Pretraining
A base model is trained on trillions of tokens with one objective: predict the next token. Everything else — grammar, facts, translation, some reasoning — is an emergent side effect of compressing that corpus.
Scaling laws and compute
Loss falls predictably with model size, data size and compute. The Chinchilla result showed most early models were undertrained: for a fixed budget, more tokens with a smaller model beats a giant model on few tokens.
Alignment
A base model completes text; it does not follow instructions. Supervised fine-tuning on instruction/response pairs adds that, then preference optimisation — RLHF with a reward model, or the simpler DPO — steers towards helpful, harmless answers. Alignment shapes style and refusals, it does not add knowledge.