Can I install an LLM in my app and train users on how to prompt it?
John Michelsen explains the potential risks and challenges of incorporating Language Learning Models (LLMs) into enterprise applications. He discusses some of the significant issues providing direct access can introduce including misunderstanding user intent due to the complexities of prompt engineering and the potential for LLMs to generate imaginative, "fairy tale" responses rather than accurate, journalistic ones.
John emphasizes that LLMs are not knowledge management systems, their purpose is to train in linguistic correctness, not factual accuracy. He stresses the need to prompt these models with accurate information in real time to prevent the spread of misinformation. Even with extensive fine-tuning, these models may still fail to adequately answer around 80% of user queries, particularly ones that require real-time information or workflow triggering.