How it works

The authors train SAEs at many scales and measure reconstruction fidelity and interpretability. Power-law relationships emerge from those measurements.