Transformer — Data Flow Through Every Block
Follow a 4-token prompt as it becomes vectors, flows through N identical blocks that read & rewrite one shared residual stream, and ends as a next-token prediction.
at rest
block / 6 idle · press Play

Input prompt

Block stack (depth)

Every block has the same structure and its own learned weights. Depth = how many times the stream is refined.

Residual stream — the data

4 tokens →d_model = 6

Operation

what changes