Granted, at least when viewing a MPR there could be an optional smoothing algorithm running, since what happens in the 'future' is perfectly well known in that case. So instead of simply letting the physics run their course with the last packet input, you could alter the input between the packets so that the car position matches the 'future' packet as close as possible when it is read from the MPR, minimizing the huge abrupt warps without resorting to the cars floating around magically. Not easy but definitely possible.
The realtime network code could also be improved to avoid warping, at the cost of delaying the car position by at least one packet. I think it would actually be preferable to have slightly more delayed reactions of your opponents than them driving or warping into you for a split second before continuing their real path.
However, the current implementation already works well enough, meaning we probably won't see that changed for a long time, if ever.