The blue-minimising robot and model splintering — LessWrong