Be careful what you make easier
I have found much peace this year going offline to shelter myself from the whiplash of LLM companies jockeying for mindshare by cosplaying as DevRel. Every now and then something gets through the filter bubble — usually because it concerns Markdown, a topic near and dear to my heart and one increasingly in vogue in the age of LLMs as a message bus.
The most recent case: The Unreasonable Effectiveness of HTML.
The dynamic half of this essay is unimpeachable. What Thariq calls "Custom Editing Interfaces" — throwaway drag-and-drop kanbans, sliders for tuning an animation, side-by-side prompt editors with live preview — I totally agree! The ability to drive down the cost of throw-away snippets and playgrounds to ~zero is one of my favorite uses of his product.
But the implicit framing of the essay is "HTML is also strictly better than Markdown for all use cases", and I think he uses the unimpeachable bits to oversell the other half of the argument.
The single most important sentence in the piece, presented as a casual aside, is this: in practice, I've found I tend to not actually read more than a 100-line markdown file. The proposed response is a richer rendering format. The actual response — the boring, unfashionable, correct one — is a shorter, denser, and more comprehensible plan, both for humans and LLMs alike.
Be careful what you make easier. Making it more pleasant to claim to have skimmed a 600-line spec is not the same as fixing the fact that the spec is 600 lines, in much the same way biking 26.2 miles is not quite the same achievement as running them.
There are other soft parts of the essay, that I feel duty-bound to highlight:
"Ease of sharing" is presented as an advantage of HTML, with uploading to S3 offered as the suggested mechanism. It is hard to interpret such a take in good faith, given that the author wrote the post itself in Markdown and had to link externally to GitHub pages for the companion HTML artifact.
As Thariq admits in the FAQ, HTML takes 2-4x longer to generate, HTML diffs are noisy and hard to review (let alone trivially edit), and HTML is more expensive for LLMs to parse.
HTML artifacts are good! But I get worried about stuff like this, and not just because I'm a Markdown dork (the beauty of being a Markdown dork is that it doesn't really matter who else uses Markdown.) When you frame stuff like this in unrigorous absolutes, I am forced to worry about what else you frame in unrigorous absolutes.