Research directions Open Phil wants to fund in technical AI safety — LessWrong