Can we safely automate alignment research? — LessWrong