The Global War of Models
Patenting like mad and investing billions. Will this be enough for China to become a global artificial intelligence empire?
AI expert and former director at Google, Microsoft and Apple, Kai-Fu Lee could have chosen to work in any corner of the world – and yet confidently picked Beijing as the venue to develop his own business. Founded in early 2023, his startup 01.AI began operating on the down-low, only to showcase Yi-34B, an open-source Large Language Model (LLM), six months later, available to programmers in English and Chinese. It clicked immediately, surpassing leading market models in key metrics, including Meta’s Llama 2.
It’s no wonder that 01.AI has been valued at over one billion dollars within eight months of its founding. While this would not have been possible without the Chinese giant Alibaba, this statistic is highly impressive – especially given that the eyes of the world, fascinated with progress in artificial intelligence, have been fixed solely on Silicon Valley over the past year. After all, this is where OpenAI presented its model (known as ChatGPT3) in the hopes of overtaking the competition in late November 2022. The move was followed by a gush of AI-based solutions.
The relative calm in China’s corner of the world should fool nobody. Artificial intelligence-related efforts have been as intense there as in the United States. Models and solutions from the Middle Kingdom not having succeeded in shining through yet is nothing but a political matter – by no means for the US only, albeit they have been trying to slow down the Chinese pace of technological development: mainly for the Communist Party of China.
Seeking certainty that they will not be used against the government, the Party has muzzled the AI industry with restrictive regulations, only allowing comprehensively tested services to be marketed. Greenlighted solutions are a litmus test for the Middle Kingdom’s aspirations and propensity for investment conducive to – as Kai-Fu Lee himself wrote in his book “AI Superpowers: China, Silicon Valley, and the New World Order” – China winning the battle for the crown of AI leader.
War of the Models
“We had over 130 Large Language Models in China in late July”, boasted Jiang Jie, deputy CEO of Tencent, a tech giant from Shenzhen, in September, during the opening night for Hunyuan – the company’s most recent Language Model, capable of, among other things, generating responses and images, and drafting meeting minutes. Jiang Jie is certain his product surpasses OpenAI’s GTP4 in terms of the capacity for resolving selected mathematical problems or writing long textual content in Chinese. “The war of hundreds of models is on”, he added.
Some of the Chinese LLMs mentioned by Jiang Jie have not moved beyond the research phase, stuck waiting for political approval. Regardless, the Middle Kingdom is currently responsible for 40 per cent of all AI models globally (the US – for half).
Tencent chose a poignant name for its flagship AI product. Referencing an ancient concept of Daoist philosophy and meditative practice, “Hunyuan” stands for primordial chaos or the original heart of the universe. The Shenzhen company intends to occupy its very core, having been the entity behind introducing WeChat, for example, the super-app that’s needed in China for everything: from boarding a bus to paying bills. This is how data is collected on 1.33 billion people.
At the outset, Hunyuan comprised a mere 100 billion parameters, having been trained on the basis of over 2 billion tokens (for any model, the token is the smallest data unit: a word can be a token – or a token set, if longer). Why “mere”? GPT3 supported 175 billion parameters in 2020. Chinese solutions are thus well behind their Western competitors, at least for the time being. The story of ERNIE Bot, an Alibaba model released in late August 2023, is similar (having awaited party approval for nearly six months). It has around 70 million users already, but OpenAI products are used by over 180 million people worldwide.
Yet much seems to suggest that China is just getting started. It is for good reason that they have been an unquestioned power in patenting AI-related innovations for years, registering 29,800 patents in 2022 alone. For the sake of comparison, Japan, South Korea, and the US were awarded 8,800, 7,800 and just over 5,000 patents during that same period, respectively.
Chinese investment in new technologies, artificial intelligence in particular, is impressive as well. According to a recent report by the company of analyst advisors IDC, the related investment outlay is to reach USD 38.1 billion (CNY 273.9 billion) for a future total of 9 per cent of global AI spending.
While data is unquestionably important, it does not specify the Middle Kingdom’s exact stage of development in terms of advanced technological solutions or, most importantly, whether they will – as declared – become the global AI leader by 2049, i.e. the one-hundredth anniversary of the People’s Republic of China.
China’s actual position is a considerably more complex puzzle than simple data would suggest. While the country enjoys multiple advantages, it is also facing enormous issues restricting and delaying its own aspirations.
Open Up or Die
China’s top research institution issued a warning in early December that things would be far from easy. The Chinese Academy of Sciences (Zhōngguó Kēxuéyuàn) based in Beijing sounded the alarm that the state was facing the threat of a “middle technology trap”. The issue of moving it higher in the value chain is closely related to the kind of country China has become under Xi Jinping’s rule. On the one hand, it has been prioritising aggressive development and can afford hugely expensive innovation; on the other – the centralised government makes all decisions. And dictatorships are not exactly famous for attracting eminent inventors, scientists, or experts.
This is why the authors of the Chinese Academy of Sciences’ analysis are appealing to the country to “open up”. The necessity of such a strategy – even if the rest of the world (i.e. the West) continues its decoupling (detachment from the Middle Kingdom) policy – was previously described by, among others, eminent political scientist and commentator Zheng Yongnian from the Chinese University of Hong Kong. “China needs to open its doors to attract international talent, and if it is not capable of attracting European and American scientists, it should at least try to attract scientists from Russia, Eastern Europe, India and other developing countries”, Zheng appealed.
People are increasingly aware that without international cooperation – in research in particular – China does not stand a chance in the global war of models. While such awareness is not necessarily common among members of the political structure, in business circles it most certainly is, as proven by Kai-Fu Lee’s initiative: he brought in one hundred international artificial intelligence experts as his startup staff, including many Chinese engineers that had until then worked in the United States.
Given the lack of international cooperation, the Middle Kingdom’s technological industry is ready to reach for non-standard means of development. In late 2023, information suggesting that ByteDance – the owner of TikTok and its domestic version (Douyin) – began using OpenAI’s technology on the sly to develop an in-house language model. As reported by “The Verge”, in-house documents have dubbed the model “Project Seed”. While the Chinese giant’s data collection-hungry algorithm has given it quite the competitive edge in social media, it has remained well behind market leaders in terms of generative artificial intelligence development. In response to the aforementioned leaked information, OpenAI cut the Chinese giant off from its API (App Programming Interface).
Reasons potentially incensing engineers working on AI development are but a single battle in the war of models Jiang Jie mentioned. And it is a global one.
Not in the Microchip
“Semiconductors are to the industry what the heart is to the human body”, Xi Jinping famously said in Wuhan five years ago, explaining why the state has to allocate all its power and resources to microchip manufacturing. Today, China manufactures a mere 15-16 per cent of microprocessors used by the national economy. The rest are imported. Yet Beijing is aspiring for that share to reach 70 per cent by the end of 2030, offering a broad array of backing to companies interested in microchip manufacturing, including tax breaks, research and development grants, and support in importing components – acquiring foreign competitors, even.
Beijing has allocated a total of over USD 150 billion to a semiconductor manufacturing stimulation campaign. The first signs are already there. Huawei introduced Kirin 9000s, a microchip designed to power AI, in May this year. Yet in a concurrent move, President Joe Biden’s administration imposed new restrictions on selling semiconductors and technologies required to build microchips to China. The most restrictive controls apply to NVIDIA, a US corporation specialising in the manufacturing of microchips used in AI. The official reason? National security concerns. The US believes that the technology could be used to train AI models intended for military purposes. Unhappy with the new restrictions, NVIDIA has been complaining that bans on exporting successive technologies can cost it up to USD 400 million per year. Considering competing with Beijing a matter of strategy, Washington is not impressed by NVIDIA’s grumbling. Nobody’s making an attempt to conceal that technological world domination is the name of the game.
An expert on American realities and the Chinese market, Kai-Fu Lee has proven himself the shrewdest of all. In anticipation of future decisions, he bought a large inventory of microchips for his 01.AI. The start-up began stockpiling semiconductors even before it fully took off. This is how Kai-Fu Lee invested most of the money investors lent him. Today, he can keep testing the Yi-34B to his heart’s content. In “The Art of War”, Sun Bin wrote: “Should victory depend on the equipment and provisions of armies, all one would need is an abacus, without any need to go into battle”, but no army has ever complained about going into battle with proper armaments. ©Ⓟ
Materiał chroniony prawem autorskim - wszelkie prawa zastrzeżone.
Dalsze rozpowszechnianie artykułu za zgodą wydawcy INFOR PL S.A. Kup licencję.