Cliffski's tweet about megatextures being a brute force approach reminded me of Chris Crawford's book "Chris Crawford on Game Design" (a good read). Megatextures allow artists to create huge, rich images to represent landscapes and backgrounds; however the images will be fixed. The opposite approach is to use procedural texturing which allows for virtually infinite backgrounds but only within the constraints of the algorythm's sophistication.
In his book, Chris talks about Process Intensity versus Data Intensity. Most modern games seem extremely biased towards Data Intensity, even though they have complex physics, game systems and artificial intelligence. This unfortunately results in games that are more like movies with limited interaction (the single player aspect of COD4 was way too short in my opinion).
I have a preference towards procedural generation for maps and enemies, as this makes for more emergent gameplay and unexpected events. Gravity Core has procedurally generated maps which worked pretty well. The difficulty is creating interesting scenarios and set pieces and reducing repetition - I didn't completely succeed with Gravity Core in these areas. I would like to take this further in future games.
The next (free) game won't be procedurally generated as I'm aiming for a simple vertical shooter however I do have another procedural game cooking in my brain.