Stop Fine-Tuning to Fix Retrieval Problems — Anant Srivastava

Why watch

An Oracle principal technologist explains where knowledge should live in AI systems—prompt, memory, or weights—and why fine-tuning is often the wrong solution for retrieval problems. Through a case study of a support assistant that invented product names after fine-tuning, he demonstrates how AI architecture accumulates by accident rather than design. For AI engineers and developers building production LLM systems who want to make intentional architecture decisions about knowledge placement.