Edit: wouldn't it be awesome if Google (they did start this, afterall) would allow sites using hashbangs to auto-update all indexed URls
Where content is the main purpose ("websites"), they are overkill at best.
They can be useful though in functionality-first applications ("apps"), where the interface can be costly to build (too slow if you reload the whole "page" on each state change). Ideally, the app differentiates between fundamental states, represented in URLs, and transient states, represented in hashbangs.
Anyway, this is good news for Twitter, if they are really going through with it.
Apparently you can't log into Twitter without 64MiB of RAM, which is what Midori uses. Dillo and Links2 use 6MiB.
Hopefully this means they're switching to a more normal HTML interface and will become a web site instead of a DHTML application again.
they constantly had problems with their redesign and the last update involved removing the hashbangs
I believe there were further tweaks for additional pages that needed to be supported (whereas I had only solved, and most likely not very elegantly, the rudimentary issue) that took additional time before the hash bang removal was actually launched.
I check the site every couple of weeks because I was interested in how the redesign would work out. It seems they eventually got it right, I like it. A lot of the other blogs have home and article pages that are just too heavy and slow.