Training
Teaching a neural network via gradient descent — example data + loss function + backprop. Multiple phases in modern LLMs.
Classic pipeline: Pretraining (massive unsupervised on raw data, builds base model) → SFT (Supervised Fine-Tuning on instructions) → RLHF/RLAIF (alignment with human/AI feedback) → RLVR (Reinforcement Learning from Verifiable Rewards, for math/code). Cost is front-loaded: GPT-4 is said to have cost ~$100M to pretrain. Later phases are cheaper but critical for quality.