The industry compares the landing of large models on the consumer side to "plain warfare", and the landing in the field of industrial manufacturing to "mountain warfare" and "plateau warfare", the latter is much more complex than the former.
Text: Zhou Xiangyue.
Edited by Zhao Yanqiu.
Recently, Wu Lianfeng, vice president and chief analyst of IDC China, shared an interesting statistic at a conference: after surveying more than 800 samples around the world and 100 samples in China, they found that only 7% of Chinese companies have not done anything in generative AI applications, which is lower than the world's 127% of the total.
This means,A large number of Chinese companies are already experimenting with the deployment of large models and generative AI。Finance, education, medical care, energy, automobiles and other industries can see the figure of enterprise exploration model. Even Mengniu released the industry's first model in the field of nutrition and health in August this yeargpt。
Among the many participants, the industrial field, which covers energy, electric power, chemical, automobile, manufacturing and other sub-industries, is also considered to be an important sector that will bring great changes to the large model.
On the one hand, it is related to the national economy and people's livelihood, and it is the basic plate of economic and industrial development, accounting for 33% of GDP in 20222%。At present, there are more than 400,000 industrial enterprises above designated size in China, covering 41 industrial categories, 207 industrial medium categories, and 666 industrial sub-categories, among which there are a large number of scenarios and business pain points.
In the past few years, the domestic industrial manufacturing field has experienced intelligent manufacturing and AI1After the baptism of stage 0, many enterprises have a considerable degree of awareness of AI applications and have completed the intelligent upgrade of many scenarios. The advent of large models is bringing new opportunities and challenges to this field.
Take a hammer to find a nail
After the large model became popular at the beginning of the year, the first to be excited and take action were various intelligent service providers who have served in the industrial field for many years. Previously, they had been looking at and thinking about how to better use AI in industry, but many functions and imaginations have been hindered by technical limitations, and it is difficult to really produce good results.
It's equivalent to the nail being there all the time, it's just a matter of whether the hammer will work. Zhi Zhen, chairman of China Industrial Internet Technology Group, told the front line of digital intelligence that the emergence of large models has brought a good "hammer", especially with the gradual advancement of multi-modal large models, it is expected to completely open up the application scenarios in the industrial field.
The general consensus in the industry is:The AI model will be integrated into the R&D and design, production process, quality management, operation control, marketing services, organizational collaboration and operation management of industrial enterprises, greatly accelerating the intelligent upgrading process in this field
Market opportunities are also expanding rapidly, with data**, based on China's industrial IT expenditure and the growth level of the global large model, it is expected that the market size of China's industrial large model will exceed 500 million US dollars in 2026, with a five-year compound growth rate of 116%.
It can be seen that for the important market of industry, Huawei, Tencent, Alibaba, Microsoft, Google and other large domestic and foreign manufacturers are accelerating the implementation of large models. Many technology companies have also taken action, China Industrial Internet launched the intelligent industrial model in June this year, and the intelligent manufacturing company SmartMore Technology also launched the industrial multi-modal model IndustryGPT V1 in early November0。
Earlier, the major industrial Internet platforms that have accumulated a lot of industrial data are almost all exploring the combination with large models. Kaos, an industrial Internet platform owned by Haier, has launched the industrial model Cosmo-GPT based on the open-source general model. The Antelope Industrial Internet Platform invested and established by iFLYTEK has also launched a large model of the Antelope industry with the technical base provided by iFLYTEK Xinghuo.
In fact, many people in the industry are optimistic about the combination of large models and industrial Internet. Yan Tongzhu, chairman of the Beijing Informatization and Industrialization Integration Service Alliance, told the Digital Intelligence FrontlineThere is terminal software under the industrial Internet platform, and there is an operating system in the middle, but there is still a brain missing, and the large model can act as this brain. The industrial Internet platform has precipitated a large amount of data from man-machine-material method environmental testing, which can also meet the big data needs of large models.
The data shared by Liu Qingfeng, chairman of iFLYTEK, also confirms this mutual promotion, in just one month and three days after the launch of the industrial model, the total number of users of the Antelope industrial Internet platform has increased by 59%, and the number of platform service enterprises has increased by 88%.
* The end action is frequent, and the attitude and action of the demand side are also changing significantly. "In the first half of the year, everyone was on the sidelines, mainly because we went to find customers. But in the second half of the year, Zhizhen obviously found that the initiative of industrial customers is strengthening, and many enterprises have begun to take the initiative to find them, ** possible scenarios, consulting the latest functions and cases that have been implemented.
There are two common concerns for industrial customers, one is:The work that people could do in the past wants to be replaced by large models；One is the pastWhat can't be achieved by other technologies, I want to let the big model realize it now。Zhizhen observes that almost all enterprises hope to reduce costs and increase efficiency through large models, but some companies take the lead in paying attention to and applying large models in order to maintain technical barriers and advancement.
Intelligent cloud veterans also told the Digital Intelligence Frontline that after the public beta of Wenxin Yiyan, a large number of central state-owned enterprises attach great importance to the exploration of application scenarios of large model technology, including many industrial enterprises. "These big customers will take some scenarios that were difficult to deal with with themselves and come to us, and want to use large models to solve problems," the person said.
EspeciallyEnergy, electricityand other traditional industries, as well asAutomotive, new energyThe high-end manufacturing industry represented by the industry itself has intelligenceHigher rigid demand, andRelatively abundant funds, in this wave of large model craze,The response was the most prompt and positive
For example, in the mining field, which has a very high demand for safe production, according to people familiar with the matter, told the front line of digital intelligenceThe large model of mining has almost become one of the fastest industries to land。, Huawei, Tencent and other large manufacturers have practiced in this field.
But except for some pioneer companies that are running ahead, the industrial manufacturing sector is still a relatively traditional field after allCompanies in the industry are generally more cautious about the application of new technologies, and most of them are still in a wait-and-see mode
Yan Tongzhu observed, especially the amount of data in the traditional manufacturing industry itself is not too large, and the threshold of the large model is relatively high at present, how much value the large model can bring, what is the input-output ratio, "they still can't see clearly", and this will affect its investment progress in the large model.
"Next year will be the year of the app explosion".
There are many subdivided scenarios in the industrial field, not only involving R&D and design, manufacturing, after-sales operation and maintenance, etc., but also fragmented and complex, and the scenarios are also very different between different subdivided industries. With the continuous promotion and exploration of large models in the industry, many companies in the industry have begun to sort out various scenarios that may introduce large model capabilities.
An industry insider admits that some companies' initial expectations for the application of large models are actually somewhat overestimated, butAt present, the application of large models in the manufacturing industry is not as good as imagined, nor is it as bad as imagined.
Zhizhen told the front line of digital intelligence that in the past six months, they have been exposed to a large number of customer feedback needs, and some of the needs put forward by customers will be very detailed, but when they disassemble them step by step to the technical level, they will find that some of these needs are still difficult to get through.
For example, in the front-end R&D and design process, some customers proposed to them that they hoped to use large models to realize the function of converting old drawings into 3D drawings. However, the practice of China Industrial Internet has found that it is relatively easy to automatically convert a ** into a 3D game at present, butThe industry has very high requirements for precisionIt is still difficult to convert two-dimensional CAD into three-dimensional CAD with a large model and solve the problem of repetitive work in the design process, but once it is realized, "it will bring great changes".
Industry insiders observe, due toThe process mechanism is complex, and there are still many pain points to be solved in the R&D and design process from truly seeing the value, but some simple functions can be realized at present。For example, some companies combine their own solution libraries accumulated over the past years with large models, and when customers put forward requirements, they can quickly match the solutions in the solution library.
In fact, in the industrial field, the first to use large models is still**The most universal and relatively peripheral scenarios such as generation, document sorting, and internal knowledge Q&A。Due to the relatively high fault tolerance rate and easy to produce results, this is also the primary position for almost all industries to reap the results in the application of large models.
Bing Jinyou, chief expert of Tencent Cloud's intelligent manufacturing, said that in the knowledge quiz scenario, they helped an automobile factory complete the knowledge of car manuals, allowing large models to replace part of the sales workThey were also approached by a domestic aircraft factoryIt is proposed to train the internal management process and management methods of the enterprise into a large model, and realize the automation and intelligence of reimbursement applications
China Industrial Internet has also achieved good landing results in scenarios such as intelligent Q&A and ** generation. Zhizhen revealed that the expert system they made for the enterprise has been able to achieve an accuracy of more than 90%, and the large model can replace 20%-30% of the programming. At present, the intelligent industrial model has been implemented in nearly 10 projects in the fields of energy and chemical industry, covering multiple scenarios such as intelligent equipment operation and maintenance, industrial network security analysis, and intelligent quality inspection.
In addition to the front-end design and development, the back-end after-sales operation and maintenance and internal management, some manufacturers are also in the middle of the endIn the manufacturing process, we will explore the use of large model generation capabilities to help enterprises detect defects in industrial scenarios。Liu Shu, co-founder of Smartmore, said for example that in industrial scenarios, it is often difficult to collect truly defective data, which will directly affect the detection rate of defective products. Now, through the technology of large model data generation, they can reduce the leakage rate of glue breakage from 1% to 001%。
Kaos, an industrial Internet company, has created an artificial intelligence assembly system based on large modelsSolve the pain point of low efficiency in discrete manufacturing。According to the official disclosure information, the system can reduce the non-processing time of the washing machine factory production line by no less than 20%, improve the efficiency of the process design link by no less than 30%, and increase the efficiency of the production change and commissioning link by no less than 50%.
In addition,A combination of large and small modelsIt is also a direction that the industry is currently seeing more results.
An automotive electronics company in Dongguan, due to the amount of PCB circuit boards used in a car as many as more than 100, design engineers often need to convert a large number of parameters involved in processing and production in the past, and then process, which is easy to make mistakes. But after the large model came, the company and Tencent explored the method of combining the large model with the OCR small model, which can directly identify the drawings, structure them into reusable things, and pass the parameters to the corresponding processing machine.
More scenes are being explored. The industry believes that with the continuous development of large models and the gradual maturity of multimodal technology, the application scenarios in the industrial field will be further opened.
"Next year may be an explosive year for the application of large models in all walks of lifeI am particularly optimistic about the future application support of multimodal large models, once the multimodal is done and lightweight, there will be too many scenarios. Zhizhen said that he is also optimistic about the application of large models on the end side and edge side, "the future volume may far exceed the current imagination."
Data is still a challenge
Since the beginning of this year, the industry has always had an image metaphor for the landing of large models in the industryThe landing of large models on the consumer side is compared to "plain warfare"., and put the landing in the field of industrial manufacturingIt is compared to "mountain warfare" and "plateau warfare".。In a word, it is much more difficult and complicated for large models to land in the industry than to land on the consumer side.
Bing Jinyou told the Digital Intelligence Frontier that they have communicated with a large number of industrial enterprises, especially manufacturing enterprises, and found that the combination of large models and industry still existsCost, talent, data, and the big model technology itselfDifficulties in these aspects.
The cost of implementing a large model includes not only the cost of computing power and deployment, but also the cost of trial and error, labor cost, etc. At present, the cost of landing large models is as high as millions to tens of millions. However, the gross profit of many manufacturing enterprises is relatively low, and they are relatively cautious in investing in new technologies.
From the perspective of the technology itself, Bing Jinyou believes that the illusion problem of the large model determines that it is still a probabilistic output, but in the industry, it is either an economic output, which needs to be 100% correct, or it needs to intervene, that is, whether the large model can be combined with the original traditional digital technology, just like people have both left brain and right brain.
Data is a more difficult pain point to solve.
On the one hand, although there are many scenarios and a large amount of data in the industrial field, the fragmentation phenomenon is obviousThe level of data collection and governance varies from enterprise to enterprise, and even a large number of enterprises may not be sufficient in terms of historical data collection
Yan Tongzhu said, for example, that in the past, there was a large amount of industrial knowledge and industrial data, which only existed in the minds and computers of old engineers and expertsIf it is not transformed into corporate intellectual assets in time, it will gradually be lost in the replacement of talents。"Just like China's oil exploration technology, it is recognized all over the world, but with the departure or death of old experts, some knowledge is broken, resulting in many new employees coming in and having to explore from scratch. ”
On the other hand, China's data element market is still in the early stage of developmentMechanisms such as data co-construction and sharing, and the definition of data property rights have not yet been establishedIn addition, industrial enterprises generally attach great importance to data securityData is a rigid demand for enterprises or campuses, and there are serious cross-industry and cross-scenario data barriers in the industrial field。In fact, the front line of digital intelligence has learned that at present, almost all industrial enterprises will require privatization and deployment when laying out large models.
The difficulties of cost, talent, and large model itself all need to be gradually filled with the passage of time and continuous progress in technology. The problem with the data, howeverIt is necessary to inject more initiative into large-scale model service providers and a large number of industrial enterprisesFrom now on, let's start to solve it.
One of the 10,000 enterprise empowerment plans launched by China Industrial Internet in October this year is to focus on data problems. "All small businesses can use it for free as long as you provide the data, and we train it. We want to try to see if there are any companies willing to take out the data and add it to a platform next year. Zhizhen said.
The report of Tianyi Think Tank, a think tank platform under China Telecom, also shows that cooperation is becoming the main way to solve data problems, and some industries have a good foundation for data co-construction and sharing, and are carrying out practice, such as Northeastern University, Ali and other units have built more than 15 public data sets for surface defects of steel, textiles and other products.
In August this year, the Ministry of Finance issued a document stating that the entry of data assets into the table will be implemented from January 1 next year. "It's a long-term plan. Zhizhen believes that based on this trend, in the future, manufacturers like them will have the opportunity to find some large data distribution groups when training industrial large models and get the high-quality industrial data required for training.
The industry observes that the application of large models in the industrial field will be a gradual development, although there are still many difficulties to break through, but its value will be gradually released in the coming cycle.
In this context, many industry insiders suggest that industrial enterprises should embrace large models as soon as possible. To put it simply, companies don't necessarily have to invest a lot of energy and money to train a model on their own, but they must pay more attention to the progress in this areaCarry out scenario planning, data storage and other work in advance