The thing is, the Google Car driver isn't a general intelligence. It's quite specialized. Watson, OTOH, is a much more general intelltigence. But it still doesn't have a hierarchy of goals that allows it to override what it is told to do. I'm not sure, however, that that counts as intelligence rather than something else.
FWIW, AI programs come up with ideas all the time. But they are designed to prune them to match their goal structure. (So are you, but your goal structure is much more self-centered.) Coming up with idea is not a problem, coming up with appropriate ideas, and knowing that they are appropriate is still a problem. Watson appears to be addressing that problem. Currently an incarnation of it has learned to diagnose cancer better than most doctors. An earlier incarnation learned to play Jephrody better than most humans. (Lots better.) And the hardware requirements have been shrinking. (I'm not sure how much is hardware improvement and how much is program improvement.)
I expect that a near-term target of Watson will be middle-management...though I also expect that it will be presented as offering advice rather than as replacing them. Basically what it will do is allow one manager to directly manage an increasing number of workers efficiently. This will prepare it for a career as an advisor to politicians.
Do note, however, that this isn't what he was talking about. He was talking about Cyborgs. These are held back by two things: The lack of a long term neural connector that won't destroy the neurons that they connect to, and the fact that installing significant Cyborg modifications requires surgery. I expect the first problem to be solved within the decade, but as for the second...