For example, when you load a page on reddit, it has to check for a vote on every single comment on the page to see if you voted on it. Chances are 99% of the time you didn't vote on an item. Using the built in bloom filter in Cassandra, it can very quickly tell if you voted on an item without having to hit the data store most of the time, which saves a ton of time.
(As a side note, there is an extra optimization that makes things even faster -- reddit tracks the last time you voted on anything as an attribute of your user, so if the page is newer than your last vote, it just assumes you couldn't have possible voted on it, skipping the lookups altogether).