IF YOU LISTEN to the bombastic rhetoric in Beijing and Washington, America and China are engaged in an all-out contest for technological supremacy. “Fundamentally, we believe that a select few technologies are set to play an outsized importance over the coming decade,” thundered Jake Sullivan, President Joe Biden’s national security adviser, last September. In February Xi Jinping, China’s paramount leader, echoed the sentiment, stating that “we urgently need to strengthen basic research and solve key technology problems” in order to “cope with international science and technology competition, achieve a high level of self-reliance and self-improvement”.
No technology seems to obsess policymakers on both sides of the Pacific more right now than artificial intelligence (AI). The rapid improvements in the abilities of “generative” AIs like ChatGPT, which analyse the web’s worth of human text, images or sounds and can then create increasingly passable simulacrums, have only strengthened the obsession. If generative AI proves as transformational as its boosters claim, the technology could give those who wield them an economic and military edge in the 21st century’s chief geopolitical contest. Western and Chinese strategists already talk of an AI arms race. Can China win it?
On some measures of AI prowess, the autocracy pulled ahead some time ago (see chart 1). China surpassed America in the share of highly cited AI papers in 2019; in 2021 26% of AI conference publications globally came from China, compared with America’s share of 17%. Nine of the world’s top ten institutions, by volume of AI publications, are Chinese. According to one popular benchmark, so are the top five labs working on computer vision, a type of AI particularly useful to a communist surveillance state.
At the same time, when it comes to “foundation models”, which give the buzzy generative AIs their wits, America finds itself firmly in front (see chart 2). ChatGPT and the pioneering model behind it, the latest version of which is called GPT-4, are the brain child of OpenAI, an American startup. A handful of other American firms, from small firms such as Anthropic or Stability AI to tech giants like Google, Meta and Microsoft (which part-owns OpenAI), have their own powerful systems. ERNIE, a Chinese rival to ChatGPT built by Baidu, China’s internet-search giant, is widely seen as less clever than most of them (see chart 3). Alibaba and Tencent, China’s mightiest tech titans, have yet even to unveil their own generative AIs.
This leads those in the know to conclude that China is two or three years behind America in foundation-model building. There are three reasons for this underperformance. The first concerns data. On the surface, a centralised autocracy should be able to marshal a lot of it—the government was, for instance, able to hand over its troves of surveillance information on Chinese citizens to companies such as SenseTime or Megvii that, with the help of the country’s leading computer-vision labs, then used it to develop top-notch facial-recognition systems.
That advantage has proved less formidable in the context of generative AIs, because foundation models are trained on the much more voluminous unstructured data of the internet. American model-builders benefit from the fact that 56% of all websites are in English, whereas just 1.5% are in Mandarin or China’s other languages, according to data from the W3Techs, an internet-research site. As Yiqin Fu of Stanford University points out, the Chinese interact with the internet primarily through mobile super-apps like WeChat and Weibo. These are “walled gardens”, so much of their content is not indexed on search engines. This makes that content harder for AI models to suck up. Lack of data may explain why Wu Dao 2.0, a Chinese model unveiled in 2021 by the Beijing Academy of Artificial Intelligence, a state-backed outfit, failed to make a splash despite its possibly being more computationally complex than GPT-4.
The second reason for China’s lacklustre generative achievements has to do with hardware. Last year America imposed swingeing export controls on any technology that might give its main geostrategic rival a leg-up in AI. In particular, that includes the powerful chips used in the cloud-computing data centres where foundation models do their learning, and the chipmaking tools that could enable China to build such semiconductors on its own.
That was a blow to Chinese model-builders. An analysis of 26 big Chinese models by the Centre for the Governance of AI, a British think-tank, found that more than half depended on Nvidia, an American chip designer, for their processing power. Some reports suggest that SMIC, China’s biggest chip manufacturer, has produced prototype chips which are just a generation or two behind TSMC, the Taiwanese industry leader that manufactures chips for Nvidia (see chart 4). But the Chinese firm may only be able to mass-produce chips which TSMC was churning out by the million three or four years ago. A professor at a leading Chinese university laments his country’s weakness in such “basic infrastructure” of AI.
Chinese AI firms are also having more trouble getting their hands on another American export: know-how. America remains a magnet for the world’s tech talent; two-thirds of AI experts in America who present papers at the biggest AI conference are foreign-born. Chinese engineers made up 27% of that select group in 2019. Many Chinese AI boffins studied or worked in America before bringing their machine learnings back home. (Few non-Chinese boffins would consider moving to a police state a wise career move.) The covid-19 pandemic and rising Sino-American tensions are causing their numbers to dwindle. In the first half of 2022 America granted half as many visas to Chinese students as in the same period in 2019.
The triple shortage—of data, hardware and expertise—has been a genuine hurdle for China. Whether it will hold Chinese AI ambitions back much longer is, though, another matter.
Take data. On February 13th the local authorities in Beijing, where nearly a third of China’s AI firms are located, said they were releasing data from 115 state-affiliated organisations, giving model-builders 15,880 data sets to play with. To liberate more data, the central government also wants to dismantle Chinese apps’ walled gardens. Most important, the latest models appear able to transfer learnings from one language to another. In the paper describing GPT-4, OpenAI said that the model performed remarkably well on tasks in Chinese despite the dearth of Chinese source material in the model’s training data. Already Baidu’s ERNIE was trained on lots of English-language data, notes Jeffrey Ding of George Washington University.
In hardware, too, China is finding workarounds. The Financial Times reported in March that SenseTime, which is blacklisted by America’s government, has used intermediaries to skirt the export controls. Some Chinese AI firms are able to harness the computing power of Nvidia’s advanced chips through cloud servers based in other countries. Alternatively, they can simply buy more of Nvidia’s less advanced semiconductors or use them more efficiently with the help of clever software. To continue serving the vast Chinese market, the American company has designed less powerful sanctions-compliant processors. These are between 10% and 30% slower than its top-of-the-range kit, and end up being costlier for the Chinese customers per unit of processing power. But they do the job.
China could partly alleviate the dearth of chips—and of brain power—with the help of “open-source” models. Such models’ inner workings can be downloaded by anyone and fine-tuned to a specific task. Most importantly, that includes the numbers, called “weights”, which define the structure of the model and which are derived from costly training runs. Alpaca, a model built by researchers at Stanford University using the weights from LLaMA, a foundation model created by Meta, was made for less than $600, compared with sums on the order of $100m for training something like GPT-4. Alpaca performs just as well as the original version of ChatGPT on many tasks.
Chinese AI labs could similarly avail themselves of open-source models, which embody the collective wisdom of international research teams. Matt Sheehan of the Carnegie Endowment for International Peace, another think-tank, says that China has form in being a “fast follower”—its labs have absorbed advances from abroad and then rapidly incorporated them into their own models, often with flush state resources. A prominent Silicon Valley venture capitalist is more blunt, calling open-source models a gift to the Communist Party.
Such considerations make it hard to imagine that either America or China could in the long run build an unbridgeable lead in AI modelling. Each may well end up with AIs of roughly similar ability, even if it costs China over the odds to keep up in the face of American sanctions. But even if the race of the model-builders is a dead heat, America has one thing going for it that could make it the big AI winner—its peerless ability to spread cutting-edge innovation throughout the economy. It was, after all, more efficient diffusion of technology that helped America open up a technological lead over the Soviet Union, which in the 1950s was producing twice as many science PhDs as its democratic adversary.
China is, of course, far more competent than the Soviet Union ever was at adopting new technologies. Its fintech platforms, 5G telecoms and high-speed rail are all world-class. But those successes may be the exception, not the rule, says Mr Ding. Particularly, in the deployment of sensors, cloud computing and business software—all complementary to AI—China has done less well.
Although American export controls may not derail all Chinese model-building, they do constrain China’s tech industry more broadly, thereby slowing the adoption of any new technology. Moreover, corporate China as a whole, and especially small and medium-sized companies, is short of technologists who act as conduits for technological diffusion. Swathes of the economy are dominated by state-owned firms, which tend to be stodgy and change-averse. China’s “Big Fund” for chips, which raised $50bn in 2014 with a view to backing domestic semiconductor firms, has been mired in scandals. Many of the thousands of AI startups created in recent years have simply slapped on the AI label in the hope of getting a slice of the lavish subsidies doled out by the state to the favoured industry.
As a consequence, China’s private sector may find it hard to take full advantage of generative AI, especially if the Communist Party imposes heavy regulations to prevent chatbots from saying something its censors do not like. Such handicaps would come on top of Mr Xi’s broader suborning of private enterprise, including a two-and-a-half-year crackdown on China’s tech industry. Although this anti-tech campaign has officially ended, it has left businesses scarred.
The result is a chill in tech sentiment. Last year private investments in Chinese AI startups amounted to $13.5bn, less than one-third the amount that flowed to their American rivals. In the first four months of 2023 the funding gap appears only to have widened, according to PitchBook, a data provider. Whether or not generative AI proves revolutionary, the free market has placed its bet on who will make the most of it. ■