So code up a prototype:
+ Implement code that computes and publishes 'balance blocks' and 'balance block hashes'. Convince a couple people with extra download bandwidth to run it.
+ Modify one of the bitcoin implementations to download the latest 'balance block' from some trusted place at startup, and use it if transactions/blocks can't be found in the traditional block database.
+ Extra credit/paranoia : query a couple of trusted places for the balance block hash, and make sure it matches the hash you got.
+ OR: randomly spot-check the balance block by requesting blocks in the traditional way, and make sure the balance block doesn't list any outputs as unspent that are actually spent.
You don't want bitcoin address balances, there are no addresses way down deep inside. You need to know which transaction outputs have not yet been spent, and the value of those outputs.
I'm not excited about this proposal, because I think it is solving a problem that doesn't need solving yet, and my priorities for bitcoin continue to be wallet security and network stability, not making it quicker for newbie solo miners to get a full blockchain so they can start validating transactions/blocks.