6月29日讯（Reporter Zhang Yangyang）At the "Empowering New Capabilities, Driving the Future with Intelligence" Young Scientists' Technology Transfer and High-Quality Development Seminar on Embodied Intelligence, Professor Lan Xuguang from Xi'an Jiaotong University was interviewed by media outlets including the Sci-Tech Daily on hot topics such as China's position in global technological competition and challenges in the industrialization of artificial intelligence.

Lan Xuguang pointed out that China is now very close to the United States in cutting-edge fields like embodied intelligence. However, when it comes to artificial intelligence technology, especially the application of large models in the physical world and industrial scenarios, it faces a core bottleneck known as "physical constraints."

Lan Xuguang stated that China's technological development has been rapid, and it has taken an important global position in areas such as artificial intelligence, green energy, and batteries. The advantage in developmental trends "may still be before the fluctuations." Focusing on his research field of embodied intelligence, he believes that the gap between China and the United States in the era of large models is now very close.

"Taking DeepSeek as an example, although its overall performance still has a slight gap compared to the leading ChatGPT, China has a significant advantage in industry application data," emphasized Lan Xuguang. He stressed that China, as a major manufacturing country, possesses a massive data scale and rich industry application scenarios, which form a solid foundation for the continuous development of artificial intelligence and its important position. This indicates that in the next step of artificial intelligence industry expansion, "China may go further."

Lan Xuguang believes that large models have developed rapidly in the digital world (such as writing, creation, programming), with a replacement rate of 50%-60%, significantly impacting related high-end professionals. However, there is a huge gap in their application in the physical world and industrial scenarios.

"The core of the Transformer architecture is to predict the next Token, following the statistical laws of data rather than the real constraints of the physical world."

Lan Xuguang explained that industrial scenarios often require strong causal relationships, must comply with real physical and chemical laws, and require extremely high stability and reliability (usually requiring over 99%). However, current large models do not pursue physical constraints, and their outputs cannot guarantee feasibility and safety in the physical world, which leads to difficulties in their implementation in industrial scenarios. "This problem is faced globally."

He further pointed out that the application of large models in the programming field works well, thanks to sufficient data and the ability to self-validate in a virtual environment. However, in the physical world (such as controlling engine valves), the lack of a reliable verification mechanism is the biggest challenge for AI implementation.

However, Lan Xuguang also mentioned that there are still some challenges in the industrialization of AI: First, technical contradictions. Academic research focuses on exploration and frontier, while industrial production requires stability and "no mistakes," creating an inherent conflict. There is a need for engineers to bridge the gap between frontier and practicality.

Second, financial issues. China's investment environment emphasizes application orientation, and while support for pure frontier technologies is increasing, it may still pose certain obstacles to the practical application of frontier technologies.

Additionally, the relationship between technology and the market is very complex, as there are certain differences between market demands and enterprise life needs and the technology of scientific researchers.

In terms of attracting investment, Lan Xuguang is also optimistic. He said that through the meeting, showcasing the technology can allow entrepreneurs to more intuitively see the potential of the technology to solve industry pain points, thereby "likely" attracting more investment attention for frontier technologies.

Lan Xuguang called on academia and industry to focus on more fundamental technical issues, rather than merely imitating the Transformer architecture. "Transformer essentially cannot create new things or new concepts; it is an optimization and combination of existing knowledge."

He believes that to solve the constraints of the physical world, it may be necessary to explore more important new models, such as looped AI, to meet the strict requirements of the industrial sector for high reliability, explainability, and compliance with physical laws.

(By Zhang Yangyang, Caijing News)

Original article: https://www.toutiao.com/article/7521263095658840603/

Statement: This article represents the personal views of the author. Please express your opinion below using the [Up/Down] buttons.