One studies neurons, attention heads, and features plus the circuits linking them; causal interventions confirm each component's role.
Large models are 'black boxes', making their behavior hard to trust and audit.
