Domestic large-scale models: both upward breakthroughs and downward roots are needed
2025-03-07
"Do you use DeepSeek?" "I do. My major is computer science, and I have a preference for artificial intelligence." After the press conference of the Third Session of the 14th National People's Congress held on March 4, Lou Qinjian, the spokesman of the congress, responded angrily to this sentence when interviewed by on-site reporters. During this year's Spring Festival, the domestic large-scale model DeepSeek left a deep impression on people with its low-cost, high-efficiency, and strong intelligent application experience. At this year's National People's Congress and Chinese People's Political Consultative Conference, "domestically produced big models" and "DeepSeek" were also frequently mentioned by many delegates and members. In recent years, with the accelerated development of China's artificial intelligence large model industry, more and more domestic large models have performed amazingly. As the core driving force of the digital age, artificial intelligence is accelerating its penetration into multiple fields of the national economy. While continuously breaking through the performance ceiling, how domestic large models should take root in industrial practice and empower industrial upgrading has become a hot topic of discussion among the representative committee members. The vertical model is rooted in the industrial soil with quick response to questions, clear and concise thinking process, comprehensive and detailed reference materials... Through the universal big model, many people have opened their first "intimate contact" with artificial intelligence. As an important breakthrough in the field of artificial intelligence, the general large-scale model has powerful language understanding and generation capabilities, and can provide intelligent support for multiple fields. With the continuous acceleration of industrial digitization in China, the demand for artificial intelligence in various industries is becoming increasingly refined and specialized. Traditional generic models often fail to accurately integrate with specific business scenarios when implemented in industrial settings, resulting in a lack of adaptability. In this context, vertical models tailored to meet the needs of industry segmentation have emerged, becoming a new trend in the application of large-scale model technology in the industry. On the first day of building the big model, we determined the strategic path of '1+N', which is' 1 base big model+N industry big models'. We have a complete solution that combines universal base, toolchain, and knowledge engineering, including 'building computing power, managing data, training models, implementing scenarios, ensuring safety, and optimizing operations' Liu Qingfeng, Chairman of iFlytek Co., Ltd., believes that the ceiling of the universal large model base is constantly being broken, and the implementation of application scenarios has entered a dividend period. It is necessary to promote the implementation with less cost, lower computing power, and higher efficiency. For example, leveraging the foundational capabilities of general models such as Alibaba's Tongyi, DingTalk has developed AI assistant products for business scenarios, covering multiple industries such as manufacturing, healthcare, retail, and education. At Jinshi Robot Changzhou Co., Ltd., through learning and accumulating a large amount of professional knowledge in robotics, DingTalk AI assistant directly serves more than 1000 distributors nationwide, efficiently answering various product after-sales questions; At the same time, it can generate work orders with just one click based on the problem description and assign them to the corresponding responsible person, reducing the solution time for after-sales problems from half a month to within 3 days, significantly improving the operational efficiency of the enterprise. Large models have great potential in vertical fields Member Zhou Hongyi, founder of 360 Group, stated that an important direction for the development of China's big models should be to leverage industry and scenario advantages, combine big models with business processes and product functions, seek multi scenario applications, and achieve vertical and industrial implementation. Accelerate the formation of the "data flywheel" effect. China has all industrial categories in the United Nations Industrial Classification. Among 500 industrial varieties, more than 40% of the products have the world's largest output, and the industry has unique advantages of being comprehensive, diverse, and large. The huge industrial scale provides fertile soil for the implementation of vertical large-scale models in the industry, but also brings risks and challenges. Industrial data is the 'nourishment' for vertical large-scale models. China has a rich variety of industries, but it also poses challenges such as diverse types and structures of industrial data, and uneven data quality. When it comes to the industrial application of big models, several representatives and committee members mentioned that we should promote the sharing of industrial data, accelerate the collection and utilization of high-quality data, and form a "data flywheel" effect in the industrial field, that is, through the continuous accumulation and utilization of data, drive the continuous improvement of big model performance. Representative Zhang Fan, Minister of Science and Technology Innovation Department of China Electric Equipment Group Co., Ltd., believes that many manufacturing enterprises started their digital transformation late, have weak foundations, and incomplete and untimely data collection in the production process; Industrial data often suffers from issues such as high noise, inconsistent formats, and poor correlation, lacking industry data standards guidance, making it difficult to provide sufficient high-quality data for training large models. At the same time, due to factors such as data security risks and protection of commercial interests, high "data walls" have been built between enterprises, and there are bottlenecks in data sharing. Based on this, Zhang Fan suggests that efforts should be made to accelerate the formation of unified artificial intelligence data format standards and specific industry standards in the industrial sector, in order to promote consensus among enterprises on the basis of data asset transactions; Leading enterprises should leverage their role as chain leaders and establish high-quality industrial datasets with demand driven and standardized formats for strategic high-value scenarios. Many regions have begun to take action around the data bottleneck in the implementation of the big model industry. For example, the "Action Plan for Promoting 'Artificial Intelligence Plus' in Beijing (2024-2025)" proposed to rely on the Beijing Data Infrastructure System Pilot Zone, create a secure and trustworthy data space, and guide enterprises and institutions to open up and gather high-value industry data. (New Society)
Edit:He Chuanning Responsible editor:Su Suiyue
Source:Sci-Tech Daily
Special statement: if the pictures and texts reproduced or quoted on this site infringe your legitimate rights and interests, please contact this site, and this site will correct and delete them in time. For copyright issues and website cooperation, please contact through outlook new era email:lwxsd@liaowanghn.com