firehose> #llmops

The organizing decision of AI engineering: given a foundation model you did not train, which adaptation technique fixes the failure you actually have? The ladder runs from cheapest to most expensive — prompting (no weight change), then RAG, then fine-tuning, then combining — and you climb it only when the rung below is exhausted. The rung is chosen by diagnosing the type of failure: if the model fails because it lacks information (private company data, recent events), RAG gives it access to that information; if the model misbehaves — outputs that are factually correct but irrelevant or wrongly formatted — fine-tuning helps more. When both are true, start with RAG (it's easier), begin with a simple term-based retriever, and add fine-tuning after; combining the two often yields the biggest gain. The whole ladder is gated by an evaluation pipeline built first, so you can tell which failure you're looking at.

Claims


Linked from