I will describe machine learning approaches that are being used to design T-cell vacccine immunogens for HIV. I will focus on one approach that has identified a hidden success in the recent vaccine trial that otherwise failed. In particular, the MRKAd5/HIV trivalent vaccine failed to reduce HIV viral load in participants who acquired infection, despite induction of HIV-specific CD8 T cell responses. I will show that participants who targeted specific "good" epitopes upon vaccination went on to have lower viral loads after becoming infected. The results suggest that the design of a successful vaccine immunogen may hinge on the inclusion of good epitopes and the exclusion of others that distract the immune system from targeting regions associated with control of viral replication.