huge flaw early in the article is assuming world selection is done using 5-dice, that is something used in academic assignments and not in most phrase based password generators.

Additionally, the cracking methods listed here are largely impractically for 90% of passwords out there, you aren't going to run that many iterations that fast, on anything but a compromised password file, you won't compromise any active system with it.

Basically this article is constructing a very narrow limited word set (7700 words, as opposed to a standard dictionary of 300,000+ words) that reduces randomness of phrase set, a very academic method of word selection, and a cracking method that is only remotely feasible in a very limited subset of possible compromises.

Additionally as acknowledge by the author of the article.

For perspective, the vocabulary of an average-educated native English speaker ranges from 20,000 to 35,000 words, while comprehensive unabridged English dictionaries contain 200,000 to 300,000 words. This means that a seven-word passphrase drawn from a sufficiently rich and well-structured wordlist could offer the same cryptographic strength as a highly random and complex password.

Although random strings remain the most compact way to achieve high entropy, passphrases offer a compelling alternative — as long as they are generated with true randomness from a well-designed, sufficiently large wordlist.

So yes, done poorly in perfect conditions in a vary narrow subset of use cases, but none of those are the reality of a situation.