Thoughts on implementing corrigible robust alignment — AI Alignment Forum