03 · How Models Learn
Training Data
What a model learns from
The text, code, images, audio, video, synthetic examples, and licensed or private datasets used to train a model. Exact mixtures are often undisclosed; quality, coverage, consent, and recency all affect results.
Concrete example
A model that never saw events after its training cut-off simply doesn't know them — which is why up-to-date answers need search or RAG.
Why it matters
The source of both capability and controversy — the copyright suits and licensing deals are all fights over this layer.