Loading slide
Loading contents...
Press run input. The drawing shows the input moving through the model's layers.
Then press generate reply. Watch each new token join the context before the next token is produced.
This is a simplified view of ordinary token-by-token generation. It shows the dependency between output steps, not a measurement of hardware speed.
A short invitation and a long report may begin with the same notes. The report still requires more output steps.
# why longer replies cost work
Why can a longer reply cost more? It keeps the hardware generating for longer. Model size, input length, and serving efficiency also affect the work. The linked chapter sources explain these costs in more detail.