According to a recent slide presentation delivered to Chinese Premier Li Qiang, China is facing challenges in keeping up with the United States in the field of artificial intelligence (AI) due to numerous obstacles in theory and technology.
As reported by state-owned China Central Television (CCTV), Premier Li was briefed on these challenges during his recent visit to the Beijing Academy of Artificial Intelligence (BAAI), a private non-profit organization founded in 2018.
The primary hurdle highlighted in China’s efforts in generative AI, as outlined in the presentation at BAAI, is the overreliance on Meta Platforms’ Llama program. This program consists of a community of large language models (LLMs) introduced just last month. LLMs play a crucial role in training services like ChatGPT and other AI applications for content generation such as music, code, images, text, simulations, and videos.
The presentation pointed out that the heavy dependence on Llama for most domestic LLMs has led to a “significant lack of self-sufficiency” in this sector of Chinese AI development. Meta, the parent company of Facebook, made its Llama 2 Artificial model open-source for research and commercial purposes in July last year.
Chinese Premier Li Qiang, positioned fifth from the left, was observed at the Beijing Academy of Artificial Intelligence on March 13, 2024, inspecting the facilities of the non-profit organization in the capital. (Image Source: Xinhua)
State entities are collaborating with specific Chinese tech firms to drive AI innovations; however, challenges persist in acquiring the necessary computing infrastructure for LLM training.
The lecture highlighted the diversity among the numerous custom chips used, leading to instability in the 100-billion-feature instruction set for Chinese LLMs. Additionally, US sanctions on China have impeded access to advanced electronics crucial for indigenous AI projects, as they are predominantly based on American technologies.
The CEO of Baidu criticized the discontent among Chinese tech companies regarding AI designs as a misallocation of resources.
While over 40 government-endorsed LLMs and related AI initiatives are already operational in China, the market boasts over 200 locally developed LLMs.
Content control in AI-generated material emerged as another critical issue discussed in the BAAI presentation. It emphasized the challenge of producing “accurate content that aligns with reality” while considering various ideologies and emotions, a unique hurdle for Chinese-developed LLMs.
AI bots like ChatGPT and Google’s Gemini are susceptible to generating misleading outputs, termed as illusions.
Premier Li Qiang of China vows to provide more flexibility to surpass competitors in the global AI race.
Although the presenter of the BAAI session was not explicitly named by CCTV, the slides displayed the logo of Beijing Zhipu Huazhang Technology Co (Zhipu AI). A representative from Zhipu AI confirmed the company’s participation in the Chinese leader’s visit to BAAI the day before, showcasing their contribution to the innovative ecosystem cultivated by BAAI, including the development of a 100-billion-parameter LLM.
