AI Competition: China’s Strategy, Biden’s Wavering Response, and the Implications for AI Safety

This edition of ChinaTalk, I’ll be featuring three takes on AI competition:

• A translation from popular WeChat influencer Boss Dai on his take on the state of AI competition

• Greg Allen on why the US should restrict China’s access to AI chips

• Bill Drexel on the specter of China competition though AI safety memes

---

Want to help figure out how to manage risks from frontier AI systems? The Centre for the Governance of AI is currently hiring for junior and senior research positions. They’re a non-profit organization based in Oxford with team members based in the US and all over the world. They do deep applied research on issues like responsible AI lab behavior, the regulation of the most capable AI models, and compute governance.

I (Jordan) have been super impressed by GovAI’s work. Their analysis has really helped ground the global discussion on AI with serious and informed analysis. Plus, their researchers make great podcast guests! Lennart Heim’s show on the geopolitics of GPUs was one of my favorites this year, and interactions I’ve had with their staff give me every indication they’re good people you’d enjoy working with. This field really needs more China expertise, and I would love to see someone from my audience tackle these questions full-time.

For more information, go to governance.ai/opportunities. Apply by August 9. 

---

How China Sees the AI Competition

I want to start by highlighting a recent viral WeChat article by Boss Dai 戴老板, a leading tech influencer who we’ve translated previously on ChinaTalk (see this piece on Huawei’s Managerial Mojo). In his recent piece, A Few Things About the Chip War 关于芯片战争的二三事, he explores the strategic landscape facing the Chinese semiconductor and AI. The following is an abridged machine translation of the final section of his article where, after recapping Alex Palmer’s New York Times feature on the chip war, he turns his attention to AI. Do note that Boss Dai is not a nationalist outlier and should be taken as a middle-of-the-road viewpoint for the tech sector.

---

It may be fate, or perhaps coincidence, but just as the battle for chips is in full swing, the fourth industrial revolution, driven by AI, has been ignited once again.

At the end of 2022, a shot fired by OpenAI brought a new frontier to carbon-based humanity, and also turned the chip war into a part of a broader contest. In fact, the United States has already restricted Chinese tech companies from acquiring the latest AI chips, and there are reportedly plans to prohibit American capital from investing in Chinese artificial intelligence companies.

As we all know, the early phase (the initiation period) of the industrial revolution is a stage when new technologies emerge in an endless stream, new models dazzle, and new giants sprout. If we cannot keep up with the changes in core technology at this time, the gap may grow larger and larger, missing by a hair could result in a loss of a thousand miles, and it would require more time and money to catch up later.

During the first industrial revolution (steam engine), China was in isolation; during the second industrial revolution (electricity), it was in the tumultuous late Qing Dynasty; during the third industrial revolution (computer), although New China was catching up, its progress was greatly influenced by political movements. [The polite way of saying that the Japanese and Mao got in the way.]

Therefore, every footnote in history tells us: we must not be left behind this time.

[See my recent interview with Jeff Ding for more on this theme. In fact, while new GPTs do raise the possibility of big national divergences, he argues that it is diffusion capacity, rather than being the first to innovate, which determines winners and losers during an industrial revolution.]

Fortunately, China has some foundations in the AI field. Since 2012, Internet giants and startups have been continuously pouring into the AI field. Although it is still the Americans who have discovered this new frontier, as far as we can see, the only country chasing closely behind in the world seems to be China. There’s no need to belittle ourselves on this.

But there are also quite a few unfavorable factors. Firstly, the methodology the United States uses in the chip war will certainly extend to the field of artificial intelligence. However, unlike the global division of labor and ultra-long industry chain of semiconductors, most areas of artificial intelligence are innovative at the software level, the chain is relatively shorter, and the probability of being choked is less than that of semiconductors (not equivalent to none).

Secondly, there is a gap in venture-capital funding. Due to geopolitical factors and the impact of Chinese concept stocks, the open secret is that US dollar funds cannot be raised. The RMB market is greatly influenced by local finance. Therefore, the current wave of AI startups in China is far less than that of the mobile Internet ten years ago, and it is also quite different from the current booming scene in Silicon Valley.

In terms of fundamental innovation, there is still a significant gap domestically. The Bell Labs that invented the transistor and OpenAI, which launched GPT, are essentially both “nurtured” by large companies (the former by AT&T and the latter by Microsoft). China currently lacks such players (albeit government-funded Beijing Academy of Artificial Intelligence and Shanghai AI Lab are potential options).

If we cannot keep pace in the field of innovation, it will result in an awkward situation where “we dare to innovate only if others dare to open source.” [Here the author is being cheeky, making fun of the fact that a lot of Chinese software firms have developed on the back of open-source projects. That said, being the second mover in a future where open-source alternatives dominate or where industrial espionage can make up the gap in innovative capacity is arguably not the worst place to be.]

In the early years of a technological revolution, countries that actively embrace it have a good chance of creating a “time difference” with their competitors. For instance, the British invented tanks in 1915, but it was Germany who first organized an armored army centered around tanks, whose combat capability far exceeded other surrounding countries. Afterward, they swept across the entire European continent with blitzkrieg.

Cooperation between the military and tech departments in the United States has always been smooth — the Minuteman Intercontinental Ballistic Missile was once the earliest major customer of integrated circuits, consuming up to 20% of the shipment volume at one point. And defense suppliers like Palantir and Ghost Robotics in the US have launched military products integrated with the latest AI technology earlier this year. [For all the challenges the US faces in bringing commercial tech into the DoD, those in charge of China’s military-civil fusion efforts wish that they had the system America did.]

Mao Zedong once said, “The Anti-Japanese War cannot be rushed; the Liberation War cannot be delayed.” The same sentence applied to the current scenario would be: “The chip war cannot be rushed; the AI war cannot be delayed.”

In 1900, the global production of steel was 28 million tons, and the production of oil was 20 million tons. By 2022, the two figures had increased to 1.88 billion tons and 204 million tons, respectively. Steel and oil have undoubtedly been the foundational materials for the construction of industrial civilization for over a century. In the twenty-first century, silicon wafers are the new steel and oil.

Before the outbreak of the Pearl Harbor incident, the US had focused its embargo on steel and oil, the backbone of a strong military victory with metallurgical and refining capabilities; after the failure of the Battle of Moscow, Germany was desperate to advance to the Caucasus, with the aim of seizing the Baku oil fields — each era has its strategic resources, and to lose them is to fail.

We are at an unprecedented intersection of the industrial revolution and major-power competition. Respect the rules, maintain awe, and be full of confidence. Dawn will surely come.

Will Biden Stay the Course?

In Washington, however, it seems as though the tide is shifting away from continued escalation of the “protect” side of the tech competition. Recent reporting around outbound investment controls has continued to minimize the final rule and push out the timing. Congress had very little China policy of substance in this year’s NDAA (here’s my Twitter thread summing up the changes, and Derek Scissors with a take I largely agree with), and there has not been much noise around Senator Schumer’s call to put together another China bill.

Of late, America’s leading semiconductor firms took to DC to make the case to, as Bloomberg put it, “study the impact of restrictions on exports to China and pause before implementing new ones” in the hopes of preserving their export markets. Said Pat Gelsinger at the Aspen Security Forum in a veiled threat in the leadup to a presidential election, “Right now, China represents 25 to 30% of semiconductor exports. If I have 20 or 30% less market, I need to build fewer factories.”

Beyond this lobbying, apparently there’s a darker game afoot in Washington to try to personally target those in the Beltway most aligned with aggressive tech controls.

In this context, I wanted to highlight a recent piece by CSIS’s Greg Allen, a former ChinaTalk guest, whose team has been putting out fantastic coverage on a whole range of export control policy, including investment screening, export controls beyond chips, and the dynamics in Japan and the Netherlands.

Some excerpts below from his recent piece making the case for why “Blocking China’s Access to AI Chips Matters to US National Security”:

Like their counterparts in the United States, China’s leaders, including Xi Jinping, general secretary of the Chinese Communist Party, believe that leadership in AI is foundational to the future of economic and military power. China’s 2017 national AI strategy describes AI as “a new focus of international competition.”

The Chinese military embarked on its AI modernization campaign at roughly the same time as the United States in 2017, and they encountered the same problem that Project Maven uncovered: a desperate need for AI chips.

But China’s military faced an additional problem: Chinese companies did not make AI chips, so, the Chinese government embarked upon a two-prong strategy. First, buy the AI chips from the United States (for now). Second, develop Chinese alternatives to the US chips (as soon as possible).

The reality of Chinese military purchases of AI chips is not up for debate. It was openly published in unclassified Chinese military procurement contracts. Researchers from Georgetown University’s Center for Security and Emerging Technology reviewed 21,088 such contracts from April to November 2020 and found that all of the purchases of AI chips specified US products. Not a single contract specified purchasing AI chips from Chinese companies. The same massive parallelization architecture that makes GPUs so attractive for AI applications also applies to other kinds of complex computational work, such as high-fidelity aerospace modeling and simulation for hypersonic missiles and nuclear weapons. A recent Wall Street Journal review of Chinese military procurement records since 2020 found that China’s top nuclear weapons research institute is also a repeat buyer of US GPU chips. The US intelligence community undoubtedly is aware of further — classified — Chinese purchases and plans.

To block China from buying advanced US chips, the export controls apply on a country-wide basis, not just for known Chinese military organizations. That is a direct response to the fact that China’s strategy of military-civil fusion has worked to deepen and obscure the linkages between Chinese commercial companies and the Chinese military. Even though US companies have reliably complied with pre-2022 “no military end user” export control restrictions, resellers networks within China have reliably been able to get the chips to the military. Blocking sales to China as a whole was overdue.

To block China from making Chinese alternatives to US AI chips, the export controls continue to allow sales of chips below certain technological performance thresholds and also restrict the sale of advanced chip-making equipment. Chinese AI chip design firms such as Biren and Cambricon were nipping at the heels of US companies such as Nvidia prior to the October 7 export controls. They were perhaps only a year behind Nvidia in terms of chip design quality. But now, those Chinese chip design companies are cut off from accessing the advanced American semiconductor manufacturing equipment (including the equipment powering chipmaking factories in Taiwan).

In May 2023, a group of AI industry and academic leaders issued a statement warning that the risks of advanced AI should be viewed in the same way as pandemics and nuclear war. None of those risks will be any easier to manage if China achieves its vision of becoming an AI-enabled authoritarian superpower.

Bill Drexel—DC policy wonks & AI safety memes on China

As the temperature rises around chip controls and AI competition more broadly, America’s AI safety enthusiasts and DC policy wonks have increasingly divergent views on China, nowhere more evident than in Twitter memewars. Despite the superficial medium, the misalignment between the two perspectives is important: the AI safety network has surprisingly deep influence across the leading AI labs that policymakers hope will give them an edge over China. Having spent some time trying to explain each tribe to the other, here are the two areas where I see each group most talk past each other­­ — as shown through AI safety memes:

If the above meme looks discordant with reality, it may be because a “race to the bottom” on safety is an abiding fear among AI safety experts — not just internationally, but most of all domestically. The AI-concerned harbor persistent anxieties that competition between OpenAI, Anthropic, Google DeepMind, and other leading labs may create incentives to produce powerful capabilities faster than safety measures can catch up.

The same AI safety groups often project a similar fear on the international stage, and fear that a Cold War–style arms race between the US and China that could imperil humanity. Given that misperceptions between countries have exacerbated previous arms races, the AI safety crowd tends to ignore or downplay the extent of US-China AI competition, despite the evidence cited in this newsletter, for example.

To be fair, there are some key differences between AI companies’ race and the US-China competition: while many of the leading labs are explicitly racing toward a common — if opaque — finish line in the form of artificial general intelligence (AGI), China and the US are competing to stay ahead in a wide range of potential AI applications across political, economic, and military domains. That said, denying the intensity of that more multifaced competition, which has been publicly highlighted by officials from both countries, tends to make the AI safety folks appear hopelessly naïve to many policymakers. The sooner both parties can accept the realities of the US-China AI competition, and work together to mitigate its worst risks, the better.

Relatedly, AI safety folks pushing for greater AI regulation in the US are often confronted by lawmakers fearful of quashing innovation and leaving the door open for China to surge ahead. In response, they often point to China’s recent push toward instituting its own AI regulations faster than the US has. Here again, the AI safety crowd lacks important context: as China watchers know well, laws in China are highly flexible on implementation, and subservient to the interests of the CCP at any given moment.

As with the internet boom, China’s regulations are likely geared to ensure that China builds its own AI companies that are competitive with Western counterparts and fully acquiescent to Party rule. Unfortunately for the AI safety crowd, their boosting of the CCP’s AI rules appears to many in DC as playing directly into the hands of the CCP, who have a history of wanting to project international leadership and responsibility on paper while disregarding their own laws at home. A better approach would be to focus on instituting regulations that would make the US more competitive vis a vis China by requiring more reliable, explainable capabilities than Chinese counterparts.

To be sure, the two communities do not diverge on everything related to China. Both, for example, have strong stated interests in shoring up leading AI labs’ cyber security from Chinese intrusions. But in the years ahead, the two groups will need to learn to communicate better on China’s AI ambitions if they are to succeed in their goals.