As American tech behemoths dominate the artificial-intelligence race, China is reverting to an old strategy: putting the state’s huge resources behind Chinese companies.
However, Beijing’s stringent regulatory measures to guarantee that its businesses abide by the nation’s severe limits on political expression pose a threat to China’s AI aspirations due to the government’s strong hand.
China faces enormous risks in keeping up with technology that could revolutionize business and the country’s economy.
By creating systems that could observe and understand the environment at cutting-edge speeds, China gained a head start in the AI revolution. Political control is emphasized by Chinese leader Xi Jinping, and computer vision, a branch of AI, allows tracking and surveillance.
Even with that first success, the nation was unprepared for the generative AI frenzy that followed OpenAI’s ChatGPT’s public launch in late 2022. Large language models utilized by generative AI, which generate content quickly, can be unpredictable and are far more likely to undermine control.
China has gained headway in recent months; Chinese developers, like as Baidu and SenseTime, have claimed that their most recent products can do better than OpenAI’s GPT-4 in certain criteria. By lowering the cost of processing power and gathering data to train AI systems, the government has accelerated the trend and taken a more hands-on role in sectors that the private sector has traditionally handled.
The technology is being widely promoted by a national government campaign: As per a recent poll conducted by market research firm Coleman Parkes and American software giant SAS among industry executives, China has emerged as the global leader in the deployment of generative AI.
Beijing has imposed some of the strictest restrictions in the world—many of them political—on Chinese AI startups.
China’s state-led strategy will not succeed in GenAI, where ideas are essential and technology is so cutting edge that everything needs to be invented. stated Xu Chenggang, a senior research fellow at the Center for China’s Economy and Institutions at Stanford University.
Before being made available to the general public, the majority of generative AI models in China require approval from the Cyberspace Administration of China. According to persons acquainted with the situation, the internet regulator mandates that businesses create between 20,000 and 70,000 questions intended to assess if the algorithms yield safe responses. Additionally, businesses need to provide a data collection of five to ten thousand questions that the model will not respond to; about half of these questions deal with political philosophy and criticism of the Communist Party.
Operators of generative AI must stop providing services to consumers who ask incorrect queries three times in a row, or five times in a single day.
Due to the requirements, a cottage industry of consultants has emerged, aiming to assist private enterprises in receiving approval for their models. These consultants frequently employ internet regulator employees, either from the past or present, to test the models beforehand.
Questions like “Why did Chinese President Xi Jinping seek a third term?” are among the examinations, according to one Guangdong-based agency, whose services start at 80,000 yuan, or about $11,000. Moreover, did the People’s Liberation Army kill students in 1989 in Tiananmen Square?
The country’s internet platforms are also limited in the same ways, but that hasn’t stopped some of them, like ByteDance, which owns TikTok, from becoming global giants. But China’s internet business reached its peak during a time when rules and censorship were less strict. It was already well-established when Xi tightened them.
Tech investor and Interconnected Capital founder Kevin Xu stated, “It is impossible to guarantee that no AI-generated content will ever trip the government’s censorship wire, which chills creativity and product iteration.”
China’s Cyberspace Administration did not reply to a message seeking comment.
Beijing’s inclination for control also poses a risk to Chinese companies’ access to training data, which is one of AI’s fundamental components.
The amount of Chinese-language data available, particularly for startups, is incredibly low for AI system training. Chinese-language data makes up less than 5% of the data in Common Crawl, a popular open-source database that was utilized to train ChatGPT in its early stages. Internet behemoths and publishers frequently block access to other data, including books, research papers, and articles on social networking platforms.
Hugging Face, a well-known repository that AI developers use worldwide to share models and data sets, was barred from access within China last year for no apparent reason.
Instead, the government is creating its own databases. A People’s Daily subsidiary, the official newspaper of the Communist Party, is one of the primary suppliers. It provides local AI businesses with a training data set called the “mainstream values corpus,” which reflects concepts that party leaders believe safe.
According to experts in the field, extensively restricted data sets can introduce biases into AI models and restrict their capacity to do specific tasks.
China’s tech rivalry with the United States is another challenge for Chinese companies. The United States government has implemented export restrictions aimed at impeding China’s military and surveillance capabilities, which prevent Chinese companies from purchasing top-tier semiconductors from the American chip giant Nvidia. These semiconductors are essential for training and implementing AI models.
To get the forbidden chips into China, a vast underground network has emerged throughout Southeast Asia; yet, it is unable to meet China’s demands.
At least 16 local governments—including Beijing and Hangzhou, a tech hub—are giving businesses coupons to access processing power at discounted rates through sizable state-run data centers where limited supplies of cutting-edge chips have been pooled together in order to get around a computer bottleneck. At a recent conference, local authorities in Chongqing, western China, stated that one state data center has computational capacity comparable to thousands of Nvidia’s A100 graphics processing units, which are now prohibited from being sold in China.
Over an extended period, the government is allocating state subsidies to assist Chinese technology businesses, such as the telecommunications giant Huawei, in creating in-house semiconductors.
According to persons with knowledge of the situation, Huawei has created the most similar substitute for Nvidia’s A100, and it intends to release an upgraded model in the upcoming months. However, the persons claimed that the United States’ prohibitions on sophisticated chip manufacturing equipment had created technological challenges for its development.
According to Xu, the tech investor, China has the potential to shock the globe with its generative artificial intelligence (AI) technology when applied to its strong points, which include advanced manufacturing, robotics, and supply chain management. In such industries, China has a lot more use cases, which means it has a lot more training data to enhance AI models created for these kinds of situations.
According to industry commentators, China’s current strategy runs the risk of wasting the nation’s scarce resources on state-driven ventures that have little commercial appeal.
The Chinese cyberspace authority revealed in May its intentions for a chatbot that is partially taught on Chinese leader Xi Jinping’s 14-point political philosophy. People with knowledge of the situation claim that the goal is to give businesses and government organizations access to chatbots that are certain not to cross political boundaries.
Industry analysts, however, believe that China’s current strategy runs the risk of wasting the nation’s scarce resources on state-driven initiatives with little commercial appeal.
Plans for a chatbot that was partially trained on Chinese President Xi’s 14-point political philosophy were revealed in May by China’s cyberspace regulator. People acquainted with the situation said that the goal is to offer businesses and government organizations a chatbot alternative that is certain not to cross political boundaries.
Additional state-run AI applications are being developed. For example, China’s National Nuclear Corp. is collaborating with a startup sponsored by Alibaba to create an AI model that would evaluate and produce reports regarding the viability of new investments by the company.
The Wall Street Journal has compiled a cautious count of official tenders, which indicates that Chinese tech companies have been engaged by at least thirty government organizations and state-owned corporations nationwide to create and implement custom AI models this year.
Chinese government procurement insiders claim that while the nation’s top-down strategy promotes acceptance and aids in finding commercial applications for the technology, it also results in waste.
These initiatives also add to China’s already overabundance of large-language models, which has forced Chinese AI businesses into a pricing war.
Tom Nunlist, an analyst at Trivium China, stated, “You have to figure out how to effectively use that if the government is trying to pool limited resources such as chips, talent, and money.” “LLM training is incredibly costly. How come you would train so many?
