Ours is essentially the twitter model, with a few tweaks. I think a hybrid approach -- flat for reading, but limited expansions on demand for context -- is best.