In recent years, there has been a great deal of excitement about 'big data' and about the new research problems posed by a world of vastly enlarged datasets. In response, the field of Mathematical Statistics increasingly studies problems where the number of variables measured is comparable to or even larger than the number of observations. Numerous fascinating mathematical phenomena arise in this regime; and in particular theorists discovered that the traditional approach to covariance estimation needs to be completely rethought, by appropriately shrinking the eigenvalues of the empirical covariance matrix. This talk briefly reviews advances by researchers in random matrix theory who in recent years solved completely the properties of eigenvalues and eigenvectors under the so-called spiked covariance model. By applying these results it is now possible to obtain the exact optimal nonlinear shrinkage of eigenvalues for certain specific measures of performance, as has been shown in the case of Frobenius loss by Nobel and Shabalin, and for many other performance measures by Donoho, Gavish, and Johnstone. We describe these results as well as results of the author and Behrooz Ghorbani on optimal shrinkage for Multi-User Covariance estimation and Multi-Task Discriminant Analysis.