Video explanation
Using 17×24, a ring holds prompt, intermediate steps and output while a central layer stack computes next-token probabilities. An exploded view closes the demonstration.
Context to next token
The model uses context to produce a probability distribution for the next token, then selects a token to continue.
Attention links information
Attention relates sequence information as layers transform representations. The tower and lines are a metaphor, not hardware layout.
Intermediate steps and answer
The demo distinguishes intermediate steps from the answer. More steps may help computation, without proving all models reason through the same sequence.