You type a sentence. The AI returns an answer. Here is every step in between — 19 stages, one concept at a time.
A sentence is split into tokens,
mapped to IDs and
embeddings,
then token representations gather context through
attention.
A stack of Transformer blocks
produces representations used to predict next-token scores.
Generation
selects a token and repeats the process;
post-training
can improve instruction following —
and a surrounding product system can turn the model into an
AI assistant.
Not magic: learned parameters, computation, and a token-by-token generation loop.