The last mile of industrial AI implementation is not to deploy AI, but to dare to hand over tasks to AI

Intelligent Things Author | Bi Weihao Editor | Mo Ying In the past few years, AI has swept the entire world at an unprecedented speed. Large models are constantly iterating, agent tools are emerging one after another, and AI programming is rapidly becoming popular. More and more people are beginning to integrate AI into their daily work and life, and more and more industries are gradually accepting and even relying on AI. But there is one industry that has always been cautious about AI - industry. The reason is simple. On your own computer, if the AI ​​makes an error or hallucinations, you can just start over. But on the production line, any wrong instruction from the AI ​​may cause economic losses or even cause a production accident. General-purpose large language models trained based on the vast and complex data in the digital world often fall into a "failure" situation after entering industrial scenarios. They understand knowledge, but do not understand industrial protocols, process mechanisms, and physical laws. This makes AI that seems to be able to answer any question become unstable and unreliable in industrial scenarios. Although the industry has extremely strict requirements for AI calculation accuracy, today's industrial AI is not a castle in the air: In the potato chip production process, AI and digital twins simulated and optimized the dough processing and potato chip filling processes, reducing production line waste by 13% and increasing production capacity by 10%; in the field of new energy vehicle manufacturing, through a combination of AI-based software and hardware products, the entire process from design to manufacturing is digitally transformed, enabling The accuracy of design data is increased by 95%, and the vehicle development cycle is shortened by 25%. In these mature production lines that have achieved a high degree of automation, large industrial models trained through massive industrial data are enough to play a role in a certain link of production and help companies continue to improve production efficiency. However, industrial AI cannot stay at the "helping" step forever. The previous stage of industrial AI was from "unusable" to "partially usable", proving that AI can enter industrial sites. The next stage should be to integrate AI capabilities in countless production links to create an industrial AI product that is "available throughout the entire process, and even easy to use" to prove that enterprises can hand over a complete task to AI. Siemens is the pioneer on this road.