Black box Interpretable
Internal process: not visible
Prompt "What is love?" ? ? ? internal process unknown Token embedding Attention: concept "affection" ↑ activated Internal state: curiosity 0.74 · warmth 0.61 Output projection AI MODEL Answer "A deep bond…"

Interpretability research tries to map what each internal layer of a model is doing. Some layers are well understood; others remain opaque even to the researchers who built the model.