Weigh the tradeoffs

Training is expensive but amortized across many uses and can change persistent behavior. Inference is repeated per request and dominates operating cost at scale. Fine-tuning can improve stable task behavior but is slower to update than retrieval; retrieval gives fresh facts without changing weights.

Tradeoffs become manageable when expressed on a shared scorecard. Record task quality, p95 latency, unit cost, operational burden, and risk exposure. Do not collapse them immediately into one number; a weighted score can hide an unacceptable safety or privacy threshold. First mark non-negotiable constraints, then optimize among the surviving options.

Consider the mechanism when judging a trade. During training, examples pass forward through the network, a loss compares predictions with targets, gradients flow backward, and an optimizer updates weights. During inference, only the forward computation is needed, often with autoregressive token generation and a KV cache. Fine-tuning is additional training; RAG and prompting are inference-time techniques. That explains why a control can improve one stage while degrading the whole pipeline. Test at the system boundary seen by the user, not only inside the component. A locally faster retriever, sampler, or model does not help if queueing, retries, validation, or human review dominates end-to-end time.

A classifier trains for five epochs on 80,000 labeled tickets and freezes a checkpoint. Every new ticket then runs a forward pass in milliseconds. Adding today’s policy to a prompt helps that request but does not alter the checkpoint.

The worked number is one epoch over 80,000 examples with batch size 100 requires 800 optimizer steps; five epochs require 4,000 steps. State the unit and denominator whenever you report it. A percentage without a denominator can conceal a tiny sample; a latency without a percentile can conceal slow users; a similarity score without a labeled task can conceal irrelevant neighbors. Compare the observed value with a threshold chosen before seeing the final test result.

Now test the tempting shortcut. Suppose the team optimizes only the most visible metric. The result may look better while the system becomes less trustworthy. The reason is concrete: Training is expensive but amortized across many uses and can change persistent behavior. Inference is repeated per request and dominates operating cost at scale. Fine-tuning can improve stable task behavior but is slower to update than retrieval; retrieval gives fresh facts without changing weights. This is why the decision record must include both the intended gain and the tolerated regression. If the tolerated regression is unknown, the change is not ready for a consequential workflow.

• Prefer a measured baseline over a persuasive demo.
• Keep versions, inputs, and thresholds reproducible.
• Separate syntactic success from semantic correctness and authorization.
• Escalate or abstain when evidence falls outside the contract.
• Re-evaluate when data, traffic, models, providers, or user goals change.

These rules turn the topic into an engineering decision rather than a slogan. They also make disagreement productive: another person can challenge the assumptions, rerun the evaluation, and reach a documented conclusion.

Write two candidate designs and place each on a small Pareto chart with quality on one axis and cost or latency on the other. A design is dominated when another is at least as good on every measured dimension and better on one. Eliminate dominated choices, then apply hard constraints such as privacy, authorization, or an SLO. For the remaining choice, define a rollback trigger before launch. Reversibility matters because estimates can be wrong: a feature flag, versioned index, pinned model, or shadow run can turn an uncertain tradeoff into a controlled experiment.