This may not be the new role of the PC to some, but I sure hope it could be mainstream.
Sounds like general-purpose Nest + a peer-to-peer (fully) homomorphically encrypted (and an ideally performant encryption scheme) network of PCs for maximum security from prying eyes with the desired functionality... unless I'm misunderstanding you. I've had similiar ideas. I think you can get pretty far right now without fully homomorphic encryption by using non-interactive zero-knowledge proofs + capability-based security to ensure that only the individuals who should get information do get information, and the people that shouldn't, don't. See the Bitcloud Project for a beginning along this path. Many of the people on #bitcloud-dev agree with me (and presumably you) that one of the most important projects to undertake after Bitcloud is complete is a P2P Plan 9, which is pretty much what the above design entails. To clarify, I mean a conceptual P2P Plan 9. The only feature necessarily alike to Plan 9 in that case would be something akin to a peer-to-peer version of the /net interface. I do agree with CrazyEyeJoe that the AI aspect of the idea will probably be in a decade. I'd be willing to bet that the decentralized aspect will be available in five years, give or take two.
Security is not the main motivation for this project. Not being an expert, I assume that anything built should be inherently secure. Security is not a feature, it is an essential component of the underlining project. Building a secure application is complexe yet it should be virtually invisible to the user. On the end user point of view, things just work. I did give the parsing the web as an example, but that's just one aspect of things. Yes it is very complicated and it might require a few years to optimize it to a useful tool. At the end of the day, it is not impossible. I haven;t heard of the bit cloud project before but it seems very interesting Bitcloud Project - fixed the link. It also contributes to the idea of your PC being yours again.
I think that a lot of this sounds like it would need sophisticated machine learning/data mining algorithms. The reason this doesn't exist is probably because it's difficult to develop these algorithms without having huge datasets and user feedback which comes from, you guessed it, companies having your data. I'm not saying it's impossible to do, but it's definitely more difficult to adapt existing datasets to your software than to collect them yourself, in a format that you dictate. Maybe in a decade.
It's true that your computer will not be processing as much information as big as datacenters. But the debate of how relevant to you the results will be could be solved by tweaking or updating the algorithm. Yes it will take time to find the best solution, but remember, this is just one feature of your machine. It could be solving other problems that have nothing to do with search. It all depends on what you want the computer to do for you.