Professor Davies is to be commended for his efforts to keep statistics and statisticians honest, a truly herculean (but hopefully not sisyphean) task. While one may perhaps not agree with everything said in the paper, it is stimulating and thought-provoking in the truest sense of the word. Professor Davies deserves the deepest gratitude of the statistical community for sharing his thoughts even if they may be unsettling to some.The paper touches on many aspects and problems of statistics, classical ones such as the (mis)use of asymptotics as well as novel ones such as a framework for statistics that does not start from the premise that the data are generated by a probability model.With regard to (mis)use of pointwise asymptotics as discussed in Section 6, I agree with the point of view taken in the paper. Our statistical forefathers were very well aware of the dangers of pointwise (as opposed to uniform) asymptotics, as is, e.g., documented by research in the middle of the 20th century (Hodges, Hajek, LeCam) that has led to a solid foundation of the asymptotic efficiency concept. The following quote from Hajek (1971, p. 153) nicely expresses this:"Especially misinformative can be those limit results that are not uniform. Then the limit may exhibit some features that are not even approximately true for any finite n".A prominent example where pointwise limit results are indeed highly misleading is provided by model selection. In recent years, there has been a wave of papers establishing what has come to be called the "oracle" property of an estimator (which should not be confused with the notion of an oracle inequality). In its simplest terms this essentially means the following: suppose that we are given a parametric model M 2 and a lower-dimensional submodel M 1 (e.g., M 1 is obtained from M 2 by imposing some zero restrictions). Then an estimator has the "oracle" property if it "adapts" to the smallest true model asymptotically, i.e., if it is asymptotically normal with an asymptotic variance-covariance matrix that equals the asymptotic variance-covariance matrix of the restricted maximum likelihood estimator if M 1 is true, and equals the asymptotic variance-covariance matrix of the unrestricted maximum likelihood estimator otherwise. Such estimators are typically obtained by some form of consistent model selection, either by a classical model selection criterion such as BIC or by a penalized maximum likelihood method such as SCAD or adaptive LASSO, etc. The "oracle" property seems to tell us that there is no price to be paid for not knowing whether or not the restrictions defining M 1 are satisfied. Of course, this is too good to be true. The crucial point is that the asymptotic framework underlying the "oracle" property is only a pointwise one. That the "oracle" property is indeed misleading already becomes clear if we recall what we have learned from the Hodges' estimator more than half a century ago. (It is somewhat unsettling that the top journals in our profession publish papers using this