If you listen to the bombastic rhetoric in Beijing and Washington, the US and China are engaged in an all-out competition for technological supremacy. “Fundamentally, we believe that a select few technologies are set to play a significant role in the coming decade,” Jake Sullivan, national security adviser to President Joe Biden, thundered last September. In February, China’s paramount leader Xi Jinping echoed the sentiment, saying that “we urgently need to strengthen basic research and solve major technology problems” in order to “face up to international science and technology competition, achieving a high degree of self-reliance and self-improvement”.
No technology seems to influence policymakers on both sides of the Pacific more right now than artificial intelligence (AI). Rapidly improving capabilities of “generative” AI like ChatGPT, which analyzes the value of the web’s human text, images or sounds and can then create increasingly passable simulacrums, have only strengthened the passion. If generative AI proves as transformative as its boosters claim, the technology could give those who wield it an economic and military edge in the key geopolitical contest of the 21st century. Western and Chinese strategists are already talking of an AI arms race. Can China win it?
On some measures of AI skill, absolutism progressed some time ago (see Chart 1). China overtakes US in share of highly cited AI papers in 2019; 26% of AI conference publications in 2021 globally came from China, while the US accounted for 17%. Nine of the top ten institutions in the world by volume of AI publications are Chinese. There are also five of the top labs working on computer vision, according to one popular benchmark, a type of AI particularly useful to a communist surveillance state.
At the same time, when it comes to the “foundation model” that lends its intelligence to buzzy generative AI, the US finds itself firmly in the lead (see Chart 2). ChatGPT and the pioneering model behind it, the latest version of which is called GPT-4, is the brainchild of OpenAI, an American startup. A handful of other US firms have powerful systems of their own, from smaller firms like Anthropic or Stability AI to tech giants like Google, Meta and Microsoft (which partly owns OpenAI). ERNIE, a Chinese rival to ChatGPT created by China’s Internet-search giant Baidu, is widely seen as less clever than most of them (see Chart 3). Alibaba and Tencent, China’s most powerful tech titans, have yet to unveil their own generative AI.
This leads those in the know to conclude that China is two or three years behind the US in foundation-model building. There are three reasons for this poor performance. The first concern is data. On the surface, a centralized autocracy should be able to marshal a lot of it—for example, the government was able to hand over its trove of surveillance information on Chinese citizens to companies like SenseTime or Megvii, with the help of the country’s leading computer— Vision Lab, then used it to develop a top-notch detection system.
This advantage turns out to be less formidable in the context of generative AI, as foundational models are trained on vast amounts of unstructured data from the Internet. According to data from Internet-research site W3Tech, American model-makers benefit from the fact that 56% of all websites are in English, while only 1.5% are in Mandarin or other languages of China. As Yiqin Fu of Stanford University points out, the Chinese interact with the Internet primarily through mobile super-apps such as WeChat and Weibo. These are “walled gardens”, so most of their content doesn’t get indexed on search engines. This makes it difficult for AI models to capture that content. A lack of data may explain why a Chinese model unveiled in 2021 by the state-backed organization, the Beijing Academy of Artificial Intelligence, Wu Dao 2.0, despite being potentially more computationally complex than GPT-4, failed to make a splash .
Another reason for China’s weak manufacturing achievements is related to hardware. Last year, the US imposed extreme export controls on any technology that could give its main geostrategic rival a leg-up in AI. In particular, this includes the powerful chips used in the cloud-computing data centers where the Foundation models its studies, and the chip-making equipment that could enable China to make such semiconductors on its own.
This was a blow to Chinese model-makers. An analysis of 26 large Chinese models by the British think-tank Center for the Governance of AI found that more than half relied on Nvidia, an American chip designer, for their processing power. Some reports suggest that SMIC, China’s largest chip maker, has produced prototype chips that are just a generation or two behind TSMC, the Taiwanese industry leader that makes chips for Nvidia (see Chart 4). But the Chinese firm may only be able to mass-produce the chips that TSMC was churning out by the million three or four years ago. A professor at a prominent Chinese university laments his country’s weakness in such AI “infrastructure”.
Chinese AI firms are also having more trouble getting their hands on another American export: tech know-how. America remains a hotspot for the world’s tech talent; Two-thirds of US AI experts who present papers at AI’s largest conferences are foreign-born. Chinese engineers made up 27% of that select group in 2019. Many Chinese AI boffins studied or worked in the US before bringing their machine learning back home. (Some non-Chinese boffins consider moving to a police state a wise career move.) Their numbers are dwindling due to the Covid-19 pandemic and rising Sino-US tensions. In the first half of 2022, the US granted half as many visas to Chinese students as in the same period in 2019.
The triple deficit of data, hardware and expertise has been a real stumbling block for China. However, it is another matter whether this will keep Chinese AI ambitions in check for long.
getting information
Collect data. On February 13 local officials in Beijing, where about a third of China’s AI firms are located, said they were releasing data from 115 state-affiliated organizations, giving model-builders 15,880 data sets to play with. Are. Kayla Blomquist, a former US diplomat in China at the University of Oxford, says the central government has previously indicated it wants to wall off Chinese apps, potentially freeing up more data. Most important, the latest models are able to transfer learning from one language to another. to another. In a paper describing GPT-4, OpenAI stated that the model performed remarkably well on the Chinese tasks, despite the lack of Chinese source material in the model’s training data. Already Baidu’s ERNIE was trained on a lot of English-language data, says Jeffrey Ding of George Washington University.
China is also looking for solutions in terms of hardware. financial Times It was reported in March that SenseTime, which has been blacklisted by the US government, used middlemen to avoid export controls. Some Chinese AI firms have been able to access the computing power of Nvidia’s advanced chips through cloud servers located in other countries. Alternatively, they could simply buy more of Nvidia’s less advanced semiconductors or use them more efficiently with the help of clever software. To continue serving the huge Chinese market, the American company has produced less powerful sanctions-compliant processors. These are 10% to 30% slower than its top-of-the-range kit, and costlier per unit of processing power for Chinese customers. But they do work.
China can partially overcome the shortage of chips and brainpower with the help of “open-source” model. Anyone can download the inner workings of such models and fine-tune them for a specific task. Most importantly, it contains numbers, called “weights”, that define the structure of the model and that are derived from expensive training runs. Alpaca, a foundation model built by Meta by researchers at Stanford University using the weights of LLaMA, was built for less than $600, compared to $100m for training something like GPT-4 The amounts on the order were compared. Alpaca performs similarly to the original version of ChatGPT on many tasks.
Chinese AI labs can similarly take advantage of open-source models that embody the collective wisdom of international research teams. Matt Sheehan of the Carnegie Endowment for International Peace, another think-tank, says China has the makings of being a “rapid follower” – its laboratories have absorbed advances from abroad and then rapidly adapted them into their own models. , often with flush state resources. A prominent Silicon Valley venture capitalist is more blunt, calling the open-source model a gift to the Communist Party.
Such views make it hard to imagine that the US or China can build up an incomparable leadership in AI modeling in the long run. Everyone could end up with AI of roughly the same capability, regardless of the costs China faces at the odds of facing US sanctions. But even though the race for model-makers is a dead heat, the US has one thing going for it that could make it the big AI winner—its unique ability to spread cutting-edge innovation across the economy. Ultimately, it was the more efficient dissemination of technology that helped the US gain a technological edge over the Soviet Union, producing twice as many science PhDs in the 1950s as its democratic adversary.
Of course, China is far more capable of adopting new technologies than the Soviet Union. Its fintech platform, 5G telecom and high-speed rail are all world-class. But those successes may be the exception, not the rule, Mr. Ding says. Particularly in the deployment of sensors, cloud computing and business software—all complementary to AI—China has fared less well.
While US export controls may not derail all Chinese model-making, they do constrain China’s tech industry more broadly, delaying the adoption of any new technology. Furthermore, corporate China as a whole and small and medium-sized companies in particular lack technologists who act as vehicles for technological dissemination. Sectors of the economy are dominated by state-owned firms, which are unstable and change-averse. China’s “Big Fund” for chips, which raised $50 billion in 2014 with a view to supporting domestic semiconductor firms, has been mired in scandal. Many of the thousands of AI startups that have been created in recent years have slapped on the AI label in hopes of getting a piece of the huge subsidies the state is doling out to the favored industry.
As a result, it may be difficult for China’s private sector to take full advantage of generative AI, especially if the Communist Party enforces heavy regulations to prevent chatbots from saying anything its censors don’t like. Such constraints will come because of Mr. Xi’s sweeping crackdown on private enterprise, including a two-and-a-half-year crackdown on China’s tech industry. While this anti-tech campaign is officially over, it has left businesses scarred.
The result is coolness in a technical sense. Private investment in Chinese AI startups totaled $13.5 billion last year, less than a third of the amount that flowed to their American rivals. The funding gap appears to have only widened in the first four months of 2023, according to data provider PitchBook. Whether or not generative AI proves revolutionary, the free market has placed its bets on who will make the most of it.
Source