I gave a quick internet-read to this blogpost by Andrew Gelman. The core of the idea is that to estimate an interaction effect a huge increase in sample size is needed compared to estimating a main effect. That is, if you power a study on main effects you could be very underpowered if you want to estimate an interaction. This is interesting to me since I work with human subject researchers that are greatly limited on their sample size for practical reasons, and they often want to estimate interactions.
Another shiny object to chase.