New scorecard evaluating AI companies on safety — LessWrong