In November, a year after ChatGPT’s release, a relatively unknown Chinese start-up leaped to the top of a leaderboard that judged the abilities of open-source artificial intelligence systems.
The Chinese firm, 01.AI, was only eight months old but had deep-pocketed backers and a $1 billion valuation and was founded by a well-known investor and technologist, Kai-Fu Lee. In interviews, Mr. Lee presented his A.I. system as an alternative to options like Meta’s generative A.I. model, called LLaMA.
There was just one twist: Some of the technology in 01.AI’s system came from LLaMA. Mr. Lee’s start-up then built on Meta’s technology, training its system with new data to make it more powerful.
The situation is emblematic of a reality that many in China openly admit. Even as the country races to build generative A.I., Chinese companies are relying almost entirely on underlying systems from the United States. China now lags the United States in generative A.I. by at least a year and may be falling further behind, according to more than a dozen tech industry insiders and leading engineers, setting the stage for a new phase in the cutthroat technological competition between the two nations that some have likened to a cold war.
“Chinese companies are under tremendous pressure to keep abreast of U.S. innovations,” said Chris Nicholson, an investor with the venture capital firm Page One Ventures who focuses on A.I. technologies. The release of ChatGPT was “yet another Sputnik moment that China felt it had to respond to.”
Jenny Xiao, a partner at Leonis Capital, an investment firm that focuses on A.I.-powered companies, said the A.I. models that Chinese companies build from scratch “aren’t very good,” leading to many Chinese firms often using “fine-tuned versions of Western models.” She estimated China was two to three years behind the United States in generative A.I. developments.
The jockeying for A.I. primacy has huge implications. Breakthroughs in generative A.I. could tip the global technological balance of power, increasing people’s productivity, aiding industries and leading to future innovations, even as nations struggle with the technology’s risks.
As Chinese firms aim to catch up by turning to open-source A.I. models from the United States, Washington is in a difficult spot. Even as the United States has tried to slow China’s advancements by limiting the sale of microchips and curbing investments, it has not held back the practice of openly releasing software to encourage its adoption.
For China, the newfound reliance on A.I. systems from the United States — primarily Meta’s LLaMA — has fueled deeper questions about the country’s innovation model, which in recent decades surprised many by turning out world-beating firms like Alibaba and ByteDance despite Beijing’s authoritarian controls.
“When Chinese companies are leveraging American open-source technologies to play catch-up, the questions become very complicated — wrapped up in issues of national security and geopolitics,” said Oren Etzioni, a University of Washington professor who specializes in A.I. and the founder of TrueMedia.org, a nonprofit working to identify disinformation online in political campaigns.
In an emailed statement, Mr. Lee, 01.AI’s founder, said his startup’s A.I. model was built on LLaMA just “like most other A.I. companies,” adding that using open-source technologies is a standard practice. He said his company had trained its A.I. model from scratch, using its own data and algorithms. Those were “the main determinants” of the “excellent performance” of 01.AI’s model, Mr. Lee said.
Meta pointed to comments by Nick Clegg, who leads global affairs, in which he said openly sharing the company’s A.I. models helped spread its values and standards, and in turn helped secure American leadership.
(The New York Times has sued the maker of ChatGPT, OpenAI and its partner, Microsoft, for copyright infringement of news content related to A.I. systems.)
A.I. has long been a priority in China. After the A.I. tool AlphaGo defeated two top players of the board game Go in 2016 and 2017, Chinese policymakers set out an ambitious plan to lead the world in technology by 2030. The government pledged billions to researchers and companies focused on A.I.
When OpenAI released ChatGPT in November 2022, many Chinese firms were being hamstrung by a regulatory crackdown from Beijing that discouraged experimentation without government approval. Chinese tech companies were also burdened by censorship rules designed to manage public opinion and mute major opposition to the Chinese Communist Party.
Chinese companies with the resources to build a generative A.I. model faced a dilemma. If they created a chatbot that said the wrong thing, its makers would pay the price. And no one could be sure what might tumble out of a chatbot’s digital mouth.
“It’s just not possible to get rid of all the problematic ways these systems can express themselves,” said Andrew Ng, who teaches computer science at Stanford and was a former executive at Baidu, the Chinese search giant.
Chinese tech giants were also grappling with new regulations that dictate how A.I. models could be trained. The rules limit the data sets that could be used to train A.I. models and the applications that were acceptable, and also set requirements for registering A.I. models with the government.
“It is both more difficult and more risky to innovate in generative A.I. in the current regulatory regime, which is still a moving target,” said Kevin Xu, the U.S.-based founder of Interconnected Capital, a hedge fund that invests in A.I. ventures.
Tech investors in China have also pushed for quick turnarounds from A.I., which has meant money has flowed to easy-to-execute applications instead of more ambitious goals focused on fundamental research, said Yiran Chen, a John Cocke Distinguished Professor of Electrical and Computer Engineering at Duke University. As much as 50 percent of China’s A.I. investment has gone into computer vision technology, which is required for surveillance, instead of building foundation models for generative A.I., he said.
Now Baidu, Alibaba, the dairy company Mengniu and the tutoring firm TAL Education have all jumped into the generative A.I. race in China, leading Chinese media to coin the phrase “the battle of 100 models” to describe the frenzy.
Some have criticized the free-for-all as publicity stunts that add unnecessary competition. In a panel discussion last year, Robin Li, Baidu’s chief executive, described having hundreds of basic A.I. models as a waste.
“More resources should be allocated to applications in various industries, especially considering the limitations on our computing power,” he said.
Success has been elusive. When Baidu introduced its chatbot, Ernie, in March, the “live” demonstration was revealed to be prerecorded. Baidu’s stock plummeted 10 percent that day.
Despite the setback, Baidu remains one of China’s few major efforts at building a foundation A.I. model from scratch. Others are being led by Alibaba and Tencent, China’s tech giants, as well as a start-up linked to Tsinghua University.
A Baidu spokesman declined to comment.
U.S. restrictions on A.I. chip sales to China pose further challenges, since many such chips are needed when training generative A.I. models. Baidu and 01.AI, among others, have said they’ve stockpiled enough chips to sustain their operations in the near future.
There are some bright spots for China with A.I., including in fields like computer vision and autonomous vehicles. Some Chinese entrepreneurs are also looking to leapfrog the United States with breakthroughs in other parts of generative A.I.
Wang Changhu, the former head of ByteDance’s A.I. lab, founded a company called AIsphere in Beijing last year to spearhead what he saw as the next major frontier in the technology: video generation. In November, the start-up released PixVerse, an A.I.-powered generator that can create video from a text description.
“We forged ahead, building our models from the ground up,” Mr. Wang said. “This gives us a significant edge as true pioneers in the realm of video generation.”
That edge may have lasted just a few months. Last week, OpenAI unveiled Sora, an A.I. tool that turns a simple text prompt into videos that look as if they were lifted from a Hollywood movie. Sora instantly went viral.