news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

On August 23rd local time, the US media "Wall Street Journal" found that although Chinese technology companies cannot obtain the world's most advanced artificial intelligence computing chips under Washington's malicious restrictions, they are now actively responding to the exciting message from industry pioneers - these chips are not necessarily needed to continue to promote AI.

According to the report, a large number of Chinese AI start-ups are improving the efficiency of artificial intelligence by writing more efficient code for large language models to solve the problem of lack of advanced chips; some other companies are focusing on building smaller and more specialized models, or adopting training methods that consume less energy and time.

The article quoted industry analysts and executives as saying that although the Chinese market generally lacks semiconductor computing power produced by global leading companies such as Nvidia, doing so could make some differences "meaningless."

On July 5, the World Artificial Intelligence Conference (WAIC 2024) was held at the Shanghai World Expo Exhibition and Convention Center. IC Photo

The Wall Street Journal pointed out that at a time when the United States is suppressing and restricting Chinese companies' access to the most advanced chips, Chinese company leaders are actively seeking to build their own artificial intelligence with efficiency as the goal of the industry consensus.

AI big model unicorn company Zero One Everything has adopted a low-precision training mode to reduce the power consumption and time required to train machine learning models. According to founder Kai-Fu Lee, Zero One Everything does not have a large number of graphics processing units, which forces them to develop very efficient artificial intelligence infrastructure and reasoning engines. The report quoted Nvidia researchers as saying that in the United States, companies including Google are using such models to accelerate model output.

Hong Kong media South China Morning Post pointed out earlier that many analysts attribute China's rapid progress in the field of artificial intelligence in part to its ability to bypass chip restrictions and develop the intelligent computing capabilities required to train local large language models. The US restrictions on the export of advanced chips are prompting Chinese companies to focus on improving the efficiency of artificial intelligence, and China's construction of computing resource infrastructure will also help alleviate people's concerns about the lack of advanced chips.

Still other companies, rather than focusing on creating the biggest and best models, are looking to develop applications in specialized fields.

“If there is no application, only the basic model, whether it is open source or closed source, is worthless.” As Baidu CEO Robin Li said at an AI industry conference last month, KPMG, one of the Big Four international accounting firms, pointed out in a recent report that information from the second quarter of this year showed that Chinese AI investors “are more focused on the practical applications of AI rather than large language models (LLMs)”, such as applications in robots and improving workplace efficiency.

Judging from the quarterly earnings reports in recent weeks, some pioneering companies in some application areas are also emerging.

The Wall Street Journal mentioned that ByteDance, known as China's "app factory", has launched more than 20 AI applications, including chatbots, video production tools, etc. It and local competitors such as Kuaishou have also launched content generation AI applications "Wensheng Video" to compete with OpenAI's Sora. The difference is that although OpenAI was the first to demonstrate this function, Sora has not yet been widely used.

How do these Chinese AI applications perform overseas? According to data from mobile app analysis company Sensor Tower, three of the top 10 AI applications downloaded in the United States this year were developed by Chinese companies.

On July 3, the latest report released by the United Nations World Intellectual Property Organization (WIPO) also showed that China is far ahead of other countries in the world in generative artificial intelligence inventions such as chatbots, and the number of patents applied for is six times that of the United States. Not long ago, a new survey by the American artificial intelligence and analysis software company SAS and Coleman Parkes Research also showed that China is currently in the world's leading position in the application of generative artificial intelligence, which is the latest sign of China's progress in this technology.

The Wall Street Journal quoted industry experts as predicting that smaller models or edge AI models (processing data and algorithms locally on physical devices) that can implement AI functions on smartphones and laptops may become a "game-changing" product in the future.

According to Winston Ma, an adjunct professor at New York University and an advisor to Dragon Global, an investment company focused on artificial intelligence, he believes that this year will be the "year of small models" because small models that use less training data are faster and are beneficial for real-time applications with specific functions. Boris Van, an analyst at Bernstein Research, also added that focusing on edge computing does not mean lowering technical requirements, but just shifting to a field where higher computing power is not the most critical requirement.

In addition, the Wall Street Journal also mentioned that in terms of computing power, the gap between China and Western chip design companies is further narrowing as Beijing promotes independent development in many aspects.

Huawei's Ascend chips have reportedly been adopted by some Chinese tech giants, national research labs and AI data centers, and half of the country's LLMs are said to be trained using the company's chips. In June, at the World Semiconductor Conference held in Nanjing, Jiangsu, Huawei announced a series of test results showing that Huawei's latest Ascend 910B AI chip outperformed Nvidia's A100 chip in many aspects, with a performance lead of 20% in some tests.

On August 9, the Shanghai Institute of Microsystem and Information Technology of the Chinese Academy of Sciences achieved a material breakthrough that can be used to develop low-power chips.

Baidu, Alibaba and Tencent are also using self-developed chips to run their artificial intelligence models and improving engineering and algorithms to make up for the lack of advanced computing power. At the same time, major Chinese companies are also studying how to combine different types of chips to avoid being too dependent on any one product in hardware.

The Wall Street Journal lamented that it seems that China will continue to promote AI development regardless of whether there are advanced chips from the United States. This can be seen from the end of the article, which specifically mentioned a speech made by a Huawei official at the AI ​​conference in July. At that time, the official encouraged his domestic counterparts: "We should not think that we cannot lead in the field of AI without the most advanced AI chips. We in China must abandon this view."