To fix precision problems that sometimes arrise when having only front or back faces cast shadows into a shadow buffer, he thought up the idea to create a shadow depth buffer for the front facing polys as well as one for the back facing polys. For every pixel find the mid-point between the front and back facing depth textures and then use that to compare against instead of only the front or the back.
Of course this has the obvious draw back of having to render everything twice to compute a single shadow depth buffer (something you may have to do multiple times per-light if the object creating the shadow resides in multiple sides of the cube map.)
Still a very powerfull technique though. I'll post some more later about some of the other things he talked about.