The connectionist triangle model of reading aloud proposes that semantic activation of phonology is particularly important for correct pronunciation of low-frequency exception words. Our consideration of this issue (Woollams, Lambon Ralph, Plaut, & Patterson, 2007) reported computational simulations demonstrating that reduction and disruption of this semantic activation resulted in the marked deficit in low-frequency exception word reading that is characteristic of surface dyslexia. We then presented 100 observations of reading aloud from 51 patients with semantic dementia (SD) demonstrating a universal decline into surface dyslexia, a phenomenon we termed "SD-squared." Coltheart, Tree, and Saunders (2010) have more recently provided a simulation of the SD-squared data within the dual route cascaded (DRC) model, achieved by varying the amount of damage to components of the lexical and nonlexical pathways. Although they suggested that these simulations provide a closer fit to the SD patients' reading data than our own, we demonstrate here that this is not the case. Moreover, we argue that the connectionist triangle model account has substantially greater explanatory and predictive power than the DRC account. © 2010 American Psychological Association.