Voice from the Two Sessions | Member of the CPPCC Made Suggestions on the Development of Large Model Industry: Provide High-Quality Data and Build Iconic Application Scenarios
The rapid development of generative artificial intelligence (AIGC) has triggered a new wave of industrialization of large artificial intelligence models, forming a brand-new industry trend. Shanghai is actively deploying the large model industry. The first large model innovation eco-community, “Model Speed Space”, was unveiled in Xuhui District of Shanghai in September 2023, and has gathered over a thousand artificial intelligence enterprises, 255 large model enterprises, and more than 100 investment institutions.
With the upcoming 2025 Shanghai Two Sessions, Shen Kaiyan, a member of the Shanghai Political Consultative Conference and Director of the Shanghai Academy of Social Sciences Economics Research Institute, plans to submit a proposal titled “Promoting the High-Quality Development of the Large Model Industry and Stimulating New Drivers of Shanghai’s Economic Growth”. The proposal argues that the large model industry is a powerful tool for Shanghai to develop new productive forces and accelerate the shaping of new drivers of economic development. However, it faces challenges such as high independent research and development costs, difficulties in obtaining high-quality corpora, and slow commercialization of technology, which require effective measures to address urgently.
During early research, Shen Kaiyan found that large model enterprises face high independent research and development costs, with the cost of high-performance computing equipment accounting for a large portion of the research and development budget. Enterprises can only rely on strategies such as innovating algorithm architecture to improve training efficiency and reduce training costs. To address these bottlenecks, direct financing is crucial. However, the amount of direct financing obtained by Shanghai’s large model enterprises is currently low.
She also noted that limited access to high-quality corpora is a significant constraint on the effectiveness of large model training. Although some corpora companies in Shanghai are already operating and providing corpora with high quality, their quantity is limited and their prices are high. Therefore, many large model enterprises prefer to obtain corpora from public sources or open data for large model training, leading to inconsistent data quality and formats, which reduce the effectiveness of large model training.
Shen Kaiyan further pointed out that currently, there are few commercialized applications of large models in Shanghai. The scarcity of composite talents who understand both large model technology and its applications within large model enterprises has led to uncertainty about potential user needs, affecting the progress of technology commercialization and the exploration of potential scenarios. In addition, the slow commercialization process of large model technology is also related to滞后相关的 legal institutions concerning technology solving, data security, privacy protection, process explainability, etc.
In response to the above issues, the proposal makes three suggestions:
Firstly, establish a full-chain equity investment system and build a public technology service platform. One should leverage the guidance role of government investment funds, establish and introduce industry funds in the large model field, guide social capital to increase investment in large model field achievements transformation and industrialization, establish a full-chain equity investment system from angel investment, venture capital to equity investment, and provide stable and continuous capital supply for industry development. Secondly, set up a public technology service platform to provide enterprises with shared access to high-performance computing equipment such as CPU, GPU, TPU, and FPGA to reduce the equipment purchase costs of large model enterprises.
Secondly, strengthen data circulation and cross-regional cooperation to solve the bottleneck of corpus supply. One should improve data circulation capabilities by collaborating with key central enterprises, municipal state-owned enterprises, data business enterprises, etc., to build a data open platform and promote the collection, access, sharing, and processing of high-quality data. Encourage Shanghai data institutions to open de-identified high-quality data and build operational data training bases. Support the construction of large model training data security rooms and pre-training corpora by Shanghai’s large model enterprises. Leverage the integration advantages of the Yangtze River Delta region to strengthen regional collaboration in solving corpus and computing challenges. It is suggested to include corpus and computing power cross-regional coordination as an annual topic of the Yangtze River Delta integration, fully leverage the complementary advantages of various provinces and cities in data resources and computing resources in the Yangtze River Delta region, establish a cross-regional allocation mechanism for computing power and a division mechanism for cost and benefits to solve the development bottleneck of the large model industry.
Thirdly, carry out innovation in large model technology and scenes to create iconic application scenarios. Establish a major technological innovation project for AIGC in Shanghai, focusing on general model theory, domain large models, and large model applications to address key issues in the application of large models in vertical fields. Based on “small cuts, deep applications,” create a batch of iconic application scenarios. Led by the Municipal Economic and Information Technology Commission, focus on three major areas to carry out innovation in large model application scenarios: manufacturing (focusing on new materials, electric vehicles, etc., and demonstrating applications in intelligent R&D design, future factories, industrial internet, and autonomous driving); social welfare (selecting scenes
