When generation is the wrong tool
A movie-streaming team uses GPT-4o to classify the sentiment of reviews. The feature is popular, but submitting every review with a lengthy prompt is expensive and raises privacy concerns. A small predictive model can handle this narrower analytical task quickly on the company’s own infrastructure. The question is not whether a generative model can do the job, but whether the job calls for generation. section 3.0
Generative modeling uses patterns learned from training data to create new content. It differs from predictive AI, which analyzes data to make predictions or solve clearly defined problems, and agentic AI, which executes multistep workflows and may use either kind of model. section 3.2.1 section 3.4
Choose the approach by the work users need done:
- If the goal is to create new content, consider a generative approach; potential application areas include healthcare and product design. section 3.4
- If the goal is to classify or predict, compare predictive approaches before accepting the cost and data-handling implications of a generative model. section 3.0 section 3.2.1
- If the goal is to carry out a multistep workflow, distinguish the agentic system from the generative model it might use. section 3.4
Model choice also shapes the user experience: decide how much work to automate, how much control to leave with people, and whether a conversational, graphical, or hybrid interface fits the task. section 3.4
