AutoInterpretation Finds Sparse Coding Beats Alternatives — AI Alignment Forum