By Pan Xutao, Chen Sida
A China-developed large climaBy Pan Xutao, Chen Sidate model using artificial intelligence (AI) technology recently drew wide attention during the 28th session of the Conference of the Parties to the United Nations Framework Convention on Climate Change (COP28) in Dubai, United Arab Emirates.
"FuXi-Subseasonal," the large AI model which features a long time range and a high accuracy for weather forecasts, is a mirror of China's burgeoning large models.
Domestic large models are on a fast track of development as various industries of the country ramp up efforts to unlock the potential of large models.
China ranks second globally in terms of the number of domestically developed large models, and a good number of pre-training large models with considerable influence in the industry have emerged in the country, according to the latest report by the Institute of Scientific and Technical Information of China, which is affiliated with the Ministry of Technology.
"The development of AI, with large models as a typical example, has shown such characteristics as rapid technological innovation, extensive application, and fierce international competition, profoundly changing the production modes and economic forms of the manufacturing sector and demonstrating how powerful AI can be in empowering industries," said Tao Qing, spokesperson for China's Ministry of Industry and Information Technology (MIIT).
Thanks to comprehensive signal control and large traffic models, many citizens who drive in the Beijing Economic-Technological Development Area have noticed a significant decrease in the frequency of being stopped by a red light. In fact, they feel that they now often see green lights all the way.
The area, also dubbed Beijing E-Town, has joined hands with Chinese online search giant Baidu to integrate various sources of spatiotemporal data, including roadside perception data and internet data.
Based on this, they have utilized large traffic models to assess the possibility of traffic congestion and conducted simulation deductions to generate plans for comprehensive signal control and automatically optimize signal timing, transforming the traditional situation of vehicles waiting for lights into the current scene of traffic lights changing for the convenience of vehicles.
It's estimated that the implementation of comprehensive signal control across the area has reduced the average time it takes a vehicle to pass through a traffic light intersection by 28 percent.
Vertical domains, from transportation to culture and tourism, and from education to healthcare, witnessed intensive launches of various types of large models recently.
Compared to general-purpose large models, large models of vertical domains feature stronger expertise and can better solve domain-specific problems and provide more targeted services.
Ctrip, China's leading online travel agency, rolled out a large model named Ctrip Wendao for the tourism industry. The model supports natural language interaction, and is able to assist users with hotel and tourist attraction reservations, trip planning and other travel decisions.
China's leading AI and intelligent speech company iFLYTEK launched an app for medical services.
IFLYHEALTH, the new app based on the company's medical large model Spark Desk, has features including self-assessment of health, medical examination report interpretation, and health record management, thus providing health consulting services for users.
A research team from Nanjing Agricultural University in Nanjing, east China's Jiangsu province, released a large language model (LLM) for ancient books.
Being able to perform lexical analysis, information extraction, intelligent poetry generation and other functions, the LLM is expected to be applied and promoted in the collation of ancient books, digital entertainment and other fields.
Data shows that over 250 companies and academic institutions in China had developed large models with over 1 billion parameters each as of October 2023, with vertical industry application becoming key track for large model utilization.
As the world's second-largest economy, China enjoys the advantages of a super-large market and rich digital resources, which enable it to provide large models with more sophisticated and richer application scenarios and an environment for continuous training, said Liu Yuanchun, president of Shanghai University of Finance and Economics in east China's Shanghai.
Computing power is the foundation of large models, massive computing power is required during training and inference.
China is vigorously promoting its east-to-west computing resource transfer project, bringing development opportunities to intelligent computing centers in the western regions of the country.
On Nov. 30, 2023, Qinghai Kunlun Artificial Intelligence Computing Center, the first intelligent computing center in the Qinghai-Tibet Plateau, was officially inaugurated in northwest China's Qinghai province.
The computing center relies on the Sanjiangyuan national big data base located in the province for the management of computing resources and operations.
On Dec. 7, 2023, northwest China's Ningxia Hui autonomous region signed its first large-scale computing center cluster construction project.
Using domestic full-function graphics processing units (GPUs) as its foundation, the cluster is expected to provide computing power for large model training and inference, as well as 3D graphics rendering and other tasks.
China aims to achieve a total computing power of over 300 EFLOPS by 2025 and increase the proportion of AI computing power in its overall computing power to 35 percent, according to an action plan for the high-quality development of computing power infrastructure issued by six Chinese authorities, including the MIIT, in October 2023. (EFLOPS is a unit of the speed of computer systems. 1 EFLOPS is equal to 1 quintillion floating-point operations per second.)