Weight-sparse transformers have interpretable circuits — AI Alignment Forum