It has been claimed that flat, linear presentation of comments appears to work better for humans than tree-structured comment threads. Without getting too deeply into whether this is true (and if so, why), I would like to offer an observation.
Conversation is never a tree; it is a general directed acyclic graph. In reality, in the commenter's mind, every comment potentially implicitly responds to an arbitrary subset of preceding comments, not to a unique parent and its chain of unique transitive ancestors.
Tree-structured threading — sometimes (erroneously!) called "true threading" — artificially imposes a tree structure on this graph. Flat, linear comment systems do not: each comment appears after all those that precede it topologically in the DAG, and it is up to the reader to reassemble the DAG based on the comments' contents.
It is true, of course, that flat comment systems fail to reify all DAG edges as explicit metadata. However, the nature of these edges is quite subtle and capturing them all explicitly is intractable. Often a comment "responds" to previous comments in indirect ways — for example, simply by omitting some aspect of the argument that has been covered by a previous comment.
(Prompted by a TC article linked off HN.)