Every other explainer here zooms into one stage. This one zooms out: type a sentence and watch the entire pipeline run, stage by stage, all at once.
Read it top to bottom, it’s the whole journey from raw text to a guess:
- Tokens, your sentence is chopped into chunks.
- Embeddings, each token becomes a vector (the coloured cells; warm = positive, cool = negative).
- Attention, every token looks back at the others; brighter cells mean “I’m paying attention to you”. This is the transformer’s core move.
- Next word (from the last token’s vector, the model scores candidates and picks) shaped by temperature and softmax.
Change the sentence and the whole chain recomputes. It’s a schematic (the shapes and relationships are real, the exact numbers are illustrative) but this is genuinely the path every word takes through a model. For the real thing per stage, follow the five-part path.