Evidence
- Sep 8Opacity per past token = attention weight times value-vector magnitude, aggregated over heads and summed over layers; author calls it a heavy simplification
- Sep 8Commenter finds it simpler than BertViz for understanding inference and a good starting point for learners