John Gruber

at least in their ability to emit markdown, forget about image and video generation, but a corpus of text that was produced for humans to read.

So LLMs consume text trying to match patterns the way humans do.

And so the fact, going back to your question, that I tried, my number one priority for markdown was to make it readable, is that it is...

and readable to someone who doesn't even know what Markdown is, is exactly how LLMs parse text, like that they can parse noisily written text.

Somebody can just bang away in a Facebook text input field that doesn't support anything like formatting and just use weird punctuation characters to sort of fake underlines or something.

And LLMs will parse that.

Well, that's what Markdown is.

So of course, LLMs really like it.

And they like sending it, too, because it's lightweight and fast.

And it's just little characters.

It's not complicated.

So for example,

Somebody did some studies.

It's much better at LLMs, I think, to this day are still better at emitting Markdown than they are at JSON because JSON has very persnickety rules.

Like a JSON file, you get a character wrong, and then the whole file is bad.

Markdown, if you forget to end an italics run, who gives a crap, right?

Yeah.

Yeah.

Doesn't matter, it's still good.

And so LLMs making one little tiny mistake or something in the markdown they admit, it's still fine, right?