Adam Scherlis comments on Consider using reversible automata for alignment research