Single SAEs trained on the residual stream activation vectors from every transformer layer simultaneously using tuned lenses.