Thanks to Mike Stay for finding me an on-line copy of Klamkin's article "On the teaching of mathematics so as to be useful". The relevant excerpts are on pages 134-135 and page 159. The former excerpt reads as follows: "In many courses, once a method is established in theory for solving a class of problems that supposedly finishes the topic. However, actually obtaining a numerical answer to a specified accuracy may be an entirely different and difficult problem. For example, in evaluating definite integrals numerically, Simpson's rule may be derived and illustrated on several 'nice' functions. However, if one assigned an integral such as $\int_{0}^{100} x^x \ dx$ to be evaluated numerically to within 10%, almost all students will use Simpson's rule 'blindly' and be lucky to come within 500%. The difficulty here is that x^x is not a 'nice' function since it increases too rapidly. Similarly for the integral $\int_{10}^{\infty} e^{-x^2} \ dx$." The passage ends with a footnote leading the reader to an explicit calculation on page 159 (which would be a bit of a pain to include here, and would not be very readable even if I did). Jim Propp