Reading a model’s internal signals is not the same as explaining what causes its behavior, one reply argues. The commenter says decoding techniques have improved but still have limitations: some hallucinate, while others provide only word-based readouts of part of the internal signal. They suggest exploring methods that decode signals across an entire context, alongside studying a simple baseline: asking the model what it is thinking about. For causality, they offer an example: it is often possible to tell that a model is aware of being evaluated, but much harder to show whether that awareness influences its behavior.