We should try to automate AI safety work asap — AI Alignment Forum