Move the query. Watch meaning bend.
The source asks readers to infer a timeline from a repeated phrase. Explore a transparent toy transformer layer, then separate association from the dated evidence that could actually verify the claim.
The visual runtime could not load. The evidence lesson remains: attention weights show association, not chronology.
Click or touch a token to move the query.Scaled dot-product attention / 3 transparent features
MISSION: EXPLORE 3 QUERY TOKENS
Which words pull together?
Select a token. The arcs will recompute from its query vector to every key vector.
ACTIVE QUERYAttention
Waiting for the first map...
1 of 3 distinct queries explored
A strong link is not a dated source.
WHAT THE MAP COMPUTES
Each arc is a normalized similarity between one query vector and the visible key vectors. Change the query and the distribution changes.
score(q,k) = q dot k / sqrt(3)
weight = softmax(score)
weight != truth
weight != chronology
weight = softmax(score)
weight != truth
weight != chronology
Transfer the lesson.
The map can surface relationships inside the sentence. What would you need to verify that the quoted phrase appeared in a book ten years earlier?
Choose the evidence that can establish chronology.