New 80,000 Hours problem profile on the risks of power-seeking AI — LessWrong