“By some metrics, China now produces more high-quality research papers in the [AI] field than the U.S. but still lags behind in ‘paradigm-shifting breakthroughs’ … In generative AI, the latest wave of top-tier AI systems, China remains one to two years behind U.S. development and reliant on U.S. innovations, China tech watchers and industry leaders have said.”
Karen Hao[1]
“Unlike the West, where anything goes on the web, China’s censors insist on strict political guidelines for CPC-conforming information dissemination. Chinese netizens are unable to pull up references to the decade-long Cultural Revolution, the June 1989 tragedy in Tiananmen Square, human-rights issues in Tibet and Xinjiang, frictions with Taiwan, the Hong Kong democracy demonstrations of 2019, pushback against zero-COVID policies, and much more… This aggressive editing of information is major pitfall for a ChatGPT with Chinese characteristics.”
Stephen S. Roach[2]
“Whoever controls information controls the world.”
Jing Tsu[3]
Today’s techno-dilemma faced by Chinese Communist Party technocrats, functionaries and decision-makers rests on two horns; first, Beijing’s leaders—led by President Xi Jinping—are chaffing to make China the self-sufficient center, the Middle Kingdom, of the technological universe; but, secondly, the announcement and exponential speed of developments in ChatGPT research and development by U.S.-based companies (as well as the increasing industry reliance on super-semiconductor chips) demonstrate how far behind they are in the race for technical dominance.
In the interest of saving space, this missive will focus on the second dilemma horn: the hurdles Chinese high-tech firms must overcome if they are to achieve AI global dominance. What is at stake? In the words of veteran China tech observer Kai-Fu Lee, former president of Google’s China operation: “developing big models of AI is a historic opportunity that China cannot afford to miss.”[4]
But China still has a long way to go. In today’s exponentially changing world, the rapid development of large language models (LLMs) such as ChatGPT and its proprietary cousins confronts Chinese tech-strategists with a clear conundrum: as one seasoned China watcher observes, “the harder the Party tries to control ChatGPT content, the smaller the resulting output of chatbot-generated Chinese intelligence will be—yet another constraint on the AI intellectual revolution in China.”[5] Of course, that did not stop Chinese firms like Baidu, Alibaba, and Kunlun from unveiling their own chatbot versions, thereby opening the latest round in the ongoing U.S.-China tech war.[6]
But there remain significant hurdles to overcome. Take, for example, Baidu’s experience. Baidu (literally “hundred times”), headquartered in Beijing’s Haidian District, is one of the largest AI and Internet companies in the world, employing some 45,000 workers. The company was incorporated in January by U.S.-educated Yanhong (Robin) Li and Eric Xu. At the present time, Baidu holds a 76% market share in China’s search engine market (with its Baidu App, based on an AI-powered Baidu Core) reaching 544 million MAUs in December 2020, based on company statistics.[7]
Baidu’s chatbot, Ernie Bot (Wenxin Yiyan—short for “Enhanced Representation through kNnowledge IntEgration”), was launched in March 2023, as Beijing’s answer to the November 2022 announcement of OpenAI’s chatbot—an event some observers in China call the “iPhone moment for AI.”[8] But it was soon apparent that Baidu’s version of the LLMs that fueled OpenAI’s ChatGPT-4 and GPT 3.5 trailed significantly far behind its American counterparts in terms of training power and performance. (NOTE: None of OpenAI’s chatbots are officially available in mainland China).[9]
Even within China, Ernie Bot’s rolling-out ceremony clearly fell far short of expectations. During the launch stream presentation, Robin Li, Baidu’s CEO, admitted the chatbot’s answers to questions and images were prerecorded.[10] The rollout was immediately accompanied by an overwhelming wave of disappointment in China: Chinese publications with testing access ridiculed the chatbot’s performance, social media users mocked it with memes, and Baidu’s stock dropped by 6.4%. (The market value has since recovered a bit as observers realized Ernie Bot fell way short in the “wow” category but was probably good enough for the Chinese market).[11] At the same time, Baidu clearly gained a first-mover advantage in China with its public rollout.
What happened? A large part of Baidu’s problem, and of LLMs in general in China, is that state censorship encumbers China with small language models because chatbots are forced to operate in “a firewalled Internet ruled by government censorship.”[12] Indeed, China’s first ChatGPT-style bot, dubbed ChatYuan and launched in January 2023, by a small Chinese startup called YuanYu (operating as a “mini-program” inside WeChat), was suspended within weeks after users posted screengrabs of its answers to political questions online.[13] Baidu’s predecessor to Ernie Bot, ERNIE-ViLG, suffered much the same fate when released as a demo in August 2022. The model was part of Wenxin, a large-scale project in natural-language processing, trained on a data set of 145 million image-text pairs and containing 10 billion parameters (the values that a neural network adjusts as it learns). Just don’t ask about Tiananmen Square or China’s political leaders.[14]
When, for example, Ernie Bot is asked about any politically sensitive issue, or about President Xi Jinping, the answer is “as an artificial language model, I cannot answer that question.”[15] (The chatbot provided the same answer when asked to write a fictional debate between Jack Ma—of Alibaba fame—and President Xi about the role of the private sector in China).[16]
Yet even with these Party-imposed limitations, Baidu’s chatbot holds a lead over other Chinese LLM contenders according to a recent Xinhua think tank survey.[17] The contenders include Alibaba Group Holding’s Tongyi Qianwen (literally “seeking truth by a thousand answers),[18] Kunlun’s Tiangong, voice recognition firm iFlytek’s SparkDesk, and image recognition company SenseTimes’s SenseChat. In another survey—using different LLM metrics—Clue, a Chinese website that follows AI research development, found cybersecurity firm Qihoo 360’s Smart Brain to be the top performing model, followed by iFlytek’s SparkDesk. These two still lagged far behind Western chatbots.[19]
In addition to problems for these high-tech firms caused by an environment of government censorship, Chinese chatbot companies face a lack of access to a new generation of superchip semiconductors, critical to today’s LLMs. Chinese technical consultants note that “China’s domestic AI industry is now generally short of computing chips, and if the U.S. further sanctions China’s chip technology, it will definitely affect computing power development in the short term.”[20] As a result, according to Chinese media reports, China’s high-tech industry is now “scrambling” for advanced chips to support its AI ambitions amid growing interest around ChatGPT chatbots, “but soaring demand, combined with US-led trade curbs, has limited supplies.”[21]
And, make no mistake about it, China is desperate for high performance superchips.[22] As of May 2023, China had at least 79 LLMs with more than one billion parameters, according to official state-affiliated statistics, each of which require hundreds if not thousands of advanced chips to train.[23] TikTok’s Chinese owner ByteDance, for example, has ordered more than one billion dollars’ worth of GPUs from Nvidia this year, (ByteDance is rumored to be testing its own AI chatbot internally), and Baidu has reportedly procured more than 10,000 Nvidia GPUs (equivalent to the amount purchased by Google).[24]
U.S. government export control laws now prohibit chipmakers like Nvidia from selling high performance superchips to Chinese firms. In September 2022, the U.S. government told Nvidia to stop exporting high performance A100 and H100 chips to China and Russia, also restricting sales of AMD’s MI250 Accelerator chip. Washington has also asked Taiwan’s TMSC—the world’s largest high-performance chipmaker—to stop producing chips for Chinese high-tech firms such as Shanghai-based Birentech Technology.[25] These prohibitions are forcing U.S.-based firms to tailor lower-end versions of the chips for China. As a result, there are only about 40,000 Nvidia A-100 chips available in China, according to the country’s leading AI officials.[26]
Why does China continue to rely on AI-enabling superchips made in the West? Why is it so difficult to attain Xi’s goal of self-reliance in this critical technology sector? When pressed for an explanation, Chinese experts point to four “challenges” in being able to catch up: gaps in training models, as discussed above a lack of access to new generation of superchip semiconductors, the availability of data sets, and the uniqueness of the Chinese language. These problems become even more obvious when discussing LLMs. At the present time, China, lags behind in targeted investment (the U.S. has 3.5 times more investment in LLMs), the U.S. also possesses of the half of the world’s LLMs, and the highest repository citation count.[27] Indeed, last year the U.S. outproduced China five-fold in fielding AI machine learning systems.[28]
Compounding these problems is a huge human resource issue—a shortage of AI-related professionals and researchers in China. According to a report in April 2023 by Hong Kong-based Renrui Human Resources Technology, China will face a shortage of 5.5 million AI engineers in 2025 (compared to 4.3 million in 2022) and estimates show that by 2025 only one out of 2.6 AI-related job positions will be filled.[29] As early as 2019, think tank MarcoPolo reported China faced a brain drain problem because most of its AI talents were choosing to stay in the U.S. after their studies (78% of Chinese AI researchers who completed graduate studies in the U.S. remained at U.S. institutions, while 21% returned to China-based institutions).[30]
There also is the issue of tech-related exponential growth. Simply put, developments in ChatGPT technologies—on both sides of the ocean—are racing ahead of control efforts. As such, even in China’s highly centralized system, we see some indications that AI-related developments are outpacing regulators’ efforts to construct meaningful guardrails. In January 2022, the Chinese government proposed a new regulation banning any AI-generated content that “endangers national security and social stability” to cover AI’s like Baidu’s ERNIE-ViLG.[31] More recently, in April 2023, Beijing’s cyberspace regulator announced a series of measures to manage generative AI services. But these measures are still in draft and must be submitted to the proper Party authorities before adoption.[32]
It is worth noting, however, that after Elon Musk returned from his recent visit to China, he mentioned that Chinese authorities recognize the need for oversight and regulation of “risky” technologies such as generative AI.[33] But Beijing’s perceptions of necessary guardrails run through a nationalistic and historic filter that is missing in the West. While in the U.S. we are obsessed with guaranteeing that future families of ChatGPT technologies abide by today’s social code—that is, they must be free of potential race, gender, and other biases—Baidu and others of its ilk are only concerned that they don’t offend Party sensibilities. Certainly, we cannot expect Beijing’s digital engineers, coders, and military strategists to be hampered (or guided) by Western legal, moral and ethical values, such as keeping a “human in the loop” to control AI-enhanced decision-making regarding machines in a battlefield setting.
In my mind, at least, that is a huge difference between the two systems.
Finally, there’s the existential issue, the danger to the human species presented by rapidly growing algorithm-driven AI models such as ChatGPT, or its future techo-permutations. As one AI expert frames the issue: “Instead of asking ‘What can AI do for us?’ we should be asking ‘What can AI do to us?’”[34]
I wonder how many Party cadres in China are asking that question?
[1] Karen Hao, “OpenAI CEO Calls for Collaboration with China to Counter AI Risks,” The Wall Street Journal, Jun. 10, 2023.
[2] Stephen S. Roach, “Opinion: China wants to tightly control AI at home. The technology has other plans,” MarketWatch, Jun. 14, 2023. Roach is the former chairman of Morgan Stanley Asia, a faculty member at Yale University, and author of several books, including Accidental Conflict: America, China, and the Clash of False Narratives (Yale University Press, 2022).
[3] Quote from Jing Tsu, “Kingdom of Characters: The Language Revolution that Made China Modern.”
[4] “Can AI give China the upper hand.”
[5] Roach, “China wants to tightly control AI.”
[6] See, among others, “Can AI give China the upper hand.”
[7] See Baidu Inc’s “Company Overview,” 2023.
[8] “Can AI give China the upper hand.”
[9] “Baidu’s Ernie Bot tops Chinese large language model rankings by Xinhua think tank, but lags OpenAI’s Chat GPT,” South China Morning Post, Jun. 12, 2023.
[10] Will Knight, “China’s ChatGPT Rival Needs to Watch Its Words,” WIRED, Mar. 21, 2023.
[11] Zeyi Yang, “The bearable mediocrity of Baidu’s ChatGPT competitor,” MIT Technology Review, Mar. 22, 2023.
[12] Ibid.
[13] Fan Yang, “AI chatbots with Chinese characteristics: why Baidu’s ChatGPT rival may never measure up,” The Conversation, Mar. 24, 2023.
[14] Zeyi Yang, “There’s no Tiananmen Square in the new Chinese image-making AI,” MIT Technology Review, Sep. 14, 2022.
[15] Cissy Zhou, “Testing Ernie: How Baidu’s AI chatbot stacks up against ChatGPT,” NikkeiAsia, Mar. 22, 2023.
[16] Ibid.
[17] “Baidu’s Ernie Bot tops Chinese large language model rankings.”
[18] See, among others, Shadine Taufik, “Meet Tongyi Qianwen, Alibaba’s answer to ChatGPT,” Tech in Asia, Apr. 11, 2023.
[19] Baidu’s Ernie Bot tops Chinese.”
[20] Ibid.
[21] “China’s Big Tech Firms scramble for advanced chips amid US sanctions and ChatGPT craze,” South China Morning Post, Jun. 14, 2023.
[22] See my missive on the topic of semiconductors as a Chinese strategic vulnerability—Jeemes Akers, “China’s Achilles Heel,” Sep. 2021.
[23] “China’s Big Tech Firms scramble.”
[24] Ibid.
[25] Jeff Pao, “China leads US in tech that matters most: report,” Asia Times, Mar. 4, 2023. The report cites a report by Australia’s Strategic Policy Institute asserting that China now leads in 37 of 44 key technologies. The U.S. maintains a leading edge in the design and development of advanced semiconductor devices, advanced integrated circuit design and fabrication, and in research fields relating to high-performance computing.
[26] “China’s Big Tech Firms scramble.”
[27] “Can AI give China the upper hand.”
[28] Ibid.
[29] Jeff Pao, “China’s artificial intelligence engineer shortage,” Asia Times, Jun. 13, 2023.
[30] Ibid.
[31] “There’s no Tiananmen Square.”
[32] Luc Olinga, “Elon Musk Asks a Question That Haunts the World,” TheStreet, Jun. 13, 2023.
[33] Angela Palumbo, “Elon Musk says China Wants to Regulate AI. The U.S. Is Still Deciding How. BARRON’S, Jun. 6, 2023.
[34] Roman V. Yampolskiy, “I’m an AI expert: Here’s my worst-case scenario,” THEHILL, Jun. 11, 2023.