Rafael Harth comments on DeepMind article: AI Safety Gridworlds