Daniel Kokotajlo comments on The “best predictor is malicious optimiser” problem