Transformer and System Layers
The Transformer in one picture
A deployed assistant has layers
The
capital
of
Australia
is
[ h₁ ]
[ h₂ ]
[ h₃ ]
[ h₄ ]
[ h₅ ]
Context-dependent representations
attention
Scores
Next-token
Base Transformer
Post-training
Retrieval
(optional)
Tools
(optional)
Attention combines context; it is not a truth check.