AI Development in China – the First Half of 2023

In this special series, we gathered a 6-month report (from January 2023 to early July 2023) detailing the transformative impact that the launch of ChatGPT has had in

• China – the most famous US rival

• India – the second-largest talent pool for AI and ML

• The UK – emerging leader in Europe

While the US media often gets caught up in self-centered narratives, we firmly believe that expanding our horizons to encompass global AI affairs, cultural nuances, and political variations can benefit one a lot.

China's rapid advancement in the field of artificial intelligence (AI) has garnered significant attention worldwide. Despite geopolitical rivalries and debates over technological supremacy, it is essential to analyze China's AI development objectively. Two key drivers have shaped China's AI sector: global resource scarcity and a whole-of-government approach. This article explores these drivers, current regulatory efforts, unresolved challenges, trends in AI development, and the key players contributing to China's AI landscape. There is also much less AI doomism in China. Why?

Introduction

AI battle between China and the US has been a topic for a long time. It's worth remembering AI Superpowers: China, Silicon Valley, and the New World Order, the book written in 2018 by Kai-Fu Lee. But the end of 2022 and the glorious marsh of generative AI changed the balance of power and it’s important to learn what is the current state of AI in China.

China's growing focus on AI, coupled with its significant investment and numerous AI publications, positions it as a key player in the field. To provide an unbiased analysis of China's advancements, this report incorporates both Chinese and English language sources. By offering a comprehensive understanding of the PRC's progress and contributions to AI, our goal is to present readers with a fresh and holistic perspective that goes beyond conventional comparative analysis.

China’s Political-Economy: Dreams of Self-Reliance, Realities of Globalization

Two key drivers stand out in China’s AI development: One global, and one domestic.

The first, global driver defining the state of China’s AI is insufficient resources and an excessive dependence on the global supply chain. COVID-induced supply chain disruptions and their knock-on effects have highlighted China’s desperation for everything related to AI development. From human capital for research and development to essential components like semiconductors and a coordinated push to control access to rare earth metals, China’s attempts to address and leverage scarce resources will continue to define her AI sector. Moreover, the relentless, international pursuit for supremacy in AI will only further aggravate China’s underlying scarcities, as other nations are likely to continue employing zero-sum economic controls to further undermine China’s already limited access to essential AI resources.

The second, domestic driver key to understanding China’s AI sector lies in her political, whole-of-government approach. China’s unique political and economic system, known as Socialism with Chinese Characteristics, allows for an integrated approach to AI, not limited by Western ideas of the rule of law or anti-trust legislation. Through state-owned enterprises (SOEs), military, and academic institutions, China’s one-party system uses military-civil fusion to give comprehensive prioritization to the development of AI technology – this both quicken its development and helps address the above-mentioned issues of scarcity. Unlike other countries that can only enact weak, term-limited legislation to temporarily support critical industries, the Chinese system’s dedication to Five-Year Plans is not limited by discontinuous terms or bound by opposing political parties. Plans can thereby compound upon one another, and institutions can coordinate with each other to work toward one national goal: Become a global AI leader by 2030.

A recent example illustrating this reality is China’s Dual Circulation "development paradigm." Introduced in response to the vulnerabilities exposed by the COVID-19 pandemic, this strategy aims to boost domestic demand, foster innovative capacity, and reduce reliance on foreign markets – effectively demonstrating how quick and capable the Chinese Communist Party (CCP) is at manufacturing industrial solutions to address structural scarcities. While it is still unclear whether China will be able to disentangle itself from AI’s incredibly globalized supply chains, signs of both success (or lack thereof) frequently emerge within this sector, due to its high priority among CCP bureaucrats and positioning as a newer, more exposed sector within the global economy.

Existing challenges

Pretty much everything about China LLMs today looks like mostly big talk and announcements (especially in their hopes for LLM industry application) but little action, as few can or have been able to make as big of a splash as ChatGPT. You'll notice that most either have invitation-based testing or that users must qualify for testing of the product. Could this have to do with the heavy degree of censorship in China? Are companies running a serious risk if they do not smooth out political kinks before releasing this technology to the public? For example, many of these LLMs are probably being trained with language that is either 1) not from within the PRC (and thus not filtered by the great firewall) or 2) language/data from before the large-scale implementation of the great firewall). How do you train an LLM to ignore politically sensitive questions about China's ethnic minorities, unseemly anti-government demonstrations, political controversies, or the blatant historical revisionism that occurs in China? Then how do LLM trainers ensure that the answers spit out by these LLMs are politically correct – not only in Chinese but in all applied languages? I'm afraid that these issues and all-around slow release to the general public will only exacerbate the existing issues of adoption, recognition, and efficiency of China's domestic LLMs. Worse yet, the question remains: Can anyone really catch up with ChatGPT 4? According to a few published announcements, most Chinese are only trying to compete with ChatGPT 3.

Further exacerbating development are hardware issues. Pan Helin, Joint Director and Researcher of the Digital Economy and Financial Innovation Research Center at Zhejiang University International Business School said, "There is currently a gap in the AIGC (Artificial Intelligence General Chip) in China. The demand is there, but the products are not. Therefore, it is the right strategy for major manufacturers to seize the opportunity and enter the AIGC market."

Current Trends in AI

Q1 2023: If you can’t beat them, join them... (Then regulate them from within)

Q1 2023: January 2023

Unsurprisingly, most of China’s billionaires are C-suite executives for tech or tech-adjacent companies. Mysterious disappearances of these business tycoons (like Xiao Jianhua, Jack Ma, Ye Jianming, Bao Fan, etc.) are not unheard of in China and have been covered extensively by Western media. In January, a notable shift in Beijing’s approach became apparent. With China approaching the second and third phases of its national objective to “make breakthroughs [in AI] by 2025” written in 2018, and “establish China as the world leader in AI by 2030,” there has been an increased sense of urgency to meet those deadlines. Instead of merely making influential technologists vanish, the CCP announced in January, that it would be acquiring stakes in two major tech firms, Alibaba and Tencent, by using a special government mechanism, often referred to as “golden shares.” While these “stakes are sometimes very small, [...] they tend to give the government board seats, voting power and sway over business decisions.” This change signals a departure from coercive disappearing acts and instead allows the CCP to places its voices in tech boardrooms and hands within AI steering committees, providing a more direct influence – likely in hopes of achieving their national goals in both a timely manner and according to their standards. It was particularly interesting to find that “Top Ten Institutions in the World by Number of AI Publications in Natural Language Processing and Speech Recognition 2021” the only two Chinese institutions found in these lists not under the direct control of the CCP were Alibaba and Tencent. It goes without saying these “golden shares” have effectively reigned them in.

February 2023

Biren and Denglin Technology join other Chinese startups in an attempt to circumvent US-led restrictions and fill the market void left by controls that have left high-end GPUs like the A100 Tensor Core GPU, upcoming H100, along with the variety of systems that incorporate them, inaccessible. China, facing restrictions on chip manufacturing technology from the US, Japan, the Netherlands, and potentially the EU, Biren has chosen to focus on general-purpose GPUs (GPGPUs) and to prioritize AI breakthroughs. Denglin’s focus, on the other hand, used the Tensor engine and programmable GPGPU engine to help develop China’s first cloud-based AI computing platform with GPU as its core technology. These advancements signify a new route to commercializing GPGPU products in China, while tactfully circumventing foreign sanctions.

March 2023

Baidu, often regarded as China’s equivalent of “Google” and a prominent search engine and artificial intelligence company, unveiled Ernie Bot, an AI-powered LLM, to the public. However, the launch fell short of expectations due to a lack of real-time demonstration and the limitation that only a small trial group could access it. Despite the underwhelming debut of Ernie Bot and the subsequent decline in Baidu’s stock price, Baidu went on to set up a 1 billion yuan ($145 million) venture capital fund specifically for startups developing artificial intelligence applications later that month.

Gunning for ever-higher frequency networks, higher capacity, and lower latency, “China Unicom, China’s third-largest wireless network operator” announced that it “expects to complete technical research and launch early applications for 6G technology by 2025.” Investors, perhaps anticipating this announcement, saw China Mobile Ltd. (China’s first-largest wireless operator) become the third-largest stock listed onshore after a trading frenzy driven by AI bets and the government’s 6G ambitions. From GPUs to GPGPUs, to 6G, even before Q1 ended this year, it is becoming increasingly evident that China’s people, investors, and most-importantly government are placing a huge bet on being able to provide both the infrastructure and breakthroughs needed to both achieve its national goals of becoming a global AI leader while achieving self-reliance. Both have difficult goals on their own. But achieving both at the same time? Truly a huge bet.

Q2 2023: April 2023

Qihoo 360 (full name 360 Security Technology Inc.) announces the launch of the chatbot 360 Zhinao 4.0 LLM. Users must qualify for testing, but the firm hopes to integrate with browsers, digital assistants, intelligent marketing, and other application scenarios to enhance user productivity and creativity.

SenseTime, a leading AI startup, announces the launch of their "SenseNova" foundational model sets and declares advancing Artificial General Intelligence (AGI) with the strategy of "big model + large computing power." SenseChat is SenseTime's own LLM and is equipped with capabilities for natural language processing, content generation, automated data annotation, custom model training, reading and comprehending lengthy PDF files, and coding. Furthermore, it supports innovative applications including assisting developers in coding and debugging more efficiently, providing personalized medical advice to users, and extracting and summarizing information from complex documents.

Beijing KunLun Wanwei Technologies, a Chinese web and gaming company, announces their launch of Tiangong 3.5 LLM through the Chinese AI startup Qidian Zhuyuan. KunLun Wanwei will provide hardware support for Qidian Zhuyuan and will begin invitation-based user testing on April 17th. It claims to be China's first domestically developed LLM to achieve "intelligent emergence."

Baichuan-Inc, an AI venture set up by the founder of the search engine Sogou, Wang Xiaochuan, unveiled a large language model called baichuan-7B. Aims to be the Chinese version of OpenAI's ChatGPT LLM and be a disruptive force in higher-level applications.

At the Alibaba Cloud Summit, Zhang Yong, CEO of Alibaba Group and Alibaba Cloud Intelligence, unveiled a proprietary LLM “Tongyi Qianwen” and announced that all Alibaba products will soon be integrated with it. User testing and experience invitations began on April 7th.

On the same day, the Cyberspace Administration of China (CAC) announced draft measures for managing generative AI services. These measures hold providers accountable for the accuracy of data used to train generative AI tools. Additionally, companies will need to undergo security assessments before releasing their AI tools to the public. The CAC emphasized that any content generated by generative AI must align with the country's core socialist values.

May 2023

iFLYTEK has launched its natural language processing (NLP) model "Xinghuo". Apart from a video presentation, there isn't much information available about this “potential rival” to ChatGPT.

June 2023

Baidu introduced its Ernie Bot and has already developed 11 Wenxin industry LLMs, covering fields such as power, gas, finance, aerospace, media, urban development, film and television, manufacturing, and social sciences.

Baidu claims that its Ernie 3.5 model has outperformed OpenAI's GPT-3.5 in general abilities and Chinese-language capabilities, surpassing even the more advanced GPT-4.

China is home to at least 79 large AI models with more than 1 billion parameters, most of them focusing on language and visual recognition, according to a recent report published by the Institute of Scientific and Technical Information of China, a government research agency.

Conclusion/Key Takeaways

The mentioned tensions, which include changing global perceptions of China’s ability to develop advanced technologies, China’s corresponding domestic posture, and concerns surrounding her ability to acquire AI end-use technology, will undoubtedly continue to shape the landscape in which this sector’s progress unfolds.

But while Western media has generally portrayed the US-led coalition's efforts to hinder China's access to advanced chips to be a step in the right geopolitical direction, I find it unlikely many DC strategists foresaw China's remarkable resilience and adaptability in response to this challenge.

Structural constraints like export controls and sanctions, while slowing China’s development in the field, have not stymied innovation. Instead of bemoaning their inability to import sophisticated chips, Chinese innovators recognize it merely as a structural change and work to reassess and adapt to their new constraints.

I think this reflects a cultural point that Azeem Azhar made in Exponential View. Azhar was surprised that East Asians, particularly the Chinese, do not espouse any of the AI fatalism that has gripped both the Western corporate suites and the general public.

The AI fatalism gripping the West today is a product of the inherent and incessant sense of opposition inherited through our myth and literature. This is not to say the Chinese have no sense of opposition, but rather their myth, literature, and culture emphasize how to create harmony out of the nature they are part of, rather than conquering the nature of which they are not part. Think about it: Western myth and literature often depict man in conflict with nature, God, and even against his kin even before Genesis ends! Confucianism, Daoism, and Buddhism on the other hand emphasize the creating harmony aspect within society, nature, and oneself respectively – thus the lack of fatalism in the East.

We can see governance in China attempting to create harmony even today with its slew of AI-related legislature. It will be interesting to see if these controls create any harmony for the innovation going forward in Q3, and whether or not it will help carve out a safer sector for AI to grow in going forward. It will also be interesting to see how Q3 develops as more LLMs get rolled out – especially as they’re marketed to China’s corporate sector. Whether or not they will be able to be monetized, real monetization of LLMs will probably happen in China before other countries thanks to its relatively mature legal environment and aims toward business application/voice assistant implementation, as opposed to general public services.

Written by Anthony Edwards