Introduction

As AI technology makes a critical leap from “dialogue interaction” to “autonomous execution,” many optimistically believe that the intelligent transformation of industrial manufacturing will follow suit. However, the reality is far less optimistic. At the 28th China Beijing Science and Technology Expo’s “Future Industry Promotion Conference,” many industry experts noted that the implementation of industrial AI is not as simple as running a model in a different scenario; it requires a complete redesign of technology from the “brain” to the “body.” The development of industrial AI must overcome challenges such as specialized models, edge infrastructure, and unified standards.

The Need for Specialized Intelligence

“Today’s general models can do everything, but they fail at that crucial 1%, which industrial applications cannot tolerate,” said Li Xiaolong, co-founder and CTO of Yuan Shan Intelligent Technology. Since 2018, Yuan Shan has focused on integrating AI into industrial control and manufacturing, deploying over a thousand models in discrete manufacturing, energy storage control, and process optimization. However, most of these models are not the popular Transformer architectures but lightweight specialized models refined through knowledge density enhancement and redundant parameter removal.

The reason is straightforward: a high-speed automated production line demands extremely strict requirements for latency and accuracy. For instance, the positioning control of welding robots cannot tolerate a deviation of 1‰, and the temperature regulation of chemical reactors cannot rely on probabilistic outputs. General models, with power consumption in the hundreds of watts and inference delays measured in seconds, have little applicability in industrial settings. More critically, industrial scenarios require models that excel at specific tasks rather than models that claim to understand everything.

Overcoming Technical Barriers

Li Xiaolong emphasized that industrial AI must overcome three critical thresholds: trustworthiness, scientific validity, and industrial applicability. While general AI can tolerate a degree of “hallucination,” any uncertainty in the operation of an aircraft made up of six million parts can lead to catastrophic failures. Industrial AI must be “white-boxed,” verifiable, traceable, and compliant with physical principles and engineering standards, which fundamentally contrasts with the “black-box” nature of current large language models.

Li revealed that Yuan Shan does not blindly pursue a large model approach in industrial control but instead adopts a “proprietary model” strategy, enhancing knowledge density and stripping away unnecessary parameters to focus effective knowledge on specific process points through knowledge distillation.

From a broader perspective, this is not a regression in technical routes but a necessary requirement of industrial reality. The development direction of industrial AI is not to pursue a “general artificial intelligence” that solves all problems with one model, but rather to create clusters of specialized intelligent agents tailored to specific processes, devices, and workflows. This judgment is increasingly accepted by frontline practitioners.

Edge Deployment and Operating Systems

If specialized models serve as the “brain” of industrial AI, then ensuring this brain operates efficiently near the production site is the key to current technical breakthroughs.

Li Xiaolong stated that many high-value industrial scenarios, such as energy storage control, HVAC energy savings, and moisture control, must perform real-time inference and closed-loop control at the edge. “Running a Transformer architecture at the edge presents significant challenges in power consumption and accuracy,” he admitted, pushing companies to break through in both model architecture and underlying operating systems.

At the model level, Yuan Shan compresses the generalization ability of large models into high-density proprietary knowledge through knowledge distillation, enabling efficient operation on resource-constrained edge devices. This is not merely about model pruning or quantization but involves redesigning the model’s expression from the perspective of knowledge density.

However, the more challenging issue lies at the operating system level. Industrial sites feature a mix of traditional industrial languages and modern AI frameworks like Python, along with a variety of industrial bus protocols. Currently, no mature operating system exists to unify the scheduling of these heterogeneous resources, achieving shared memory, efficient parameter transfer, and real-time collaboration.

Liu Bing, president of Chengmai Technology, pointed out that while China’s digital development has ranked among the best globally for over a decade, the level of intelligence across industries varies significantly. A key reason is that the capabilities of operating systems, as the foundation of digitalization, have not kept pace. “If we develop new products that integrate AI natively into operating systems, while all edge devices can run the next-generation system, the level of intelligence will increase significantly, even exponentially.”

This judgment reveals a reality: the infrastructure for industrial AI is far from mature. Numerous foundational tasks, from model lightweighting to operating system adaptation, still need to be filled by the industry. This is both a bottleneck for the development of industrial AI in China and an opportunity for local technology companies.

The Challenge of Standards and Protocols

In addition to the technical constraints of operating systems, the lack of standards is a deep-seated barrier to the large-scale development of industrial AI.

Li Xiaolong pointed out an awkward situation in the industry: the calling protocols of large models from mainstream vendors like OpenAI and Anthropic are different and relatively closed. “Today, when we work with lobsters, the protocols generated by OpenAI and Anthropic differ when accessing tokens,” he noted, making collaborative scheduling among multiple agents exceptionally challenging.

He believes that a fundamental experience in the development of the computer industry is that standardization is essential for forming a shared and prosperous ecosystem. “What standards do we currently use in China to define the protocols for agents and large models? There are no particularly good standards yet.” Although discussions about relevant standards began years ago, a complete set of protocol specifications is still lacking at the execution level. The questions of who will lead and guide this effort remain unanswered.

Wei Liang, deputy director of the China Academy of Information and Communications Technology, stated that there has been ongoing work to promote the development of related standards in China. However, due to commercial barriers among major vendors, the pace of advancing interface interoperability is much slower than the industry expects. Without unified standards, the so-called ecological integration is difficult to realize, and the free scheduling and collaborative operation of multiple agents in industrial settings can only remain a conceptual level.

Li Xiaolong advocates that traditional execution software should be clearly delineated into standards such as MCP (Model Context Protocol) or CRUD (Create, Read, Update, Delete) formats for agents to access. By clarifying business boundaries, agents can effectively schedule tasks. In other words, old systems will not be eliminated but need to be “interface-ified” and “standardized.” The software ecosystem in the industrial sector has accumulated numerous specialized tools and systems over decades. Attempting to “start over” with agents is neither realistic nor economical. A more feasible path is to incorporate these existing assets into the scheduling range of agents through standard protocols, allowing new and old systems to work together.

China possesses the world’s most complete industrial system and the richest application scenarios, providing a unique testing ground for the iteration of industrial AI. Many experts at the conference indicated that to move from “experimentation” to “scale,” standardization is a necessary hurdle. As Li Xiaolong stated, only when standards are clarified can we achieve symbiosis and mutual prosperity in the industry.