Talent Needs of Technical AI Safety Teams — LessWrong