A technical note on bilinear layers for interpretability — LessWrong