Value alignment problem — LessWrong