Dmitry Vaintrob comments on The “best predictor is malicious optimiser” problem