It's more a showoff of a js idiosyncrasy - they found an ugly looking subset of characters that is Turing complete and wrote a translator to it.
If you're new to the concept of Turing tarpits, then this should blow your mind. On the other hand, this is a sufficiently advanced bit of CS thinking that no future employer should question the author's basic competency. I certainly have never written a translator.
You can get all the primitive values by taking advantage of unary plus, binary +, empty arrays, array dereferencing, function calls, and the standard strings returned by some basic expressions. The numbers are straightforward. The strings are dereferenced with numbers to get some individual letters. You can get methods by using array dereference on objects with strings. You get the rest of the letters with btoa and atob. Then you get eval, and you're off to the races!
I'm sure, but I think the encoder just goes token by token, eval'ing string conversions to get identifiers. You'll notice that for pretty simple expressions you get absurdly long strings out.
Well, okay, so it's still strictly speaking javascript but if we consider the functionally complete subset to be our 'target' language we end up in the same place, methinks :).
Is there a requirement that HN articles be ingenious? It's a novelty, that's all. And sometimes that's enough.