\n
![](\"https:\/\/fortune.com\/img-assets\/wp-content\/uploads\/2024\/07\/GettyImages-2160026935-e1720714745735.jpg?w=2048\")
<\/p>\n [ad_1]
\n<\/p>\n
Last week, Shanghai hosted China\u2019s largest AI event: The World Artificial Intelligence Conference (WAIC), with 500 exhibitors, 1,500 exhibits, over 300,000 attendees, and even an appearance from Chinese premier Li Qiang.<\/p>\n
But despite its impressive scale, the conference left me disenchanted. I\u2019d hoped to witness the sector\u2019s technological advancements. Instead, WAIC confirmed my suspicions: There\u2019s a gap between what China\u2019s AI can do and the cutting-edge innovations emerging from Silicon Valley.<\/p>\n
WAIC exhibitors focused on robotics and large language models (LLMs), with only a few generative AI companies in the mix. Over half the companies at WAIC, including big tech companies and even some state-owned telecommunications companies, were showcasing their new models.<\/p>\n
In Shanghai, Baidu founder Robin Li encouraged<\/a> attendees to start developing practical AI applications rather than continue to refine their LLMs. He stressed that a powerful and widely-used AI application will benefit society more than another model that can process vast amounts of data yet had no practical use.<\/p>\n The generative AI applications on display in Shanghai were mostly ChatGPT-like chatbots, except for Kuaishou\u2019s text-to-visual application<\/a> Kling, a Sora-like product that I found genuinely impressive.<\/p>\n As I wandered the showroom, I noticed that most chatbots required prompts in English, instead of Chinese. That leads me to suspect that many of China\u2019s AI programs are, in fact, running on models developed outside of China.<\/p>\n It\u2019s obvious that the models still need some fine-tuning. One consumer prompted a text-to-visual app from Moore Threads with \u201ca cute baby boy with brown hair, sitting in the garden.\u201d The result was a baby with bright fuchsia skin, eyes that didn\u2019t align on the face, and a disproportionately small body.<\/p>\n I left the conference agreeing with Alibaba chairman Joe Tsai\u2019s candid admission<\/a> earlier this year that China\u2019s generative AI development is at least two years behind the U.S. That means U.S. and Chinese companies aren\u2019t really playing in the same leagues, and so it\u2019s difficult to directly compare them.<\/p>\n The critical problem is that China\u2019s LLMs are limited to using data within the Great Firewall. As investment bank Goldman Sachs noted late last year<\/a>, \u201cLLM performance improves with scale\u2014more parameters, more and better training data, more training runs and more computation.\u201d There is simply less information in the isolated Chinese-language internet compared to an open internet with sources in many different languages.<\/p>\n AI companies outside of China just have far more data they can use for training. An AI developer in China will struggle to keep pace.<\/p>\n The constraints caused by limited access to advanced GPUs are also glaringly apparent. U.S. policies that curtail access<\/a> to cutting-edge chips and chipmaking technology will mean that Chinese companies are lagging behind<\/a> their non-Chinese peers.<\/p>\n Yet despite these limitations, China\u2019s AI developers are searching for opportunities to innovate.<\/p>\n A lot of strong talent from the country\u2019s mature consumer tech ecosystem is pivoting to AI. Most of the founding members of the hyped \u201cfour tigers\u201d<\/a>\u2014Baichuan, Zhipu AI, Moonshot AI and MiniMax\u2014had a stint at a big tech company. Their strong intuitions regarding consumers and products are why they\u2019re now leading China\u2019s AI application space. From a consumer\u2019s perspective, their products are on par with many of the leading U.S. applications.<\/p>\n There\u2019s progress on the hardware front too. Huawei\u2019s Ascend AI processors, in particular, seem to be miles ahead<\/a> of their competitors. The Chinese tech giant, now using SMIC\u2019s manufactured chips, claims its Ascend 910B AI chip can outperform Nvidia\u2019s A100 chip in some tests, especially in the use of large AI model training.<\/p>\n Chinese AI developers face some fundamental hurdles, such as a challenging environment, a lack of advanced chips, geopolitical isolation, and national security concerns that limit talent and capital mobility.<\/p>\n Together these constraints will create two parallel AI ecosystems: one inside of China, and one outside of it. The U.S. is going to maintain its lead in developing this transformative technology.<\/p>\n But just because the U.S. has the technological edge doesn\u2019t mean that China\u2019s AI developers will be left behind. Chinese companies have always started off a step behind their non-Chinese peers, yet fierce competition and a willingness to experiment helped them catch up to\u2014and in the case of consumer internet companies, even outcompete\u2014the rest of the world.<\/p>\n In the world of AI, the U.S. and China are both frenemies and competitors. We should hope that the geopolitical competition between them doesn\u2019t get in the way of innovation and collaboration.<\/p>\n The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.<\/em><\/p>\n<\/div>\nRecommended Newsletter:<\/strong> CEO Daily provides key context for the news leaders need to know from across the world of business. Every weekday morning, more than 125,000 readers trust CEO Daily for insights about\u2013and from inside\u2013the C-suite. Subscribe Now<\/a>.<\/div>\n