Watch the match change

This is a teaching example, not a reading taken from a real trained model.

Attention can run more than once. An earlier use of attention has already brought the clue from big or small into the number list for it. The bars show what happens when attention updates it again.

Switch the last token between big and small.

With big, the pattern "the thing was too big to fit" makes trophy the stronger match.

With small, the pattern "the container was too small to hold it" makes suitcase the stronger match.

The written token it stays the same. The learned relationship around it changes which noun receives more weight.