It should be faster for higher-order AD, among other things. Keno talked about some of the theory in [1], but also has a Juliacon presentation later this week.
Seems kinda silly to post this to HN before the public announcement and before there's any actual documentation or good example materials available.
Why not just let Keno post this once he's ready? This stuff annoys me because it's probably the reason that this repo was private for the last year instead of public.