How models are trained

The builder's-eye story of how an LLM is made: pretraining (self-supervised next-token prediction over a huge text corpus) produces a raw base model, then post-training turns it into the assistant you chat with. Explains why a model has a knowledge cutoff, why it's biased, why it doesn't learn from your chats, and why served model versions are frozen. The conceptual story for builders, not the ML internals.