HTML > Markdown

Elad Gil posted something this past weekend that's obviously true once you see it. His rough hierarchy of how AI knowledge is distributed today: people inside the major AI labs are three to four months ahead of Silicon Valley startup engineers. SV founders are three to six months ahead of New York. New York is six to twelve months ahead of the rest of the world. Most people, he argues, are a year or two behind the frontier and have no idea.

He closed with the Heinlein line that William Gibson made famous: "The future is already here, it's just not evenly distributed."

What Elad described is the ripple effect of technological diffusion. It's probably been true for a long time, and it's more noticeable now because the pace has compressed. This week I want to walk through one specific example of what that diffusion looks like in practice.

A few days ago Thariq, a member of technical staff at Anthropic, published an X article championing HTML. Thariq writes primarily about Claude Code and is one of the best technical follows on X if you want to see how the people building these tools actually use them.

His argument: he's stopped writing markdown files for almost everything and switched to having Claude generate HTML instead.

Needless to say, this post caught my attention. I've been banging the table on markdown libraries for weeks. Build a knowledge base. Put your work into .md files. Assemble a company brain the model can read and reason over. I still believe in all of that. But Thariq's article challenged my thinking in the best way possible.

Quick orientation for non-technical readers. Markdown is a plain text file with a few formatting symbols for headings and bullets. If you're using AI regularly at work, you're already generating markdown files whether you realize it or not. Skills are markdown files. The memory feature in Claude generates markdown files. Project instructions and system prompts get stored as markdown. Open your settings and you'll find them.

HTML is the language browsers read. Most non-technical professionals only encounter HTML by accident. You try to save an image, you end up with a file that opens in your browser. That's HTML.

Markdown gives you structure. Headings, lists, the occasional table. HTML gives you structure plus presentation plus interaction. Colors, fonts, layout, collapsible sections, embedded charts, simple interactivity. A markdown file is a transcript. An HTML file is a finished deliverable.

Andrej Karpathy has a hypothesis about how humans actually want to interface with language models. Audio in, visual out. We speak faster than we type, and we process visual information faster than we read prose.

I buy it. Most of what I write starts as voice dictation. It's faster, it captures the way I actually think, and the words come out closer to how I'd say them in conversation. On the output side, I'd rather scan a well-designed page than read three paragraphs to find the same point.

Markdown handles the audio-in side fine. It's just text. On the visual-out side, markdown taps out quickly. Headings and bullets get you maybe ten percent of what a human actually wants to look at. HTML gets you the rest of the way.

Here's how I've started thinking about it. HTML is a longer lever than markdown.

Picture a physical lever. The longer the lever, the more force you can apply with the same effort. HTML produces more output per unit of input. You describe what you want, the model generates an HTML file, you open it in a browser, and you have something that looks like a real deliverable. A styled brief, a comparison table, an interactive prototype, a dashboard mockup. The markdown version of the same exercise gives you a wall of bullets.

The cost of the longer lever is tokens. HTML files burn more tokens than markdown files because there's more structure to generate. Free users of Claude or ChatGPT will hit usage limits faster. Acknowledge the tradeoff and move on. At this stage of adoption, the more capable workflow is the right call, and costs come down.

Think about the deliverables you produce in a typical week. A status update. A strategy memo. A meeting prep doc. A plan you're trying to get approved. Most of them get drafted in Google Docs or Word, sent as attachments or links, and disappear into someone's inbox.

Pick one and try it as HTML instead. Open Claude or ChatGPT, paste in the content or describe what you need, and ask for a single-page HTML file. Specify what you want it to look like. Clean typography, a header that frames the document, two or three collapsible sections so the reader can scan first and dig in where it matters, a styled callout for the recommendation. Download the file, open it in your browser, and see the difference.

If it lands the way I think it will, send it to whoever you'd normally send the Google Doc to. Pay attention to their reaction. Your non-technical boss won't know how you made it. They'll just notice it's better than what everyone else sends them.

Cloudflare cuts 20% of its workforce. Matthew Prince announced Cloudflare is laying off roughly a fifth of its employees, citing AI-driven changes to how the company operates. This is the third major tech layoff in recent weeks where AI was cited as a factor. Block, then Coinbase, now Cloudflare. The pattern is no longer speculative. Real companies are restructuring around the assumption that AI changes the shape of the workforce. If you're in a knowledge-work role and you're not actively building AI fluency, the math is getting harder.

Anthropic's secondary market wrinkle. Inside baseball, but a look for anyone tracking private markets. A recent X post flagged language in Anthropic's stock policy that constrains how shares can move through SPVs and similar structures. For anyone who's been eyeing Anthropic secondaries as a route to exposure, the mechanics are more restrictive than they appear on the surface. Buyer beware.