[lyle_nel]

Yes it is completely useless against random strings as one would expect, but with word salad it not completely useless. For example, it can find words like "tequieromuchogerald" and "paulaalejandrailove" easily enough. I haven't fully explored its ability to do this, but I suspect starting with a dictionary of words instead of a random population might aid it in finding XKCD like random passwords. In general I don't disagree though, passphrases if chosen correctly can be completely intractable to a password cracker.

[saganus]

It might be interesting to use the word frequency of the target language as some sort of weight for each word, or maybe as a multiplier for the number of times that word will get pushed to the back of the queue, i.e. the words with the most frequency get to be repeated more than 1 time when pushed to the back, or something like that.

[lyle_nel]

I see what you are saying. To some extent this is already true ,since high frequency ngrams(substrings) already dominate the gene-pool if they happen to facilitate the production of a large amount viable offspring. That is why you would see swearwords with high frequency since that is what many people use for passwords(in the case of the myspace list).