[Interim research report] Evaluating the Goal-Directedness of Language Models — LessWrong