When I was working on mSIGNA I became a little torn on the whole filtering mechanism. I fully support connection filtering...but in practice always run my own full node instances to connect to due to the three fatal flaws: 1) no mechanism for short proofs of tx nonexclusion, txout unspentness, block validity, nor the ability to find the first instance of the use of a scriptPubKey without full blockchain scanning, 2) poor privacy, 3) lack of incentives to run servers.

I always felt that BIP37 was necessarily a step towards a client/server architecture.

Having said that, I have found the filter mechanism useful, if only because no "special" server is required. However, in practice I'd rather make the distinction between trustless peers and a client/server model more explicit.