How far can China go with AI?

IF listen To the bombastic rhetoric in Beijing and Washington, America and China The battle for technological supremacy unfolds. Last September, President Joe Biden’s national security adviser, Jake Sullivan, said, “Fundamentally, I think a select few technologies will be very important in the next decade.” Supreme Leader Xi Jinping said, “In order to cope with international science and technology competition, achieve a high level of self-reliance, and fulfill social responsibility, we will strengthen basic research and solve major technological problems.” There is an urgent need to do so,” he echoed this sentiment. self-improvement”.

At the same time, regarding the “underlying model” that gives the generative of the topic, AITheir resourcefulness puts America firmly in the lead (see Figure 2).chatGPT The pioneering model behind it, its latest version GPT-4, is an open brain childAI, an American startup. A few other American companies from smaller companies such as Anthropic and Stability AI Google, Meta, Microsoft (OpenAI), which has its own powerful system. Erniechat Chinese rivalGPT What Chinese internet search giant Baidu has built is widely regarded as less smart than most (see Figure 3). Alibaba and Tencent, China’s most powerful tech giants, have yet to reveal their own Generative. AIseconds.

From this, those in the know conclude that China is two to three years behind the United States in building the basic model. There are three reasons for this poor performance. The first is about data. On the surface, a centralized dictatorship should be able to sort out much of it. For example, the government was able to hand over mountains of surveillance information on Chinese citizens to companies like SenseTime and Megvii. Used by leading computer vision labs in the country, it was later used to develop a top-notch facial recognition system.

Its advantages have proven less formidable in a generative context AIThis is because the underlying model is trained on much larger amounts of unstructured data from the internet. US modelers benefit from the fact that 56% of all websites are written in English, while only 1.5% are written in Mandarin or other Chinese languages. I’m here. W.3Techs, an Internet research site. As his Yiqin Fu of Stanford University points out, the Chinese primarily interact with the internet through his mobile super apps such as WeChat and Weibo. Since these are “walled gardens”, much of their content is not indexed by search engines.This ensures that its content is AI A model that sucks. Wu Dao 2.0, a Chinese model launched in 2021 by the Chinese government-backed Beijing Academy of Artificial Intelligence, failed to attract attention despite potentially being more computationally complex than China’s. The reason may be explained by lack of data. GPT-Four.

The second reason China’s generative performance has been so poor has to do with hardware. Last year, the United States imposed strong export controls on any technology that could give its major geopolitical rivals an edge. AIIn particular, it includes powerful chips used in cloud computing data centers where underlying models learn, and chip-manufacturing tools that allow China to build such semiconductors on its own.

This was a blow to Chinese model builders.Analysis of 26 major Chinese models by the Chinese Governance Center AIBritish think tank . Some reports SMICChina’s largest chip maker produced a prototype chip that was just a generation or two behind TSMCis an industry leader in Taiwan that manufactures chips for Nvidia (see Figure 4). But Chinese companies may only be able to mass-produce chips. TSMC Three or four years ago it was being produced in millions. A professor at a top Chinese university laments his country’s weaknesses in such “basic infrastructure.” AI.

Chinese AI Businesses also have more trouble getting to know-how, another American export. America continues to attract the world’s tech talent. two-thirds of AI American Experts Publishing Largest Papers AI Conference is foreign-born. The Chinese engineer made up his 27% of that select group in 2019.many chinese AI Boffin studied or worked in the United States before bringing machine learning home. The COVID-19 pandemic and rising tensions between China and the US have reduced that number. In the first half of 2022, the US issued Chinese students half as many visas as in the same period in 2019.

The triple shortage of data, hardware and expertise is a real hurdle for China.Whether to keep Chinese AI Whether Ambitions will return for much longer is another matter, though.

Information acquisition

take data. On February 13, local authorities in Beijing said nearly a third of his people in China had AI We publish data from 115 state-owned organizations, providing model builders with 15,880 data sets. The central government has previously suggested it may dismantle China’s walled garden of apps, releasing more data…to another.In the paper describing GPT-4, openAI stated that the model performed very well on the Chinese task, despite the lack of Chinese source material in the model’s training data.Baidu already Ernie Jeffrey Ding of George Washington University says it was trained on a lot of English data.

Even in hardware, China is finding workarounds.of financial times SenseTime, which has been blacklisted by the US government, reported in March that it used intermediaries to circumvent export controls.some chinese AI Businesses can harness the computing power of Nvidia’s advanced chips through cloud servers based in other countries. Alternatively, you can buy Nvidia’s less advanced semiconductors or use them more efficiently with the help of clever software. In order to continue to serve the vast Chinese market, the American company designed a less powerful sanctions-compliant processor. has a higher cost per unit of processing power. But they do their job.

China can partially alleviate its chip and brain shortages with the help of an “open source” model. The internal structure of such models can be downloaded by anyone and fine-tuned for specific tasks. Most importantly, it contains numbers called “weights” that define the structure of the model and are derived from costly training runs. Alpaca is a researcher at Stanford University, LLaMAthe underlying model created by Meta was created for less than $600, while the training total was about $100 million GPT-4. Alpaca works just like the original version of ChatGPT in many tasks.

Chinese AI Laboratories can likewise take advantage of an open source model that embodies the collective wisdom of international research teams. Matt Sheehan of another think tank, the Carnegie Endowment for International Peace, says China is a “fast follower.” Chinese laboratories have absorbed advances from abroad and rapidly incorporated them into their own models. A prominent venture capitalist in Silicon Valley, his capitalist is more blunt in saying that the open source model is a gift to the Communist Party.

With such considerations, it is hard to imagine that either the US or China could build an unbridgeable lead in the US or China in the long run. AI modeling.each eventually AIIt has broadly similar capabilities, even if China may sacrifice itself to catch up in the face of US sanctions. But even if the race for model builders is dead heat, America has one thing that can make it big. AI Winners—unparalleled ability to spread cutting-edge innovations across the economy. After all, it was the spread of more efficient technology that helped America open up a technological edge over the Soviet Union, which in the 1950s produced twice as much science as she did. P.timeD.s as its democratic enemy.

Of course, China is much better at adopting new technology than the Soviet Union. Its fintech platform, 5G. Telecommunications and high speed rail are all world class. But these successes may be the exception, not the rule, says Ding. AI— China is not doing very well.

U.S. export controls won’t derail all of China’s model building, but they will constrain China’s technology industry more broadly, thereby slowing the adoption of new technologies. There is a shortage of engineers who can serve as a conduit for popularization. Part of the economy is dominated by state-owned enterprises, which tend to be stocky and change-averse. China’s “Big Fund” for semiconductors, which he raised $50 billion in 2014 to back domestic semiconductor companies, has been plagued by scandals.thousands of many AI Startups Created in Recent Years Just Slapped AI Label in hopes of getting some of the fancy subsidies the country offers to preferential industries.

As a result, China’s private sector may find it difficult to make the most of its generating power. AI, especially when the Communist Party imposes strict regulations to prevent chatbots from saying things censors don’t like. It will come in addition to his extensive reliance on private enterprise. The anti-tech campaign is officially over, but it left a scar on the business.

As a result, sentiment about technology has cooled.Private investment in China last year AI The startup reached $13.5 billion, less than a third of the money that flowed to its American rival. In his first four months of 2023, the funding gap appears to have widened, according to the data provider’s PitchBook.Generative or not AI Proving to be revolutionary, the free market bet on who would make the most of it. ■