Are Large Models in Industrial Manufacturing a Real Demand or Just Following the Trend?
With the rapid advancement of AI technology, the wave of “large models” has inevitably reached the industrial manufacturing sector. These large models, which are complex AI systems trained on massive datasets, have demonstrated powerful capabilities across various domains, such as natural language processing, image recognition, and autonomous driving.
As the country with the most comprehensive range of industries, China is currently transitioning from digitalization to intelligent manufacturing. In this wave of digital transformation, large models hold great promise, with businesses widely hoping to leverage intelligent technologies to enhance production efficiency, optimize processes, reduce costs, and drive digital transformation forward.
However, while the vision is ambitious, the reality is somewhat less so. Unlike the heated competition in language models like ChatGPT, Baidu’s ERNIE Bot, or Alibaba’s Tongyi Qianwen, which are driving fierce price wars, the development of industrial large models has remained relatively quiet. The reason lies in the fact that industrial large models are still in their early stages and require further development and refinement.
On the afternoon of August 11, 2024, the Shenzhen Branch of the China Computer Federation’s Young Computer Scientists & Engineers Forum (CCF YOCSEF Shenzhen) hosted a viewpoint forum at the International Conference Center in Shenzhen University Town. The theme of the forum was “Are Large Models in Industrial Manufacturing a Real Demand or Just Following the Trend?”
Leanplans Participates as a Guest Moderator at the 2024 China Computer Federation Viewpoint Forum. The forum gathered over 80 attendees, including representatives from institutions such as Tsinghua Shenzhen International Graduate School, Peng Cheng Laboratory, and 48 companies including CGN, Leanplans Manufacturing, TCL CSOT, and Amazon.
Cognitive Understanding and Outlook
AI large models have already made significant strides across various industries. In 2022, DeepMind’s AlphaFold2 model successfully predicted nearly all known protein structures. More recently, the China Academy of Information and Communications Technology (CAICT) released the 2023 Large Model Application Casebook, selecting 52 exemplary cases of commercial large model deployments, covering sectors such as intelligent manufacturing, education, fintech, media, healthcare, and transportation.
During the introduction session, Professor Li Wenxian, Vice Dean of the Sino-German College of Intelligent Manufacturing at Shenzhen Technology University, used intelligent robotics as an example to explore the prospects and challenges of large models in driving industrial automation.
Professor Li highlighted that large models are currently being applied in manufacturing, logistics, and energy sectors. These applications include autonomous navigation and path planning, machine vision for quality inspection, and flexible assembly with collaborative robots.
When discussing the value of large model applications, Professor Li Wenxian identified three key areas of impact. First is efficiency improvement, where large models optimize production processes and increase equipment uptime. Second is enhanced product quality, achieved by detecting defects through automated systems to ensure consistency. Lastly, large models enable production flexibility, empowering businesses to quickly respond to market changes and meet demands for personalized and customized manufacturing.
Professor Li also pointed out several challenges currently faced by intelligent robots powered by large models, including data scarcity, high variability, quantifying uncertainty, safety assessments, and real-time performance issues.
He believes that future research should focus on optimizing computational efficiency, improving model generalization, and strengthening data security measures. Additionally, new ethical guidelines and regulations will need to be established to address the challenges posed by these emerging technologies.
While large models show significant promise, Liang Xiaojun, Director of the Industrial Intelligence Research Office at Peng Cheng Laboratory, issued a reminder that current industrial internet development remains focused on the network and platform layers, with insufficient progress in upgrading the underlying industrial control systems. He urged the industry to strengthen independent innovation and build a quality control system for the new industrial internet, ensuring sustainable growth in the sector.
The Symbiosis of Large and Small Models
Currently, due to the diversity of industrial scenarios, limitations in computing resources, the challenges of collecting and organizing training data in industrial fields, as well as concerns around the safety and reliability of large models, the application of large models in China’s industrial sector remains in an early exploratory phase.
Most existing applications of industrial large models are focused on knowledge-based Q&A systems. However, this basic functionality is far from sufficient. The true value of large models lies in their ability to penetrate deeper into the core execution processes of industry, which is currently the most pressing need in the manufacturing sector.
A fully realized industrial large model should cover all critical aspects of manufacturing, including solution planning, R&D design, experimental validation, production and manufacturing, operational management, and even marketing and after-sales services.
At present, large models struggle to comprehensively capture the characteristics and rules of specific industries or sectors, and they cannot fully meet the specific application needs of certain fields. To genuinely integrate large models into industrial applications, three core challenges must be addressed: first, the lack of deep understanding of the industry; second, insufficient familiarity with individual enterprises; and third, the issue of hallucination (i.e., generating incorrect, meaningless, or inaccurate outputs).
Overcoming these obstacles is essential for the deep application of large models in real industrial scenarios. Advancing the use of large models requires a gradual, step-by-step approach, with strategies that involve categorization, layering, and phased implementation. This process cannot be rushed.
During the debate session, Chengmin Tan, founder of Leanplans, served as the moderator, leading an in-depth discussion on the topic, “Which types of industrial enterprises need large model support?”
He pointed out that China currently has only 483,000 large-scale industrial enterprises, leaving approximately 4 million small and medium-sized factories facing significant challenges in digital transformation.
For these businesses, small models, due to their lightweight, efficient, and easy-to-deploy nature, often provide a quicker and more flexible solution to meet their practical needs. He further emphasized that while large models are well-suited for handling complex tasks, in many specific industrial scenarios, small models may actually be more appropriate.
For example, for small and micro enterprises with low production volumes and diverse product types, the flexibility and cost-effectiveness of small models are particularly attractive. According to a statistical analysis by the China Academy of Information and Communications Technology (CAICT) on 507 AI small model application cases, these models are predominantly applied in the manufacturing sector, accounting for as much as 57% of the use cases.
Looking ahead, it is expected that large models and small models will coexist in the industrial manufacturing domain for the long term, with their integration creating even greater value for enterprises.
On this topic, the panelists at the event engaged in in-depth discussions. They unanimously agreed that companies in the Greater Bay Area have a strong demand for advanced technologies, but when implementing them, they are more concerned with whether large models can reliably and effectively solve their pressing problems and the associated cost of investment.
Due to the varying challenges across different industries and sectors, large models need to be adapted to local conditions to overcome the “last mile” barrier from technology to practical application. This process can be accelerated through policy support and close collaboration between the tech sector and industry.
Currently, small models, with their efficiency, flexibility, and low resource consumption, continue to play an irreplaceable role in specific environments and tasks. Looking ahead, reducing the cost and technical barriers of deploying large models will be key to enabling these models to assist industrial enterprises in their upgrade and transformation efforts.
Data Quality and Sharing
During the discussion on the topic “How can industrial enterprises seize the opportunities brought by technological advancements in the era of large models?” Ms. Ding Liu’an from Huaxing Optoelectronics shared her insights.
She emphasized that companies should assess the necessity of digital transformation based on their development stage, with the key consideration being whether the technology can create tangible value for the business. Companies should not focus solely on the sophistication of the technology but should place greater importance on the actual benefits it brings.
Whether it is large models, small models, or other technologies, the core goal is to address the problems faced by enterprises and enhance production efficiency. Huaxing Optoelectronics is planning to establish its own computing power platform and aims to create models specifically tailored for the panel industry through data sharing within the industry.
Existing large models are primarily trained on non-industrial corpora, while the lack of sufficient and accessible data in the industrial sector limits their application in this field.
Data quality directly impacts the accuracy and reliability of models, yet industrial data is often fragmented and diverse in structure. Companies, due to security and commercial considerations, are unable to share data, resulting in a lack of high-quality data support for industrial large models during training.
The successful application of industrial large models relies on the precise extraction of high-quality data unique to the industry and effective data processing techniques. The key lies in extracting industrial knowledge, particularly in transforming engineers’ experiences into usable datasets, and enhancing data quality through data cleaning and standardization to improve the models’ generalization capabilities and predictive performance.
Leanplans, as a component manufacturing factory and online order platform with years of production experience, not only excels in production capacity but also leads the industry in the accumulation and application of data.
By continuously gathering and optimizing business data, LeanPlans has built a highly competitive proprietary database. This data encompasses various aspects, from design drawings and price assessments to production processes, forming a comprehensive and in-depth manufacturing database.
LeanPlans’ proprietary models can accurately analyze customer needs, intelligently match production resources, and provide optimal solutions in a short timeframe. This not only enhances production efficiency but also significantly reduces costs, offering customers a higher cost-performance service.
Through the continuous refinement of these proprietary models, LeanPlans has gradually established a unique competitive moat in the manufacturing industry, securing a favorable position in the market.
Conclusion
After more than four hours of intense discussions, the forum came to a close. Huang Dian, the chair of CCF YOCSEF Shenzhen and associate researcher at the Guangdong Academy of Intelligent Technology, provided a summary of the event.
She emphasized that the purpose of this activity is to strengthen the connection between academia and industrial enterprises. By delving into discussions about large model technologies, the forum aims to help businesses better understand their potential while also bringing academic insights closer to industry needs, thereby promoting the practical application of technology in industry.
Huang further stated that CCF YOCSEF Shenzhen will distill the insights from the forum into several outcomes, providing constructive suggestions for government decision-making, industry development, and the intelligent transformation of enterprises, thus aiding industrial manufacturing and upgrading.
The development of industrial large models relies not only on technological breakthroughs but also on addressing key issues such as data acquisition, technology adaptation, and cost control. Only when these challenges are effectively addressed can industrial large models truly unleash their potential and drive the process of industrial intelligence.
This requires collaborative efforts from the government, enterprises, technology providers, and various sectors of society. With continuous technological advancements, industrial large models will bring more innovation and transformation to the industry, enabling enterprises to respond more efficiently and accurately to market demands, achieve sustainable development and competitive advantages, and promote the deep integration and transformation of artificial intelligence with traditional industries.
Introduction of Leanplans
Leanplans focuses on the integration of AI and manufacturing applications, developing proprietary AI models to empower the manufacturing of non-standard parts, thereby reducing costs and increasing efficiency in design, DFM, pricing, transaction, and production collaboration.
With a self-operated factory at its core, Lean Plans integrates high-quality production capabilities such as CNC machining, sheet metal fabrication, 3D printing, die casting, and injection molding. Through an “AI-powered full-scenario + cloud factory” approach, it provides a one-stop solution for the design, development, and flexible production of non-standard parts for domestic and international clients.
This approach meets the demands for small batch sizes, diverse product varieties, high frequency, short delivery times, and flexible manufacturing.
Get a Free Quote!
Leave your contact details, or directly visit our online quoting platform to experience the future of material selection and production. Get expert material evaluations, tailored DFM analysis, and fast 24-hour production turnaround.
- Free Quote: Upload your designs, and our AI-powered engine will generate a custom quote in seconds.
- Talk to an Expert: Connect with one of our engineers via WhatsApp for immediate assistance.