I'm also curious about scaling and collisions. Not only do you have a birthday problem with normal usernames, but what about special classes? Why do I not take all "nyt" and similar usernames to honeypot the actual NYT's contact?
It's not fully baked, but my expectation is that it will work similarly to how the .eth, namecoin, and other systems work, where you'll be able to register a user discovery service on a blockchain which the clients will recognize and use for searching. In this model, NYT registers "nyt.xx" and "rcarback" pops up in the interface as "rcarback@nyt.xx". As it stands, we've rolled out a basic version with a single central user discovery point for now.
That's a useful feature. Others (above) are telling me this is useless (I disagree). Are these identities separate?
The discovery service sounds useful