Interpretability
Research into what happens inside a model: which parts represent which ideas, and how they produce an answer.
Also: mechanistic interpretability, explainability
Models are trained, not written, so even their builders do not know in detail how they work. Interpretability researchers study attention heads, neurons and directions in hidden states to find circuits that perform specific tasks. In 2024 Anthropic used sparse autoencoders to find millions of interpretable features in Claude 3 Sonnet, including one for the Golden Gate Bridge that, when amplified, made the model talk about the bridge constantly. The hope is to audit models for hidden goals or flaws; the field is young and its tools are partial.