It's just time consuming, and compute time has gotten a lot cheaper. Advancement in GPU computing, cheap clouds, etc have really changed the game on this one since the invention of MD5 (which is why unsalted MD5 is such a bad idea now).
It's important to note that brute force is always technically possible, and the best hashing algorithms just consume more time. When we say "impossible" we really just mean "impossible to bruteforce before the heat death of the universe".
> And isn't there the possibility of collisions, so that even if you find a string that maps to the same hash, it might not be the original password?
That's correct, but if the hash is what's stored then the server is checking hash vs hash and a collision will result in login.
This is all coming from one computer security course, but I'm sure someone on HN can correct me if I've got any of this wrong.
I think strings that maps to the same hash are just inintelligible garbage. If you find something that looks like human then it's certainly the original password.
[1] https://www.youtube.com/watch?v=I915tOiR9sM
If it weren't a song title, it would probably have been impossible to crack. That sentence has 12 words. People say that most English conversations only use 3000 words. 3000^12 is 2^138. It has quite a bit more entropy than what we can crack nowadays. Besides, "stripper" isn't part of the 3000-word dictionary.
http://arstechnica.com/security/2013/05/how-crackers-make-mi...
Not really. 2^128 is a stupefyingly big number. We'll likely never find a random collision in MD5. We know how to generate collisions in MD5, but this requires knowing the preimage (e.g., the m in `h = H(m)`).