Convergent strategies of self-modification — AI Alignment Forum