"Computing power is a major bottleneck restraining Chinese companies in terms of LLM applications. More efforts are needed to boost the overall planning and coordination of resources and accelerate basic research and technological innovation to actively participate in global AI governance and further promote the orderly development of LLMs," said Zhao Zhiyun, head of the ISTIC.
To help SMEs with computing power, Jiang said more than 10 computing power partners are expected to offer low-cost and high-quality computing power to AI SMEs.
A total of 10 data partners have opened 18 high-quality data sets of nearly 500 terabytes for companies to do LLM training, he added.
In addition to Beijing, cities including Shanghai, Shenzhen, Hangzhou, Nanjing, Suzhou and Chengdu have all unveiled supporting policies to drive LLM development in terms of computing power, data and ecology.
Shanghai, for instance, will offer a maximum 30 percent subsidy of companies' computer power investment, or up to 5 million yuan, to drive the research and development of LLMs. Shenzhen has said it plans to launch a guideline for the openness and sharing of public data by the end of this year.
"Though local governments have poured great investment into the development of large language models, it is not enough. More efforts are needed from the government to increase investment so that China can truly stand at the forefront of LLMs," said Dai Qionghai, an academician at the Chinese Academy of Engineering.
Compared with ChatGPT, current LLMs of Chinese companies are more "industry-driven", as seen from the dozens of large models launched currently.
Tencent Cloud, the cloud subsidiary of Chinese internet firm Tencent Holdings, launched its industry-specific large model in June.
"General or universal large models are mostly trained based on public information that may contain errors, rumors and biases and lack professional know-how and industry data," said Dowson Tong, Tencent's senior executive vice-president and CEO of the company's cloud and smart industries group.
Data from such models contain too much "noise" and are likely to cause huge legal liabilities or public relations crises. Therefore, companies need more industry-specific models, Tong said.
He said Tencent's one-stop industry-specific large model solution has covered 10 major industries, such as finance, culture and tourism, government affairs, media and education, and is able to offer 50 different kinds of solutions.
"In addition, the larger the model is, the higher the cost of training. Companies also tend to choose a suitable model at a reasonable cost," he added.
Compared to general or universal large models like ChatGPT, industry-specific large models are basically industrial versions of ChatGPT focused on niche sectors. They can better leverage industry data to offer more targeted solutions, industry experts said.
Huawei Technologies Co unveiled the latest version of its LLM last week that is also more industry-driven.
It is already being used in more than 10 sectors such as finance, manufacturing, government affairs, power, coal mining, healthcare and railways, supporting the implementation of AI applications in over 400 business scenarios, the company said.
Zhou Hongyi, founder of 360 Security Group, has said that only by empowering hundreds of industries can LLMs truly promote the revolution brought about by AI.
The real potential of LLMs comes from an enterprises-oriented market and the country should seize the opportunity of industrial development to develop LLMs, he added.
