Explore exposes every technical detail. Classroom gives a concise presenter-led walkthrough. Story builds the intuition as one visual narrative.
Running example: “I love deep learning” → 我 · 爱 · 深度学习
Self-Attention + FFN, bidirectional attention, contextual output H.
Tip: click a token (bottom of the stack) or a row of the matrix to change the query token.
Masked Self-Attention + Cross-Attention + FFN, generating one token at a time.
Two stacks, one bridge: the Encoder output H is wired into every Decoder layer.
How the Encoder output is injected into the Decoder — Q from the Decoder, K/V from the Encoder.
Tip: click a Decoder token (right) or a row of the matrix to change which token is doing the querying.
Padding mask vs. look-ahead mask — where they apply and how their shapes differ.
Bidirectional in the Encoder, causal in the Decoder — and what breaks if you swap them.