FREE LLMs & agents LESSON · LLMs & agents
Shape model behavior deliberately
Pretraining, post-training, inference, and capability trade-offs
An LLM product inherits behavior from several distinct optimization stages.
Pretraining learns broad next-token structure from a data mixture; supervised post-training demonstrates desired interaction; preference optimization shifts outputs toward comparative judgments; tool and domain tuning add narrower capabilities. At inference, decoding, context, caching, quantization, batching, and serving limits change cost and behavior. Evaluation must identify which layer can plausibly fix a failure.
Do not ask prompting to repair a missing capability or retraining to repair a deterministic permission boundary.
Preference data can reward surface style over underlying truth.
Annotators may prefer confidence, length, agreement, or familiar framing, and a reward model can amplify those patterns. Preserve objective task checks, disagreement analysis, diverse evaluators, and adversarial examples. Inspect regressions in refusal, uncertainty, cultural context, and minority cases rather than accepting one aggregate preference win.
Optimizing what evaluators like is not identical to optimizing what is correct or useful.