Intensive technical knowledge, I wish he gave a bit of overview, its getting long. I'm guessing he figured out how to make it faster or he was just exploring it the bad driver code.
He shows some ways to poke memory managed by the iphone driver and saturate the gpu using only ~20% of cpu instead of the 64% that you use with the conventional OpenGLES path.
This necessarily introduces some fragility since you're essentially poking at magic locations in memory in a totally non-safe manner, but he seems to think the driver's internal structures are stable enough that this isn't a problem.
In fact, if enough top quality games do this, then the stability might be a self-fulfilling prophecy, since messing with the driver internal structures' layout would break all the really nice looking games.