tom4everitt comments on “Designing agent incentives to avoid reward tampering”, DeepMind