Of course after Peggle, Portal and Crayon Physics hit the mainstream it suddenly became apparent bigger doesn't equal better. Or more sales.
I've always loved the approach Introversion has taken with development. They're truly dedicated garage developers, spending more time trying to perfect a fractal tree than they really should. But I can respect spending an inordinate amount of time trying to wrap ones head around a concept like procedurally generated landscapes.
When I heard that Gamasutra was hosting an event with Chris Delay speaking on the topic of procedurally generated content, I definitely wanted to jump on the opportunity. While they had a fairly unrehearsed HP shrill asking the questions, Chris had some great points.
Chris emphasized that the main reason his titles have procedurally generated textures and meshes was because artistic content is just not a space he feels Introversion can compete within, since other companies have mastered that area. He saw it as neither a positive or negative thing, it's just the case for Introversion. Should artists be afraid? Chris doesn't think so. Procedural content cannot replace people, since it ultimately can't produce those unique items that make an environment distinct. While you can generate the landform that the world consists of (mountains, hills, streets, clouds, etc) it cannot add fine-grained details to the world.
You automagically gain several efficiencies with procedural content:
While the idea of using fractal algorithms for landscape generation or building trees, I hadn't thought much about procedural animation. Of course Spore uses it for their character builder, but introducing this as a new way of rigging meshes would again immensely help developers. Not needing an entire team of dedicated animators or texture artists would make things much more palatable.
There are tradeoffs of course, and Chris repeatedly mentioned that procedurally generated content requires a different way of thinking about memory management. Rather than loading assets off disk, you load them in memory at runtime - so you don't worry about texture compression, but you do now have to worry about LOD given to your algorithm and how much memory the resulting data structure will reside within. You can't let your procedure go willy-nilly and create too many verts.
Introversion's latest undertaking, Subversion, sounds interesting. Right now Chris describes it as more of a thought experiment, so who knows if we'll actually see it. But what he's pursuing is procedurally generating cities from 10 kilometer view all the way down to pens and desks inside a building. Not only does this employ a landscape generator for hills and mountains, but also will procedurally generate streets and buildings based on markets and traffic demand. Each procedural algorithm feeds its brothers, affecting its ultimate output. For example, more traffic makes more roads which can make bigger buildings.
One difficulty Chris found with this approach was that it was often ard to find out bad results - sometimes you would have cities being built on entirely one side of an area, with another being completely blank. Or sometimes a fire escape would open into nothingness on 30th floor. It's all a matter of finding a way to re-seed or compensate when these failures occur. Or maybe it just makes the whole concept quaint.