I'm not saying that slowing down the checking of a password has no value; Clearly it has much value. Without it, even a strong password would be brute forced quickly. I'm just saying, I am missing how it solves the problem of weak passwords.
The real problem here is that I know I can't remember a sufficient degree of entropy. Sure, there are all sorts of ways to fake it; take the first letter of every word in a sentence, but that's not exactly random, either. I know a guy who, for a pass phrase, would use a 24 line block of text from some project Gutenberg title. I doubt that was any more secure than 4 or 5 random characters.
10000 * 1s = ~2 hours cpu time, but that's a terrible password -- common words plus 3 digits means 1000 * 10000. 26^6 * 1s = ~10 years CPU time -- a few thousand bucks on EC2 -- that's getting better for what is still a pretty weak password. In any case, straight MD5s are super-fast. At 1e5-1e6 MD5s/second or more on a modern CPU, paying the compute cost for a 1 cpu-second KDF can be several orders of magnitude stronger than MD5(password+salt).
Can you tell me more about allowing the user to cache the derived key within the context of web applications? that would mitigate the limitation I was describing (where how slow you made the password check was limited by what delay a user would tolerate when logging in.)