Photo: VCG

The recent controversy surrounding Stanford University students plagiarizing a Chinese AI large model has sparked heated discussion online after the model drew global attention for its powerful performance. Liu Zhiyuan, the co-founder of ModelBest, whose model was allegedly copied, told the Global Times that the case should not be over-interpreted, but it does highlight how China is gradually becoming an important player globally in AI model research and development.

ModelBest is a well-known domestic startup that has raised hundreds of millions in financing and has over 100 research and development personnel, more than half of whom graduated from top Chinese universities such as Tsinghua University and Peking University.

Many people see the incident as a case of reverse plagiarism by a traditionally dominant player. Others argue that this indicates a narrowing gap between China and the US in large models, at least on a technological level.

In the 2024 Artificial Intelligence Index Report released by Stanford, Tsinghua University was noted as one of the academic institutions outside the West that has released the largest number of foundation models.

Where does China's AI large model ecosystem currently stand relative to global developments? How did China's cutting-edge AI large model research team transition from beneficiaries of open source to contributors? What is the future direction of China's AI large model development? The Global Times interviewed Liu Zhiyuan, an associate professor at the Department of Computer Science and Technology, Tsinghua University, to explore how China is rapidly emerging as an important player in global AI development.

Being a global leading player

Silicon Valley Venture Capitalist Vinod Khosla expressed his concerns in May: "We are in an AI war with China, and whoever develops the best models will dominate the world economically, socially, and politically."

The competition between the US and China in the field of AI has become one of the most talked-about topics, where even minor developments can spark debates.

The involved team from Stanford University has issued an apology following accusation of using the open-source work of Chinese scientists without proper attribution in the development of a new AI model. The accusation came after the team announced Llama3-V on May 29, claiming it had achieved comparable performance to GPT4-V and other models with the capability to train for less than $500.

The Llama3-V quickly ranked in the top four on the trending list on Hugging Face, the most popular AI platform, soon after its release, and yielded over 1,000 result pages on Google Search.

But netizens uncovered evidence that the Llama3-V project code was reformatted and from MiniCPM-Llama3-V 2.5, a multimodal large language model developed by ModelBest and Tsinghua University. The two models were verified to have highly similarity in terms of providing answers and even the same errors, and that some relevant data had not yet been released to the public, according the CEO of ModelBest Li Dahai. He noted that the MiniCPM-Llama3-V2.5 is currently the world's most powerful end-side multimodal large language model.

Li said the team hopes that their work will receive more attention and recognition, but not in this way. He also called for an open, cooperative and trusting community environment.

In a recent exclusive response, Liu told the Global Times that the plagiarism incident, as an isolated case, is not worthy of extended discussion, but it does serve as a window for everyone to see the creativity and influence of China's open-source models.

Liu said that in the wake of the incident, they received a lot of support and solidarity, including from scholars and practitioners in the US. "Everyone upholds the most basic sense of fairness and justice, based on the consensus of breaking through national and political boundaries to jointly enhance the development of artificial intelligence. There is a lot of room for mutual learning between China and the US in the development of open-source large models."

Lucas Beyer, a researcher at Google DeepMind, commented on X that "MiniCPM got a lot less attention, including from me. Given the similar results, the main reason seems to be not coming from an Ivy-league Uni, but from a Chinese lab?"

In fact, prior to the plagiarism incident, Liu's team had been working on natural language processing and the development of large models for many years, and had already established a certain level of influence in global open source field.

"When we initially released MiniCPM-Llama3-V 2.5, it rushed the top of the trending list on Hugging Face soon and was on the list for about a week, and the series had around 370,000 downloads. This largely reflects the recognition we have received from the AI academia and industry, especially in the innovative end-side models, which has allowed more people to see the important value of China's contributions to open-source technology."

Liu said that on Hugging Face's Trending list, it is common to see that about half of the top 10 open source models are developed by Chinese research teams. "It should be said that China is rapidly becoming a more important player in the global development of artificial intelligence."

Liu also emphasized that China is still in a state of "catching up," but the "gap between us and the most advanced levels internationally is getting smaller."

"Some scholars and practitioners believe that the strength of large models from China is greatly underestimated and overlooked, which I think is understandable. In terms of academic research, Chinese researchers were already able to publish many high-quality academic papers internationally as early as 10 years ago, but it is still clear today that well-known teams from American universities receive higher levels of recognition and attention," Liu said.

In Liu's view, China has become a very important and dynamic participant in AI research on a global scale, but he also emphasizes that China is still in the process of catching up the top. "There is a metaphor that says China's current development in AI is on a plateau, meaning that it has already formed a very strong research and development basement, becoming increasingly powerful in terms of data, technology, and application scenarios. However, it has not yet reached its peak, because we can see that several key technologies that have driven the rapid development of AI in the past decade are still introduced by the US."

"Objectively speaking, we are still in the catching-up stage, but the gap is getting smaller," Lee Kai-fu, an AI expert who is also the CEO of 01.AI, has stated to the media before that the difference between China and the US in AI lies in breakthrough innovation, such as the Transformer technology invented by Google, and many new large models and new practices developed by OpenAI.

However, Li pointed out that China also has unique advantages. Li believes that China's biggest advantage is its ability to develop applications in practice, such as WeChat, TikTok, Feishu, and other products that basically outperform American applications.

Li explained that this is because in the era of mobile internet in China, the biggest skill that entrepreneurs and product managers have learned is how to iterate quickly, launch, optimize, continuously improve products, a skill that he believes the US still lags behind China.

Following the plagiarism incident, Liu spoke out on the Chinese Quora-like platform Zhihu stating that the rapid development of AI relies on the global sharing of algorithms, data, and models. He explained that their team developed the MiniCPM-Llama3-V 2.5, which uses the latest Llama3 as a language model base.

The cornerstone of open source sharing is compliance with open-source licenses, trust in other contributors, and respect and admiration for the achievements of predecessors, Liu said.

Recently, ModelBest announced that its MiniCPM large model will be free for commercial use, once again demonstrating how it has evolved from a beneficiary of the open-source spirit to a contributor.

Liu told the Global Times that the concept of a community of shared future for mankind advocated by China is a very good philosophy in the open-source community.

Future trends of China AI models

At the national two sessions held in March 2024, "new quality productive forces" became a popular topic. The government work report also pointed out the need to vigorously promote the construction of a modern industrial system, accelerate the development of new quality productivity, and emphasized the deepening of research and application of big data, AI, and the launch of the "AI+" action to create a digitally competitive industrial cluster. AGI is becoming an important engine for new quality productive forces.

There has been a heated discussion on how to truly apply AI in practical scenarios to help improve productivity.

Data shows that there are at least 130 companies in China researching large model products, with 78 of them focusing on general large models. Liu believes that China's advantages in large model research and development are also obvious, especially in the development of stronger end-side models, which will be an important trend in the future.

He calls for China to further strengthen its ability in the most original and grounded innovation, shifting its focus from being a follower to being an innovator.

Liu believes that the future development of AI mainly involves three major directions. The first is how to produce high-quality large models using more scientific methods, specifically how to achieve more efficient and smaller models on end-side devices, similar to turning a supercomputer that fills several rooms into a small chip in a phone. The second is the efficient integration of scientific research and engineering manufacturing, requiring more advanced engineering to manufacture large models. The third is how to apply AGI to everyday life to make our lives more efficient.

Liu Yuanchun, President of Shanghai University of Finance and Economics, stated that as the world's second largest economy, China has a super-large market and digital resource advantages, providing a deeper and richer landing scenario and learning environment for large models. At the same time, this also requires large models to be more practical and support the development of the real economy.