AI passes Go: where next for China’s artificial intelligence ambitions?
Image credit: Getty Images
China has announced ambitious plans to lead the world in artificial intelligence, but it will have to overcome some major obstacles to achieve that. How realistic are its aspirations?
In July 2017, China published its Next Generation AI Development Plan. As analysts such as Jeff Ding of Oxford University point out, it was not a green-field programme. It rather sought to gather and focus a diverse set of existing initiatives in response to what local academics called a ‘Sputnik moment’: the 2016 defeat of the world’s best Go player, Lee Sedol, by a Google-owned AI.
The plan is aggressive and has three main economic milestones:
- By 2020, China’s AI industry should be developing in step with the rest of the world and generating RMB150bn (£17.4bn) from its core operations.
- By 2025, the industry should be a world leader in some key AI segments with the core generating RMB400bn (£46.5bn).
- By 2030, China should be the overall AI world leader, with the core generating RMB1tn.
For context, China’s AI industry was estimated to be worth around RMB15bn (£1.7bn) when the plan was released. The plan often provokes sceptical responses outside China but the country has made significant advances toward the 2020 goal (indeed large parts of it have arguably already been achieved). Notwithstanding the tendency to see Communist goal-setting as authoritarian bluster, much of the criticism is well-founded. More recently, analysts have also been trying to judge the technological and economic implications for the plan and the wider global AI economy of growing Sino-US tensions.
China’s main challenges fall into three broad, interlinked categories: know-how, independence and investment. First, based on the most common global AI benchmarks, China lags behind the US in all but one: data. It has fewer ‘world-class’ researchers, fewer companies, less experience in advanced algorithm development and, increasingly important, less technical prowess in hardware development.
In a ‘first-pass’ ranking of various regions’ AI capabilities, Ding, a researcher at Oxford’s Future of Humanity Institute, awarded the US 33 and China 17 points (100 representing overall global AI capacity). This put China in second place but by some margin.
Second, benchmarking tends to highlight China’s current dependence on foreign technology – as has the addition of Huawei to the US Entity List, a list of organisations to which exports are restricted for security reasons. Made in China 2025 – the earlier, broader plan under whose umbrella national AI ambitions sit – has been designed to seed and, by its conclusion, deliver 70 per cent national self-sufficiency in high technology.
But consider just how even now, Huawei, a world-class player in 5G and AI (largely thanks to its Kirin handset chips), is being forced to rethink its research and product development strategies as it faces constrained access to US-owned and US-derived technology (for example, the ‘open-source’ Android OS and the Arm processor core family).
The Trump administration took similar action in April 2018 against ZTE, China’s other major communications infrastructure company, which brought it to the brink of collapse. It has also banned several Chinese acquisitions on national security grounds (for example, the Canyon Bridge investment fund’s bid for Lattice Semiconductor in 2017).
Threats to the international links maintained by most Chinese AI players are therefore making technological independence a more pressing issue. To get some idea of how deep those links run, consider the case of SenseTime, the country’s leading computer vision company and one of the government’s five ‘national champions’ for AI. It uses Arm IP at the heart of its products, including the Parrots developer framework now being promoted as a ‘local’ alternative to the likes of Google’s TensorFlow and Apache’s Spark.
Third, 2019 has seen mounting evidence that China’s technology sector faces a capital crunch. Government guidance funds operate mostly at the provincial level and look to combine public and private investment. They support technology clusters around cities such as Shanghai, Beijing and Tianjin, with an increasing number targeting AI. But they are finding it harder to attract outside investment, while traditional venture capital (VC) and private equity (PE) funding are also in decline.
A recent analysis by the Chinese Academy of Science and Technology found that VC/PE investment fell 87 per cent in the first quarter of 2019 to RMB241bn (£28bn). The number of VC/PE funds participating in fundraising rounds also fell, by 53 per cent to 110 per cent.
Several factors are thought to be contributing to the slump. The trade war and mounting unrest in Hong Kong are making investors jittery. After a number of financial scandals, China has tightened its asset management regulations. The VC/PE sector has seen phenomenal growth – until recently, Chinese companies were securing angel investment within 10 months of incorporation, against 15 months for US counterparts – and this was always likely to slow.
With China still running the ‘catch up’ stage of its AI race, these are serious hurdles. But they remain just hurdles, not insurmountable obstacles. Moreover, there is even some suggestion that China has already made sufficient foundational progress and is well placed as the AI battleground shifts away from R&D towards commercialisation.
One leading proponent of this China-optimist view is Kai-Fu Lee. Today, he leads an investment group, Sinovation Ventures, but he has a rockstar reputation in Sinotech. It comes from his time as the head of Google China and, before that, founding director of Microsoft Research Asia (whose alumni are scattered throughout the Chinese AI business) and a pioneering researcher in speech recognition.
Lee’s latest book, ‘AI Superpowers’, argues that deep learning and neural network technology is relatively mature (still leveraging ideas described by Hinton, Rumelhart and Williams in 1986), and unlikely to undergo major short-term disruption. With those concepts now supported by enabling software, hardware, data and algorithms, it has reached the stage where it is ready for monetisation.
Lee’s position is that Chinese entrepreneurs are ahead of the game here, based on their existing ‘gladiatorial’ track record in developing online apps and services. For example, where major US app developers have tended to deploy one-size-fits-all products globally, their Chinese competitors might have originally been ‘copycats’ but they have also learned to pay more attention to tuning them to local tastes.
As growth – particularly for increasingly AI-powered mobile platforms – moves from the West into Asia, South America and Africa, the Chinese players therefore have an edge in the market.
Another potentially important advantage for China is that its lead in data to feed machine learning has gone hand-in-hand with the market’s more enthusiastic embrace of ‘super apps’ that can make greater use of ML. A big help here has been Asian consumers’ far greater willingness to embrace these platforms and accept the technologies inside them.
The best-known example is Tencent’s WeChat. What began as a WhatsApp clone has evolved to become so broad a platform that it has been described as much as an operating system on an operating system as an app. It is everything that Facebook wants to be, including the e-wallet system (and the comparison between how ubiquitous WeChat payments have become in Asia with the ongoing controversy over Facebook’s Libra cryptocurrency proposal is pertinent).
An analysis by Tsinghua University estimated that out of 2017’s global AI talent base of 204,575 researchers and engineers, the US had 28,536 and China had 18,232. However, it then made a further distinction by carving out those who could be regarded as world class. It concluded that there are 5,518 in the US, but only 977 for China. That supports the traditional view about the know-how gap.
But Lee’s counter argument is that if China has more applications-facing engineers proportionally, ones with more talent for customisation and with greater access to data, this kind of ranking falls way short of telling the whole story.
If the question of technological prowess is not as simple as might first appear, nor is that of technological independence. In a November 2018 speech, Dr Tan Tieniu, deputy secretary-general of the Chinese Academy of Sciences, reflected on the preceding technology confrontation between Washington and Beijing. “The US ban on ZTE fully demonstrates the importance of independent controllable core, high and foundational technologies,” he says. “In order to avoid repeating this disaster, China should learn its lesson about importing core electronic components, high-end general-purpose chips and foundational software.”
When the US later followed its move against ZTE (subsequently rolled back) by adding Huawei to the Entities List, He Tingbo, chief executive of Huawei’s semiconductor design subsidiary HiSilicon, echoed Tan’s words when she wrote in a note to employees: “All the spare tyres we have built turned to Plan A overnight.”
Nobody believed that Huawei could immediately replace all US-related content overnight. But the company was not caught as unawares or unprepared as some reports suggested.
China generally has speeded up its development of a national technology since Donald Trump took office. AI is a strong indicator.
Alibaba is generally thought of as an e-commerce company. But it is also challenging Amazon in the cloud and building out its AI capacity as another of Beijing’s ‘national champions’. Part of its efforts include a massive investment in quantum computing – it launched a 10 qubit-capacity cloud service in February 2018 and has also developed the Tai Zhang quantum circuit simulator (quantum computing is seen as an important enabler for increasingly complex and processor-hungry machine-learning algorithms).
Meanwhile in the start-up world, companies such as Cambricon and Yitu are developing their own architectures for AI accelerators. Cambricon recently launched the second generation of its cloud-aimed MLU (machine learning unit). Earlier this year, the Yitu Healthcare division announced that it had used natural language-processing techniques to diagnose a wide range of paediatric diseases.
Perhaps most controversial of all, computer vision specialist SenseTime has pulled out of a surveillance venture based in Xinjiang province (the Chinese government continues to use technology developed by another local specialist, Megvii/Face++). The company nevertheless remains a global supplier of surveillance technology, even counting the New York Police Department among its clients even though there are US alternatives.
Analysis of the technological content of these systems will still uncover important western contributions. IP within a system-on-chip’s processors or for the network-on-chip are common examples. But China’s larger players are generally using AI’s nature as a leapfrog technology to, as Dr Tan suggested, push foreign suppliers out of what might be considered core slots.
There is a long way to go but the AI plan and Made in China 2025 had already placed this shift on the agenda. It is now widely thought that the US Huawei ban (and the ZTE one before it) simply accelerated the trend.
What about investment? As mature as fundamental machine learning and neural network technology may now be, developing systems around them remains expensive.
Alibaba is planning a massive $20bn (£16bn) share offering in Hong Kong. There has been plenty of speculation as to where that money could go (and there would be plenty to go around) but some popular guesses include internal AI R&D, acquisitions of AI unicorns, and the establishment of a ‘defence fund’ in case the US chooses to widen its attacks on Chinese technology companies.
The role played by China’s BAT triumvirate – Alibaba alongside Baidu and Tencent – in enabling and financing AI is both massive and increasing. Huxiu.com, a Chinese online science and technology platform, recently published a study into the national AI sector which found that around 60 per cent of the companies within it are linked to one of the three BATs (Baidu: 25 per cent; Tencent: 19 per cent; Alibaba: 16 per cent). Another 5 per cent are connected to Huawei.
It is also worth noting that since Huxiu created this map in January, SenseTime has expanded promotion for its Parrots platform and both iFlytek (the local leader in computer speech) and DJI (the world’s largest drone company) have sought to expand their partner networks. iFlytek is also reportedly looking to raise $300m-$350m (£240m-£280m) of investment from non-US sources. This has been seen as another war-chest application and hedge against any intensification of the Sino-US trade war.
So, as the Chinese AI sector has matured, there are signs of the main commercial players moving to a more prominent funding role, taking up some of the VC slack.
At the same time, analysts such as Ding and Greg Allen, of The Center for A New American Security, point to continuing traditional ‘top down’ support from government, running alongside this private-sector strand. The government, national and local, is investing to deliver on both socio-economic/whole-of-society goals and military ones.
The latter is particularly important. The 2017 AI plan did not emerge from nowhere. It was an attempt to bring together initiatives because of that AlphaGo ‘Sputnik moment’.
Go’s significance in China is sometimes not fully understood in the West. It is not simply the national game. To this day, proficiency on the board is seen as reflecting proficiency in the battlefield. Within a month of AlphaGo’s victory, the People’s Liberation Army had organised a “Workshop on the Game... and the intelligentisation of military command and decision-making”.
China faces many social issues to which it sees AI making a contribution. Smarter cities will reduce energy consumption and thereby pollution. Rapid urbanisation and an ageing population – one that by some estimates could go into an irreversible decline before the end of the next decade – mean there is much China aims to secure from e-healthcare and the promotion of smart manufacturing and agriculture.
However, the concept of the ‘intelligentisation’ of warfare is never far from the discussion.
China claims that it does not want to engage in an AI arms race – but it is developing drone and surveillance technologies, and researching various ways in which new ‘asymmetric’ digital warfare capabilities could help it counterbalance the more traditional capabilities of the West (there is even the argument now that the US’s need to maintain traditional standing armed forces could be a drag on its military capacity).
So, if pure investors are in retreat, here is another alternative that is not merely capable but also driven to take up the slack. One persistent concern remains how efficiently government can invest to stimulate new technology in comparison with the private sector. Progress might well be slower, but funds are likely to persist.
So, three major challenges, but not all ones that can be overcome. And all three – know-how, dependence and funding – are linked by a fourth theme that faces AI globally. Just how ready is AI for commercialisation.
It is becoming noticeable in the West that its own big technology suppliers – not just Arm, but also players such as Intel and Microsoft – are now themselves placing a lot more emphasis in their sales pitches on enabling systems rather than individual breakthroughs.
And there are numerous dedicated AI hardware architectures that are at or close to commercialisation themselves, taking over from the standard GPU and CPU platforms that are seen as being tested to their limits. Google has its Tensor Processing Unit. Graphcore of the UK has its Intelligence Processing Unit. Cambricon and Yitu are very much part of a pretty well evolved sector.
So, perhaps another question arises. Rather than addressing scepticism towards what China’s AI sector wants to do in the future, perhaps the West should be looking more closely at what it is already capable of doing on platforms that are more stable than is often acknowledged.
One of the biggest problems in reporting on China’s AI ambitions is that little of what is published in Mandarin is automatically available in English. Oxford University Rhodes scholar and PhD candidate Jeff Ding researches China’s development of AI at the university’s Future of Humanity Institute, and is working hard to reddress that balance through his ChinAI mailing list.
It captures much of the main Chinese content around AI (including most of that cited in this article) and provides English translations. It has thus become an invaluable resource not just for those keeping a close eye on Beijing, Shanghai and beyond but also developments in AI more generally.
As a supplement to this article, please go online to read an interview with Ding and other supporting articles for which there was not enough space to include here in print. These go into greater depth about such topics as China’s roles in the development of AI standards and ethics, and its ambitions for the so-called ‘intelligentisation’ of warfare.
And, perhaps most important of all, you’ll find a link to sign up for Ding’s mailing list.
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.