Sparse autoencoder feature labeling is AI interpretability's scaling bottleneck. Tsinghua University researchers posted ...