Sparsify: A mechanistic interpretability research agenda — AI Alignment Forum