Evaluating SAE interpretability without explanations

Open in new window