•Don’t hold your breath
By The Economist
South of the Huai river few geese can be seen through the rain and snow.” In classical Chinese this verse is a breakthrough—not in literature but in computing power. The line, composed by an artificial intelligence (ai) language model called Wu Dao 2.0, is indistinguishable in metre and tone from ancient poetry. The lab that built the software, the Beijing Academy of Artificial Intelligence (baai), challenges visitors to its website to distinguish between Wu Dao and flesh-and-blood 8th-century masters. Anecdotal evidence suggests that it fools most testers.
The system, whose name means “enlightenment” and which can emulate lowlier types of speech, derives its power from a neural network with 1.75trn variables and other inputs. gpt-3, a similar model built a year earlier by a team of researchers in San Francisco and deemed impressive at the time, considered just 175bn parameters. As such Wu Dao represents a leap in this type of machine learning, which tries to emulate the workings of the human brain. That delights fans of classical literature—but not as much as it does the Communist authorities in Beijing, which have put ai at the heart of China’s technological and economic master plan first set out in 2017. It spooks Western governments, which worry about ai’s less benign applications in areas like surveillance and warfighting. And it intrigues investors, who spy a huge business opportunity.
On the face of it, the plan is off to a good start. The logistics arm of jd.com, an e-commerce group, operates one of the world’s most advanced automated warehouses near Shanghai. In May Baidu, China’s search giant, launched driverless taxis in Beijing. SenseTime’s “smart city” ai models—urban surveillance cameras that track everything from traffic accidents to illegally parked cars—have been deployed in more than 100 cities in China and overseas. China has been deploying more ai-assisted industrial robots than any other country. And in 2020 it surpassed America in terms of journal citations in the field.
The five most prominent listed Chinese ai specialists are collectively worth nearly $120bn (see chart 1). The biggest of them, Hikvision, has a market value of $60bn. SenseTime, which went public in Hong Kong on December 30th, is worth $28bn. Two more are expected to list soon. In 2020 investments in unlisted ai startups reached $10bn, according to the ai Index compiled by researchers at Stanford University. In its prospectus SenseTime forecasts that revenues from ai-assisted image-recognition and computer-vision software, the most mature part of the market, could hit 100bn yuan ($16bn) by 2025, up from 24bn yuan in 2021 (see chart 2).
Look beyond the headlines or Wu Dao’s elegant verses, however, and things look more complicated. Yes, China has made progress on ai, and even the occasional big splash like Wu Dao. But it almost certainly still lags behind America in terms of both investment and cutting-edge innovation. In 2020, three years into the master plan, privately held Chinese ai firms received less than half as much investment as their American counterparts. And a lot of the public and private money pouring into the sector may end up being wasted.
China’s five-year-old ai master plan set out a number of goals. For example, by 2025 the country is to create an industry with global revenues of 400bn yuan, achieve “major breakthroughs” in technology and lead the world in some applications. Five years later it is to dominate the industry (by then worth $1trn in sales), having written its ethical code and set its technical standards, just as Europe and America defined the contours of the Industrial Revolution.
Elements of the Communist Party’s approach are characteristically prescriptive. The Ministry of Science and Technology has instructed China’s tech giants with existing ventures in certain subdisciplines of ai—Tencent in medical image recognition, Baidu in autonomous driving—to double down on these. That said, the plan is less hands-on than some of the country’s other development projects, observes Jay Huang of Bernstein, an investment firm. In the words of Huw Roberts of Oxford University and five co-authors, the blueprint acts chiefly as a “seal of approval” which “derisks” assorted ai initiatives championed by central-government entities, local authorities and the private sector.
In practice, the derisking involves doling out lots of public money. Some of this takes the form of tax breaks and subsidies, as in the “little giants” programme to nurture 10,000 promising startups across various sectors, including ai. Local governments, even in poor rustbelt provinces such as Liaoning in the far north-east, have also dangled similar incentives in front of ai-curious companies.
Another type of support comes from government procurement. Firms do not disclose how much revenue they derive from public-sector contracts. But the share is likely to be significant. Central and local authorities use SenseTime’s surveillance technology. Megvii, which also specialises in image recognition, has extensive dealings with state-owned enterprises.
The state is also investing in ai companies directly. The central government runs several tech-investment vehicles. Local governments are increasingly creating their own, often armed with billions of dollars. Tianjin, a coastal metropolis, announced a $16bn ai fund in 2018.
Government capital is increasingly helping plug a gap left by foreign investors scared away by American sanctions against some of China’s ai darlings, which are seen as being too close to the Communist Party. A fund run by the Cyberspace Administration of China, a regulator, has acquired an undisclosed stake in SenseTime, which last month was hit by another round of American sanctions over its alleged involvement in government repression of the Uyghur ethnic minority. (SenseTime says that the sanctions are based on a “misperception” of its business.) A separate vehicle, the Mixed-Ownership Reform Fund, accounted for $200m of the $765m that the firm raised in its initial public offering (ipo). Local governments chipped in another $220m.
Lost in translation
State dosh, combined with access to plentiful public data, has helped turn Chinese ai firms into powerhouses in certain niches. According to Bain, a consultancy, by last June the cloud division of Alibaba, China’s e-commerce behemoth, was offering 62 ai-enabled services, from voice recognition to video analytics, compared with 47 from its closest Western rival, Microsoft. SenseTime and Megvii mass-produce computer-vision software and hardware that can be adapted to and installed in individual factories. Despite being locked out of most Western markets by the American sanctions, SenseTime raked in 762m yuan in overseas revenues in 2020, compared with 319m yuan two years earlier, mostly from South-East Asia.
For all these successes, though, China’s ai industry trails the West in important ways. Despite leading America in the overall number of ai-related publications, China produces fewer peer-reviewed papers that have academic and corporate co-authors or are presented at conferences, both of which are typically held to a higher standard. It ranks below India, and well below America, in the number of skilled ai coders relative to its population. These shortcomings are likely to persist, for three reasons.
First, capital may not be being allocated efficiently. It is unclear, for example, how much of Tianjin’s $16bn kitty has actually been deployed. More damaging, Beijing has created a system for rewarding local officials that favours debt-fuelled spending and seldom punishes wastefulness.
Many state ai investments have been “reckless and redundant”, says Jeffrey Ding of Stanford University. Zeng Jinghan of Lancaster University has documented the rise of firms that falsely claim to be developing ai in order to suck up subsidies. One analysis by Deloitte, a consultancy, estimated that 99% of self-styled ai startups in 2018 were fake. Such boondoggles not only burn through public cash, Mr Ding notes, but also consume scarce human capital that could more usefully have been deployed elsewhere.
China’s second problem is its inability to recruit the world’s best ai minds, especially those working on high-level research. A study in 2020 by MacroPolo, a Chicago-based think-tank, showed that more than half of top-tier researchers in the field were working outside their home countries. America and Europe look more appealing to such footloose brainboxes, including many Chinese ones. Though about a third of the world’s top ai talent is from China, only a tenth actually works there. A shortage of non-Chinese researchers further handicaps China’s capabilities, notes Matt Sheehan of the Carnegie Endowment for International Peace, a think-tank in Washington.
Even more problematic for the party, its master plan ignored the cutting-edge semiconductors that power ai. Since its publication Chinese companies have found it ever more difficult to get their hands on advanced computer chips. That is because virtually all such microprocessors are either American or made with American equipment. As such, they are subject to restrictions on exports to China put in place by Donald Trump and extended by his successor as president, Joe Biden. It will take years for Chinese companies to catch up with the global cutting-edge, if they can do it at all.
These challenges will continue to bedevil all of China’s high-tech industries for years to come. It could leave its ai businesses stuck in a rut—successfully rolling out relatively unsophisticated products while trailing Europe and America in paradigm-shifting developments of greater financial and strategic value. Consider Wu Dao 2.0. Although it was a huge improvement on gpt-3, it did just that—improve an existing technology rather than break new ground. No amount of Chinese taxpayers’ money is likely to change that.
Credit| The Economist