohui, chief engineer of Guangdong Topstar Technology Co., Ltd., said: "Taking the implementation of the manufacturing model as an example, the digital and intelligent level of many industrial control systems is not high, the industrial data collection is difficult, the standards are not uniform, the value is not high, and the data utilization is not sufficient, which makes it difficult for many enterprises to optimize the production process with the help of large model technology." Wu Tongning, deputy director of the Artificial Intelligence Research Center of the Chinese Academy of Information and Communications Technology, believes that the emergence of many innovative applications has put forward a higher demand for artificial intelligence software and hardware support systems. The growth in demand brought about by new technologies can drive the development of domestic computing power, algorithms, and data-related innovation subjects.

　　The interviewed experts said that the creation of new value, adaptation to new industries, and the generation of new kinetic energy are the key elements to accelerate the development of new quality productivity. In the future, the mainland needs to further strengthen the coordination of resources and R&D forces, strengthen the role of large models in the development of scenarios, achieve breakthroughs in the high-quality application of large model technology, and drive high-quality economic and social development.

Source: People's Daily