I'm a big fan of that architecture, if only because it greatly reduces the payoff of a successful attack. When everything is stored unencrypted (or with a common master key), there's an absolutely massive payoff for the hacker who breaches the security.
Dropbox is incredibly easy to use, that's where its power comes from. It's "secure enough" for casual use.
Additionally I believe you can't share data between users with tarsnap.
There is a market opportunity for a corporate-level secure data exchange infrastructure.
The overall design, including how it is possible to do secure delegation, is fairly well described in a paper from the 'Storage Security and Survivability 2008' workshop - http://tahoe-lafs.org/~zooko/lafs.pdf
That's besides the point. Tarsnap is more than useable enough for its target audience, which is why cperciva (rightfully) doesn't give a rat's ass about making it as user-friendly as Dropbox and its ilk.
The two issues are completely separate. Implementation of client-side encryption and user-friendliness are not mutually exclusive. Tarsnap is simply proof that client-side encryption is entirely within the realm of possibility for cloud-based data storage services.