Hennawi, Tsalmantza, and I had a long conversation about why our likelihood optimization does better at measuring quasar redshifts than our posterior-PDF optimization. In the latter, we use a highly informative prior PDF: That any new quasar spectrum must look exactly like some quasar we have seen before. This is the hella-informative data-driven prior. It turns out it is too strict for our problem: We end up over-weighting quasar spectra that fit the continuum well, at the expense of the narrow features that best return the right redshift. This raises a great philosophical point, one I used to discuss with Roweis extensively: You don't necessarily want to model all of the features of your data well. You want to model well the parts of your data that matter most to your questions of interest. So if we want to use ultra-informative priors, we ought to also up-weight the informative features of the data, and remove the uninformative. This is done, traditionally, by filtering—which is terribly heuristic and hard to justify technically—but which has been done more quantitatively in some domains, once notably by Panter and collaborators in MOPED.