Part 4/11:
Limitations in Knowledge Representation: Fine-tuned models cannot reliably store new factual knowledge. They are prone to confabulation and hallucination—making up facts or providing unreliable answers.
Task Teaching, Not Knowledge Storage: Fine-tuning effectively encodes patterns—like how to generate emails or write fiction—but not facts.
Semantic Search: The Efficient Alternative
Speed and Cost: Semantic search (also called neural, vector, or neural retrieval) is fast, easy, and cost-effective.
Scalability: Semantic embedding-based search can handle trillions of data points. Adding new data is trivial—no retraining needed.