Explaining grokking through circuit efficiency — AI Alignment Forum