One misconception is worth clearing up before going deeper: chatting with a model is normally inference, not training. Training changes learned numerical parameters. Inference uses those parameters, together with request-time context, to produce output. A product may separately save conversations, build memory or use data in later training under its terms, but those are separate mechanisms.
During pretraining, a model learns from broad data against an objective such as token prediction. A loss measures error and an optimizer updates parameters through gradients. Post-training then shapes behaviour with demonstrations, preferences or reward signals. These processes change the model itself. RAG and conversation context do something different: they change what the deployed model sees when a request arrives.
Fine-tuning continues training from an existing model for a narrower purpose. Some methods update only a small set of additional parameters rather than all of the weights. The result still needs held-out task evaluation and checks for regressions in behaviour that previously worked. [9]
Distillation trains a model using a teacher's outputs or other supervisory signals. It can transfer useful behaviour to a smaller model, but it also inherits limitations from the teacher and training data. Quantization reduces numerical precision for storage or computation. It is a separate choice from teaching the model a task, and it can change quality.
Northstar illustrates why the distinction matters. Leave rules belong in an authoritative source and retrieval pipeline. Employee balances belong in the HR system. Training either into model weights would make freshness, permission changes and deletion harder to manage.