Interpretability of large models

The full issue offers readers a systematic map of interpretability methods — from single-neuron analysis to a mechanistic breakdown of reasoning chains in large language models. The authors compare approaches by required resources, reliability of conclusions and applicability to frontier models.

A dedicated section covers practice: how teams use interpretability to debug model behaviour, find undesirable strategies and prepare systems for a safety audit. The conclusion lists open questions and the Forum's direction for the year ahead.