工业分销商价值创造者 2023（英）

© 2023 Boston Consulting Group1Generative AI is one of today’s hottest business topics, with companies exploring its potentialapplications and benefits across industries and functions, including manufacturing. But despite therecent buzz, manufacturers should recognize that simply applying tools like ChatGPT on their own willnot revolutionize factory operations.Instead of replacing traditional AI, GenAI offers complementary use cases in the areas of assistance,recommendations, and autonomy that pave the way to the factory of the future. It does so through itsGenerative AI’s Role in the Factory of theFutureDECEMBER 08, 2023 By Daniel Küpper, Kristian Kuhlmann, Monika Saunders, John Knapp, Kai-Frederic Seitz, JulianEnglberger, Tilman Buchner, and Martin KleinhansREADING TIME: 15 MIN© 2023 Boston Consulting Group2capacity to generate content, such as text and images, tailored to specific tasks or inquiries. (See “HowGenAI Works.”)To discuss the applications of GenAI, it is essential to first define how it differs from“classical” machine learning (ML). Classical ML algorithms discern patterns withinobserved data, enabling them to generalize these insights to new, previously unseendata. For instance, an ML model might be trained using specific text fragments—suchas operator incident reports in which machine breakdown descriptions are classifiedinto specific root causes such as “end of tooling life” or “operator error.” Based on thistraining, the model can process previously unseen text fragments of incident reportsand judge what caused the incident. The basis for such models may be deep neuralnetworks, support vector machines, or other methods. GenAI takes this approach further. Beyond merely classifying existing text, it cangenerate new text based on specified criteria—such as operator instructions thatoutline a process to resolve a particular root cause of a machine breakdown. Althoughthe progression from classical ML to GenAI might seem incremental, it poses afundamental technical challenge. In classical ML, the model merely needs sufficienttraining to confidently categorize a text fragment. In contrast, GenAI must construct atext fragment from individual words and letters, ensuring that it is grammaticallycorrect, comprehensible, and accurately represents the process. The number of potential outputs from GenAI is virtually limitless. Considering thatthere are roughly 170,000 English in current use, a mere five-word text has more than140 septillion potential combinations. On the other hand, only a fraction of themwould be grammatically correct and understandable. Among those, an even smallerfraction would accurately describe a given process to fix the root cause of a machinebreakdown. Consequently, the margin for error in GenAI models is incredibly narrow, necessitatingextremely precise models. To attain this precision, GenAI must use “foundationalmodels” instead of being trained only on context-specific data. Foundational modelsare trained on extensive datasets