I'm not sure if position finding for the head is very separate from rendering, just because it is so smooth when he moves. I think head positioning is very separate from gesture positioning -- which is why cursors are jumpy. Maybe 13of40 is right on getting used to the interface -- maybe the user is holding his hand out too straight, not straight enough or something to that effect, but that still seems more of an interface problem than user problem. I do wish it was priced closer to oculus. At that oculus pricing I would probably get one, but at 4x it seems not quite there (interface and lacking apps).
13of40 - I think it is more "version 1" issues with the shake and a lack of the best human-to-computer interface rather than an "interface that nobody's used".
Put-that-there[0] is an interface from the early 1980's that seems like it would do well to reduce the janky interactions. Speech recognition probably isn't quite where it needs to be for a generic setup (as opposed to the structured setup of put-that-there).
Like I said, the computer to human side -- rendering virtual objects over a real field -- is beautiful. The quirks with incomplete map information aren't too distracting because we are good at filling in the gaps. The human to computer side needs work because it takes 30s to do a simple resizing operation. If you could say "make that screen 90 percent of my field of view" at any time the vid author wouldn't be complaining that getting too close causes clipping of his field of view.
That being said, maybe the hololens can be good enough nonetheless, because interaction latency is separate from pose/presentation latency and the latter is certainly far more make-or-break.
I suspect that the main achievement (or magic sauce) of the hololens is strict separation between position-finding plus persective-correct content rendering (which both together are quite high-latency, relatively speaking) on one hand and the final push to the actual framebuffer on the other, where a last corrective shift(/and maybe rotate(/and even more maybe zoom)) correction can be applied independently. The latter would be very low latency, based on the difference between camera frame at the beginning of the rendering process and most recent camera frame. This would be closely related (if not derived from) software-based shake reduction in cameras, which in turn might be traced back to the algorithms used in optical mice (which lack rotation/zoom).