In recent years, Artificial Intelligence (AI) has been a hot topic at the annual sessions of the National People's Congress and the Chinese People's Political Consultative Conference, and this year is no exception. "Deepening research and application of big data, artificial intelligence, and launching the 'AI+' action" has been included in this year's government work report.
Large-scale models are currently the focus of attention and represent the future direction of artificial intelligence. Wang Jinqiao, Executive Deputy Director of the Big Model Research Center at the Institute of Automation, Chinese Academy of Sciences, told the China Science Daily that among the three paths to General Artificial Intelligence (AGI) – brain-like intelligence, information intelligence, and game intelligence – information intelligence represented by large models is "advancing the fastest".
From large language models like ChatGPT to AI video generation models like Sora, the United States has been leading the way in the field of artificial intelligence. Despite various domestic efforts in recent years to catch up, the emergence of Sora raises concerns: Is the gap between Chinese artificial intelligence and the world's advanced level widening? Will we lag behind in the era of AGI?
Let's hear what representatives and experts in the field of artificial intelligence have to say.
China is narrowing the gap with the forefront
In an interview with the China Science Daily, Zhao Xiaoguang, a member of the National Committee of the Chinese People's Political Consultative Conference and a researcher at the Institute of Automation, Chinese Academy of Sciences, stated, "Since the reform and opening up, it has taken us just over 40 years to complete the path of industrial revolution and technological revolution that the West has gone through. Whether China's research and development and industrial development in artificial intelligence are diverging from the advanced level, I believe this question should be answered in this context."
She mentioned that in some cutting-edge technological fields, it seems like we are always catching up with Western countries, but this does not necessarily mean that China is falling behind in advanced research. "In some areas, we are even leading."
Zhao Xiaoguang has long been committed to the research of advanced robots and intelligent robots, and the research on "humanoid robots" has become popular. "In 2022, American entrepreneur Elon Musk proposed to launch 'Optimus Prime,' and we gained a new understanding of humanoid robots. Now, our laboratories and domestic companies have produced humanoid robots with excellent functionality. What does this mean?" Zhao Xiaoguang asked and answered herself, "It means that we have mastered many key technologies, but the problem lies in not fully understanding and integrating them. Large AI models are the same."
"So my answer is that the gap between China and the forefront of artificial intelligence is not widening but narrowing," she said. "We need to continually refine our innovative ideas, make good use of the advantages of the new national system, and gradually lead the way."
In contrast to Zhao Xiaoguang's views, Liu Wei, Director of the Human-Computer Interaction and Cognitive Engineering Laboratory at the School of Artificial Intelligence, Beijing University of Posts and Telecommunications, was more direct.
"People who say that the gap between China's artificial intelligence field and European and American countries is widening probably don't understand artificial intelligence," Liu Wei said, emphasizing that artificial intelligence has many research directions, and generative AI large models are just one of them, in which China has leading aspects.
Both Zhao Xiaoguang and Liu Wei mentioned that even with powerful models like GPT-4 and Sora, the goal of achieving AGI is still very distant, "and may even be impossible." Therefore, there is no need to worry about "China missing the bus for AGI".
Large models need to "walk on two legs"
Undeniably, as the most attention-grabbing technology in the field of artificial intelligence, every iteration of large model technology signifies changes in the competitive situation among countries. According to Liu Qingfeng, a member of the National People's Congress and chairman of iFlytek, the competition between China and the United States in the deep application and strategic demand of large models will be "critical" by 2024.
To this end, Liu Qingfeng suggested formulating a national plan for the development of general artificial intelligence to systematically promote the development of China's general artificial intelligence. The first goal is to leverage the advantages of the new national system, increase and maintain continuous investment in the "main battlefield" of general large model bases.
"We need to face the gap squarely, focus on the independently controllable base large models 'main battlefield,' concentrate resources at the national level, accelerate catching up, and systematically construct the ecosystem and applications of general artificial intelligence to build comprehensive advantages," Liu Qingfeng said.
Wang Jinqiao told the China Science Daily that the biggest limitation of the development of domestic large models is computing power.
To address this issue, Zhang Yunquan, a member of the National Committee of the Chinese People's Political Consultative Conference and a researcher at the Institute of Computing Technology, Chinese Academy of Sciences, conducted a special investigation. After assessment, he proposed, "In addition to continuing to tackle artificial intelligence chips, can we aggregate China's supercomputing power to provide computing power support for large model pre-training?"
Zhang Yunquan told the China Science Daily that the characteristics of computing power required for training large models, such as parallel computing, high-speed network interconnection, and communication technologies, are all possessed by supercomputing systems. It is feasible to organize talent to conduct technical research and develop a dedicated supercomputing system for large model training.
"As NVIDIA's founder and CEO Jensen Huang said, every country should have its own sovereign artificial intelligence infrastructure. Who will train it? I think it can be achieved by establishing a special project to develop supercomputers that support the training of large models," Zhang Yunquan said.
He told reporters that developing such supercomputing facilities would entail higher costs, but it would solve the problem of whether or not there is sovereign large models. At the same time, we should also "walk on two legs" – keep up with the research and development of domestic artificial intelligence chips. After breakthroughs are made, costs will naturally decrease.
Wang Jinqiao agrees that large model computing power needs to "walk on two legs," and he believes that the development of large models also needs to "walk on two legs." "In addition to universal basic large models, China may need multiple dedicated large models that meet the needs of different scenarios," Wang Jinqiao said. China has the largest market demand, such as chemical discovery, molecular simulation, weather forecasting, etc., which do not require particularly large model scales but need to be fully integrated with industry scenarios and data. Such dedicated models have their uses.
He further explained that the development of universal basic large models is to maintain resonance with the progress of cutting-edge technology and not miss the "window period" of technological development. Specialized large models are oriented toward market demand and utilize the capabilities of large model technology through scenario applications.
"2024 may be the year when large models land," Wang Jinqiao said. "We need to explore a path of characteristic development of large models." Central State-Owned Enterprises (SOEs) Entering the Arena: Opportunities and Challenges
Recently, the State-owned Assets Supervision and Administration Commission of the State Council (referred to as the State-owned Assets Supervision and Administration Commission, or SASAC) held a special promotion meeting on "AI Empowering Industrial Renewal" for central SOEs, emphasizing the need to accelerate the layout and development of the artificial intelligence industry.
According to Zhang Yunquan, the entry of central SOEs, which possess capital, talent, and scenarios, into the field signifies a positive development for industries like artificial intelligence, which are capital and talent-intensive. However, it's crucial to distinguish priorities and avoid rushing in blindly.
Explaining to China Science Daily, Zhang pointed out the difference in thinking patterns between the East and the West. While the East tends to strategize like playing chess, focusing on seizing territories and capturing leaders, the Western approach is more akin to playing Go, preferring to surround and suffocate opponents. This analogy extends to the competition in large model development, reminding us to have strategic composure and not to hastily pursue commercialization, as it may lead to missing out on original innovation opportunities.
He suggested that China could have hundreds of institutions exploring the application of large models in various industries, gradually investing more resources after demonstrating practical results through combining them with industry data. Simultaneously, it's essential to establish an elite "sharpshooter team" to drive the research and iteration of core large models, ensuring sovereignty and preventing future technological bottlenecks.
Liu Wei cautioned that while central SOEs have the advantage of concentrated resources for significant undertakings, they must avoid superficial collaborative efforts between academia, industry, and research. He emphasized the replicable success pattern of OpenAI, which involves long-term stable support, a non-profit-oriented approach, and unwavering focus on goals. "I hope central SOEs can learn from OpenAI, cultivating a 'quiet yet far-reaching' atmosphere and an open system."
Zhao Xiaoguang highlighted the diversity in collaboration within the Silicon Valley humanoid robotics startup scene, where, apart from Musk's "Colossus," there are many other startups like Figure AI focusing on humanoid robot limbs. "Categorized collaboration can better facilitate the implementation and transformation of cutting-edge AI technology," Zhao noted. "Leading enterprises participating in advancing AI development should leverage their 'chain-leading' advantages, combining research achievements of scientific research institutes with industry strengths to form synergies and bridge the gap between technology and industry."
As an insider in the large model community, Wang Jinqiao welcomed the entry of central SOEs. Specifically, he hoped they would intensify efforts in developing smart computing centers, opening up more scenarios, and sharing industry data to fully embrace large models.
"I hope central SOEs can create more space for the development of large model technology by absorbing the latest AI technology, coordinating large and small models, and taking the lead in experimentation," Wang stated.
Source: China Science Daily (March 8, 2024, 4th Edition, Two Sessions)
