Some motivations to gradient hack — AI Alignment Forum