← Themes

Why AI Needs a Native Web Protocol

Or: how we keep asking AI to do something it was never designed to do.

Three waves of media

Look at how information has migrated through the last fifty years.

The first wave was digitization — paper became files. Books, photos, music, contracts: all of them got new addresses inside computers. This wave's foundation was containers — TXT, JPG, MP3, PDF — formats that let machines store what humans once filed in cabinets.

The second wave was vectorization — text became embeddings. Once we figured out how to represent meaning in high-dimensional space, AI started understanding what we wrote, not just storing it. This is the wave that gave us LLMs, semantic search, retrieval-augmented generation.

The third wave is happening right now: APIfication — interfaces become programmatic. Every UI that used to require human hands and eyes is being rebuilt so AI can operate it directly. This is the wave that's just beginning, and it's where the entire next decade of AI products will play.

The web is not yet on the third wave. That's the problem this essay is about.

HTML is for humans

Look at HTML for a moment. The angle brackets, the nested div soup, the inline styles, the aria-* attributes for screen readers, the <noscript> fallbacks, the carefully ordered <head> tags for browser compatibility. HTML is one of the most beautifully human-shaped formats ever designed: it's edited by humans, read by humans, and structured around the way human eyes scan a page.

When we ask AI to "build a website," we're asking it to operate on this human-shaped format. And it works! GPT-4, Claude, Gemini — they can all generate functional HTML. But every interaction is brutally inefficient. Here's what happens when you ask an AI to "make my hero section more energetic":

  1. The AI loads your entire HTML file into its context window — often 3,000 to 5,000 tokens for a single landing page.
  2. It re-generates the whole file, character by character, just to change one section.
  3. You paste the new HTML back, hoping nothing else broke. (Spoiler: things usually break.)
  4. You repeat this loop 5–10 times for every meaningful change.

A typical AI-driven website-building session burns 30,000–50,000 tokens. Most of that is the AI re-generating context that didn't need to change. It's like asking someone to copy a 200-page book just to fix a typo on page 47.

This isn't a bug in any specific model. It's the format. HTML was never meant to be edited by something that thinks in JSON.

What does an AI-native web protocol look like?

Here's the design question I started obsessing over earlier this year: if we were starting from scratch — if HTML didn't exist, if browsers didn't exist, if the constraint of backward-compatibility was lifted — what format would we design for AI?

The answer, I believe, has four properties:

1. Structured, not stringy. JSON, not character soup. AI models speak JSON natively. Every piece of content has a name; every layout decision has a property; nothing is hidden in style attributes or class names that need to be inferred.

2. Composable, not monolithic. A page is a tree of sections. Each section is a tree of components. Each component is a bag of properties. To change one button's label, you patch one property — not regenerate a paragraph.

3. Operable through API, not file-edit. Want to add a "testimonials" section? POST to /api/page/:id/section. Want to change the headline? PATCH /api/page/:id/section/:sid. The AI never touches a file. The AI never re-renders 3,000 tokens of context. The AI sends a 50-token diff.

4. Renderer-separated. The protocol describes what. The renderer decides how. This is the same separation TeX made between content and typesetting fifty years ago. It means a single page description can render in any visual style — a "theme" is just a different renderer plugin.

When you put these four properties together, you get something that doesn't look like a CMS, doesn't look like a no-code builder, doesn't look like Webflow. It looks like a protocol — the same way HTTP is a protocol, the same way JSON-RPC is a protocol. A spec that any AI client can speak.

This is what Mold is.

Mold as a reference implementation

I built Mold in four weeks earlier this year. The goal wasn't to ship "another website builder" — there are already a thousand of those. The goal was to establish the protocol and then provide one open implementation so anyone could see what it looks like in practice.

Here's what came out:

The five biggest open AI models — Grok Flash, Gemini Flash, ChatGPT, Claude, GPT-5 (via M365 Copilot) — all built complete, working sites with Mold on the first try. Average token cost dropped from ~3,500 (HTML approach) to ~100 (Mold approach). A 97% reduction.

That number isn't magic. It's just what happens when you stop asking AI to redo work it already did.

What this enables

The most exciting downstream effect is not that website-building gets cheaper. It's that website-building gets accessible to people who shouldn't need a developer.

If you want a portfolio site, you should not need to know what a <div> is. You should describe what you want. An AI that speaks an AI-native protocol should produce it. You should be able to revise by speaking, not by editing files. And the cost should be low enough that a free-tier AI can do it.

This is the same shift that happened when desktop publishing replaced typesetters in the 1980s, when WYSIWYG editors replaced HTML hand-coding in the 1990s, when blogging platforms replaced static HTML in the 2000s, when no-code builders replaced blogging platforms in the 2010s. Each shift removed a layer of expertise from the bottleneck. AI-native protocols remove the next layer.

The interesting question is not whether this happens — it's already happening. The interesting question is what protocol wins.

A note on what comes next

Mold is open. The protocol spec, the engine, the themes — all on moldpage.dev. The deeper philosophical project is laid out in Tokenomic: REPRICED by AI (Amazon, Feb 2026), the book I wrote on what happens when intelligence becomes a commodity.

If you're an AI company, an AI developer, or just someone who's tired of pasting HTML into ChatGPT — try it. Tell me what breaks.

The third wave is here. The web hasn't caught up yet. We can fix that.

— Gerald Yang Toronto, May 2026