I don't think you understood the comment you're replying to, but are doing some kind of pattern matching.

The point it, he does it much much better than ChatGPT, to the point that his understanding of the comment appears seamless, whereas ChatGPT's is crude and mechanic in comparison...

I'm just referring to the fact that they said "But it does not understand anything here" even though the parent comment made no such claim. The parent was saying ChatGPT made a bullshit answer that even tricked OP into thinking there was understanding.

I think the disagreement is about the fact that it would be considered specifically impressive. It doesn't seem more impressive than everything else ChatGPT does when it repeats a boilerplate and tries to find parts of a joke using probabilistic weights, or tries to categorise the joke into either puns or something else.

Maybe. But human's pattern match capabilities regarding jokes seem to still be on another level compared to the attempts this LLM makes.

Well, it does try to identify the setup, punchline, and what the subversion of expectations is. It does provide an (often wrongheaded) explanation of what's being subverted. I find that impressive.

Setup is in the beginning. Punchline is in the end. Subversion of expectations is a standard template.

But, like you point out, it fails to grasp the subversion.