I) Introduction: The race as of today

The California AI bill SB-1047 has ignited a fierce debate among policymakers and the public. Safety concerns are at the forefront of arguments for regulation, with some urging caution before advancing AI technologies. Opponents of regulation argue that such measures could hinder AI progress, posing their own safety risks by potentially allowing China to overtake the US in AI development. The recognition of an AI race between the US and China is widespread among American policymakers, with a significant focus on maintaining AI supremacy as a matter of national security. A US government report highlights the potential dangers of AI-enabled capabilities, stating they could be used for threatening critical infrastructure, amplify disinformation campaigns, and wage war. The report warns that global stability and nuclear deterrence could be compromised by these AI applications. Beyond these immediate threats, the prospect of AGI (Artificial General Intelligence) or transformative AI, capable of revolutionizing society, also poses a significant concern for the US. However, less attention has been given to the existential risk of uncontrolled AI development within the US, exacerbated by competitive pressures. According to a 2023 Expert Survey on Progress in AI conducted by AI Impacts, the median response from 655 AI experts indicated a 19.4% probability that human inability to control future advanced AI systems could lead to human extinction or severe disempowerment. To address these concerns, an accurate assessment of Chinese AGI efforts is crucial. Policymakers must determine whether to prioritize outpacing China in the race to AGI or focus on governing domestic AI labs and guarding against other AI threats from the PRC (People's Republic of China). This essay examines China's competitiveness in general intelligence models and its potential to compete in the AGI race. After analyzing the AI market, particularly the capabilities of Chinese large language models (LLMs), it becomes evident that while Chinese AI development is impressive, its LLMs lag behind. For the future, despite the challenges China faces in scaling up toward AGI, it remains a potential competitor in the race towards transformative AI. This analysis underscores the need for a balanced approach in US AI policy, weighing the importance of maintaining a lead over China against the imperative of robust governance and safety measures at home.

II) PRC vs USA: Current AI

a) Market Size and Growth Projections

Currently, the United States holds a significant size advantage in the AI market. With tech hubs like Silicon Valley attracting both business and talent, the U.S. has become a global leader in AI technology. In 2023, the U.S. artificial intelligence market was valued at USD 37.01 billion. This market is expected to grow rapidly, reaching approximately USD 369.34 billion by 2033, which translates to a compound annual growth rate (CAGR) of 28.83% from 2024 to 2033. Meanwhile, China, though also a global leader in AI, does not yet match the power of the U.S. market. In 2023, China’s AI market was estimated at USD 29.02 billion. Despite impressive growth, it is projected to expand at a CAGR of 20.12% between 2024 and 2029, reaching a value of USD 104.7 billion by 2030. While the U.S. clearly maintains a substantial lead, China remains a notable force in the AI sector. Its steady growth indicates that it will continue to be a significant player, and as such, it poses potential risks and harms that policymakers should carefully monitor.

b) Innovation and Advancement in Chinese AI

China demonstrates a robust capacity for innovation, often surpassing the recognition it receives. As a global leader in technologies such as facial recognition (albeit for controversial uses) and humanoid robotics, China's technological advancements are noteworthy. Moreover, China's role in AI research contributions is substantial. Despite criticisms from both within and outside China regarding their perceived lack of innovation, Chinese researchers are increasingly competitive on the global stage. According to a CSET data brief, the output of highly cited AI publications by Chinese researchers has grown significantly. Over the past decade, the share of top 5 percent AI publications from China has risen from half that of the U.S. in 2010 to achieving parity in 2019. This growth highlights China as a powerful innovator and contributor to the global AI research community. Even though China’s technology operates under the strict control of the Chinese Communist Party (CCP), its impact on AI innovation remains significant. Given this context, China’s AI industry is a formidable force, comparable to the US in leadership. U.S. policymakers have legitimate concerns regarding the potential applications of innovative AI from China, including disinformation, military uses, cyber warfare, and other strategic areas.

c) China is not a Leader in LLMs

When it comes to general intelligence models, particularly large language models (LLMs), the landscape shifts significantly in favor of the United States. According to benchmarks of publicly accessible LLMs, the top 20 models are non-PRC. While it is true that many of China’s LLMs are not publicly released nor have their performance verifiably tested, this lack of transparency creates uncertainty about their true capabilities. For instance, Wu Dao 2.0’s performance remains speculative as its developers have not released the model or whitepapers detailing its training and performance metrics. Based on the current state of available models, it is unlikely that China is on pace to produce transformative general intelligence models before the US. The transformative success of LLMs such as GPT-3 and subsequent models since 2020 suggests that the US maintains a leading edge in this area. Nonetheless, it is plausible that China and Chinese AI companies are investing heavily in scaling up the production of LLMs in response to these developments. However, using the potential of China’s LLMs as a justification to keep extremely large US models deregulated is likely unfounded. As it stands, China is undoubtedly a significant player in AI technology, posing potential concerns for US policymakers. Yet, the race towards artificial general intelligence (AGI) appears to be a different story. Current evidence does not support that China will surpass the US in creating transformative AGI through LLMs.

III) Unique Challenges Facing the Future of PRC LLMs

While the United States currently leads the world in artificial intelligence, China has expressed a strong desire to overtake this supremacy. Since 2016, the Chinese Communist Party (CCP) has aimed to become a global AI leader by 2030. This ambition is reflected in China’s 14th Five-Year Plan (2021-2025), where AI ranks “first among frontier industries,” However, the ability of the PRC to scale up the performance of LLMs faces several unique challenges.

a) Training data

China faces significant bottlenecks in accessing high-quality training data for large language models (LLMs), more so than the US. Many foreign datasets restrict users with Chinese IPs from full access or do not provide services to Chinese users at all. For example, according to Liza Lin of the Wall Street Journal, less than 5% of the data in Common Crawl, a widely used open-source database for training ChatGPT in its early days, is Chinese-language data. Other valuable data sources, from articles on social media platforms to books and research papers, are often inaccessible due to restrictions by internet giants and publishers. This challenge is likely to persist for PRC LLMs. Furthermore, Liza Lin also writes that “most generative AI models in China need to obtain the approval of the Cyberspace Administration of China before being released to the public. The internet regulator requires companies to prepare between 20,000 and 70,000 questions designed to test whether the models produce safe answers, according to people familiar with the matter. Companies must also submit a data set of 5,000 to 10,000 questions that the model will decline to answer, roughly half of which relate to political ideology and criticism of the Communist Party.” This regulatory environment exacerbates the problem of accessing high-quality training data, which China must address to scale up its LLMs. Without sufficient high-quality data, models will struggle to produce reliable outputs, and using lower-quality data as a substitute may lead to "hallucinations" and biases.

b) Compute

Compute power is crucial for creating more intelligent models. Dedicating more computing power to model training can yield significantly better results, as demonstrated by OpenAI’s Sora. The AI race is characterized by a hypercompetitive trend in training run compute of state-of-the-art models, which has doubled every 3.4 months since 2012. China's ability to keep pace with this trend is essential for its success in AI development. However, China faces significant challenges due to fierce export controls that impede its ability to scale up operations. In response, China is focusing on domesticating key parts of the GPU supply chain. 

US export controls in 2023 have severely impacted China's access to advanced semiconductor chips, cutting off certain chips made with US equipment globally and barring exports of Nvidia's A800 and H800 chips to China. This forces China’s LLM labs to find alternatives to Nvidia GPUs, which are the clear first choice for many. The inability to acquire state-of-the-art Nvidia chips means China relies more on local chip production, but this too is hampered by export controls. For instance, the Dutch government has blocked exports of lithography tools from ASML, a key player in GPU production, further constraining China's chip-making capabilities.

In response, China is ramping up domestic investments in advanced chips and GPU supply chain development. State-backed investments are being directed towards companies like YMTC and SMIC, as well as the development of lithography equipment by SMEE. Additionally, China is providing 'computing vouchers' worth between $140,000 and $280,000 to AI startups to subsidize data center costs, combating rising GPU costs due to US sanctions. These vouchers aim to mitigate the difficulties startups face due to the scarcity of crucial Nvidia processors in China. Notably, the Huawei Ascend 910B, a domestically produced GPU, is only around 2-3 times worse in performance per dollar compared to an equivalent Nvidia chip (A100), according to Aschenbrenner. He notes that while the yield of SMIC’s 7nm production and the maturity of China’s capabilities are debated, there is a reasonable chance China could produce these 7nm chips at large scale within a few years.

c) Industrial mobilization

Again Quoting Aschenbrenner:

“The binding constraint on the largest training clusters won’t be chips, but industrial mobilization— perhaps most of all the 100 GW of power for the trillion-dollar cluster. But if there’s one thing China can do better than the US it’s building stuff. In the last decade, China has roughly built as much new electricity capacity as the entire US capacity (while US capacity has remained basically flat). In the US, these things get stuck in environmental review, permitting, and regulation for a decade first. It thus seems quite plausible that China will be able to simply outbuild the US on the largest training clusters.''

d) Algorithmic breakthroughs

The USA holds a significant lead in innovation and success with large language models (LLMs). Key breakthroughs, such as the transformer architecture, which has driven the generative AI and LLM boom in recent years, have primarily been achieved by US AI labs. While it was previously mentioned that Chinese researchers are highly competitive, an analysis found that roughly three-quarters of the Chinese authors in the study currently work outside China and 85% of those work in the US—at tech giants or universities. Talent retention or “brain drain” would likely be a huge problem for attaining a lead over the US in algorithmic complexity and sophistication. As a result, any future breakthroughs in general language models are more likely to originate from the USA, with China continuing its efforts to catch up.

e) Political will

Censorship and political will is a problem for China in terms of building LLMs. The politics of the CCP and the strict control of the media mean that LLMs are fiercely controlled as shown earlier with the ideology question dataset. Adherence to core socialist values restricts the ability for AI labs to include broad datasets that could potentially cause the model to produce anti-CCP content. High-quality data is already a concern as also mentioned earlier. This means that at the moment, LLMs are lacking the necessary freedom in China. For example, very recently in May 2024, the Cyberspace Administration of China announced that it rolled out a large language model trained on Xi Jinping Thought. While this is a concern for Chinese companies at the moment, it is very possible that these regulations are rolled back in the near future due to increased pressure to compete. Especially as US AI is becoming more and more threatening to China’s socio economic stability, they would likely deregulate and accept some level of self-imposed risk. Furthermore, there is already talk of heavily learning about LLM scaling.

At the CPPCC (A CCP conference), a leading Peking University computer scientist and the director of the Beijing Institute for General Artificial Intelligence is advocating for an AGI moonshot to secure China’s global leadership (source).

Censorship and political control pose significant challenges for China in building large language models (LLMs). The Chinese Communist Party (CCP) maintains strict control over media and information, which translates into stringent oversight of LLMs. As highlighted earlier, the requirement for LLMs to adhere to core socialist values restricts the inclusion of broad datasets, potentially leading to the production of anti-CCP content. This adherence is enforced through mechanisms such as the ideology question dataset, further limiting the freedom necessary for developing robust and fully general LLMs. For example, the Cyberspace Administration of China announced the rollout of a large language model trained on "Xi Jinping Thought" in May 2024.  While this points towards the current restrictive environment, it is possible that these regulations could be relaxed in the future. Increased pressure to compete with the US, whose advancements in AI pose a growing threat to China’s socio economic stability, might prompt deregulation and acceptance of some self-imposed ideological risks. 

Additionally, there is a growing advocacy for LLM scaling within China. At the CPPCC (a CCP conference), a leading Peking University computer scientist and director of the Beijing Institute for General Artificial Intelligence has called for an AGI moonshot to secure China’s global leadership in AI. This could potentially cause a shift towards more aggressive and open AI development strategies in the near future.

IV) Conclusion US Policymakers Shouldn't Be Worried, for Now…

While China's ability to scale LLMs is not out of the question, the country faces significant challenges that require substantial attention. Issues like censorship and ideological thought-policing severely hinder China’s ability to leverage high-quality data and fully commit to artificial general intelligence (AGI). Additionally, China struggles with domestic LLM innovation and talent, which limits its potential to surpass US algorithms and architectures. Finally, export controls present major obstacles in the GPU supply chain, preventing China from accessing the most cost-efficient GPUs and components, making scaling difficult.

However, the potential power of nationwide mobilization in China could help overcome many of these hurdles. If the CCP feels sufficiently pressured and threatened, it might roll back ideological constraints. Should China succeed in fully domesticating and scaling up its GPU supply chain, its multi-billion-dollar data centers could operate at cost efficiencies somewhat comparable to those using Nvidia technology. The CCP’s political style could be highly effective in driving an industrial mobilization effort to reach AGI before the US.

As Aschenbrenner points out in “Situational Awareness,” underestimating China in this race could be a significant mistake. Ceding potentially the next industrial revolution to a political adversary would be devastating. The threat of China scaling up is real, but it has not yet materialized. US policymakers should focus on addressing the non-AGI threats posed by China while ensuring that domestic governance structures are robust enough to avoid destabilizing or destroying the nation and the world through AGI and non-AGI threats.