The launch of OpenAI’s ChatGPT marked a pivotal shift in the field of artificial intelligence. As a result, many global AI enterprises, including those from China, are working to develop generative AI technologies, largely encompassing general large models and large language models (LLMs). This seismic movement has effectively rendered the development and utilization of large models a globally competitive endeavor.
According to recent figures from the Institute of Scientific and Technological Information of China, the United States and China collectively account for over 80% of the world’s large model distribution, with the U.S. leading the race in terms of quantity, and China rapidly gaining momentum since 2020.
The Trio Dominating China’s Large-Scale Model Development Landscape
China’s large-scale model development is led by three key player categories: top Internet platform companies, AI companies specializing in computer vision, and new startups fueled by the ChatGPT boom. Let’s take a closer look at each of these players:
- Chinese Tech Giants
Chinese tech giants are making substantial strides in large model development, each carving a distinct path. Baidu leads the pack with its chat robot Wen Xin Yi Yan, dubbed the “Chinese version of ChatGPT,” reflecting a robust investment in product and ecological capabilities.
Alibaba’s approach centers around their Tongyi large-scale model series, aiming to infuse future products with large models for sweeping upgrades. Huawei is banking on its industrial chain advantages and computing power to launch its Pangu Chat, spanning multiple AI domains. Tencent, having debuted the HunYuan AI Large Model, leverages it in advertising, showcasing commercial applications of large models. Lastly, ByteDance is facilitating large model training through its Volcano Engine, signaling an ongoing focus on developing large models, particularly for language and image applications.
- Computer Vision AI Enterprises
ChatGPT’s influence has stimulated fresh AI investments in China, prompting activity from computer vision companies, notably the “Four AI Tigers”: SenseTime, Yuncong Technology, Megvii Technology, and Yitu Technology. As of now, only SenseTime and Yuncong have launched large models.
In April 2023, SenseTime unveiled its strategic plan for AI development, which hinges on its SenseNova large model system, demonstrated with applications in content generation, data labeling, and 2D/3D digital human generation.
Meanwhile, CloudWalk Technology publicized its Comfort Model for beta testing in May 2023, following ambitious funding plans to bolster research and development. The company demonstrated the basic functionalities of the model, which include applications in quizzes, Chinese-English language translation, programming, and reading comprehension.
Despite financial concerns and skepticism about their ability to produce large models, these companies strive to capitalize on ChatGPT’s momentum. Concurrently, speech recognition companies like iFLYTEK are entering the scene, showcasing applications of their Spark Cognitive LLM, released in May 2023, across multiple sectors.
- New Startups
The ChatGPT phenomenon has spurred the rise of numerous startups in China’s large-scale model sector, led by two distinct groups of founders. The first group comprises high-level executives who have left prominent AI and Internet companies, while the second group consists of individuals with academic backgrounds from universities and research institutions.
The first group has witnessed numerous executives leaving their roles at prominent AI and Internet companies to seize the opportunity presented by the ChatGPT wave. For instance, Wang Huiwen, co-founder of the original Meituan, established Beijing Guangnian Beyond Technology Co., Ltd. and subsequently acquired AI startup First-class Technology Oneflow in March 2023. Likewise, former Sogou CEO Wang Xiaochuan initiated Baichuan Smart Company, and Yan Junjie, former vice president of SenseTime, co-founded Minimax, all with a focus on large-scale model technology.
Meanwhile, several startups with academic roots have emerged from universities such as Tsinghua University and Fudan University, and other research institutions. For example, Lingxin Intelligent Company, founded by Huang Minlie, an associate professor of the Computer Department at Tsinghua University, has been exploring super-anthropomorphic large-scale models since late 2021. Similarly, Shenyan Technology Company, with a founding team largely drawn from the NLP laboratory of Tsinghua University, has enlisted Professor Sun Maosong, the academic leader of the lab, as the company’s chief scientist.
Government Incentives and Regulations Mold China’s AI Model Landscape
In the global race for AI dominance, nations are implementing various policies to gain an edge in the upcoming industrial transformation. China is no exception, with the government introducing guidance programs and incentive support policies for the advancement of artificial intelligence during the “14th Five-Year Plan” period. These policies are designed to chart the overall development direction and technological priorities of artificial intelligence.
Major cities, including Beijing, Shanghai, and Guangzhou, have also put forth specific plans. Notably, on May 30, 2023, the Beijing Municipal Government issued “Several Measures to Promote the Innovation and Development of General Artificial Intelligence in Beijing,” which includes proposals to construct a “Beijing Public Computing Power Center for Artificial Intelligence” and a “Beijing Digital Economy Computing Power Center.”
China is not only implementing incentive policies to support large-scale model development but also refining regulatory measures. The Cyberspace Administration of China recently released the “Generative AI Service Management Measures,” the first regulatory document for the generative AI industry in the country. The draft consists of 21 articles covering various requirements, including admission criteria for AI service providers, algorithm design, data selection, content generation methods, user privacy, and business secrets protection.
Challenges for China’s Large-Scale AI Model Development
China’s large-scale AI model development is progressing rapidly, with Baidu and HKUST Xunfei stating that they are closing the gap with ChatGPT. Baidu’s CEO, Robin Li, believes the difference between Baidu’s large model and ChatGPT is minimal, while HKUST Xunfei’s “Spark Cognitive Big Language Model” matches ChatGPT in key areas. However, China’s large-scale model development is still facing significant challenges.
One such challenge is the constraint on chip supply. In October 2022, the U.S. Department of Commerce imposed an export restriction on high-end GPUs, including Nvidia’s high-end chips such as the A100 and H100, preventing their export to China. As a result, this constraint threatens to slow down China’s large-scale model training and put the country at a disadvantage in the technology competition.
Another significant issue pertains to data security. The global use of ChatGPT has led to a number of data breaches, prompting regulatory authorities in Italy, Canada, and other nations to scrutinize the data security risks associated with ChatGPT and OpenAI, the company behind it. Furthermore, the acquisition and use of freely available public data and internet resources, which commonly inform AIGC product training, are controversial due to concerns about authorization, ownership, and usage rights. As such, China’s large-scale models are expected to face similar regulatory scrutiny regarding data sources and usage security in the future.