My understanding is this: the genesis for fold (as opposed to foldl) is from the linked talk: http://vimeo.com/6624203.
If fold (a concurrent reduce) is used instead of foldl (a sequential reduce), then the mapping function can be pushed down and applied concurrently to the various elements. A generic mechanism can then be used for dispatching concurrent operations. These functional operation are composed of first mapping (and/or) filtering and then reducing.
My understanding is this: the genesis for fold (as opposed to foldl) is from the linked talk: http://vimeo.com/6624203.
If fold (a concurrent reduce) is used instead of foldl (a sequential reduce), then the mapping function can be pushed down and applied concurrently to the various elements. A generic mechanism can then be used for dispatching concurrent operations. These functional operation are composed of first mapping (and/or) filtering and then reducing.
[EDIT: typo and re-phrase]