“The "Towards Monosemanticity" paper demonstrated success in using sparse autoencoders to extract interpretable features from a one-layer neural network, including features for Arabic, Hebrew, and Base64.”