AI Is Having Its
HTML Moment
Every decade or so, a new primitive rewires the whole stack. In 1993 it was the hyperlink. Now it is the prompt. And if history is any guide, we are only at the beginning of a very long wave.
In 1993, a physicist at CERN released a specification that almost no one read and almost everyone eventually used. HTML was not a programming language. It was not even particularly elegant. It was a way of saying: here is some content, here is how it relates to other content, and here is roughly how it should look. That was enough to change everything.
What made HTML powerful was not the technology itself. It was what the technology assumed. It assumed that any machine, anywhere, running any operating system, could parse a document and render it into something useful. It was a universal agreement about how to exchange meaning. Before HTML, software was compiled, siloed, and closed. After HTML, everything was composable, linkable, and readable by anyone with a browser.
We are in that same inflection point right now, but one layer up. The new primitive is not a markup language. It is natural language itself, mediated by a model that can parse intent and return action. Instead of a document saying “here is a button,” we now have a system that says “here is what you want, and here is what I did about it.” The interface layer is collapsing into the instruction layer.
The architectural parallels run deep. The prompt is the new HTML: a human-readable format for expressing structure and intent. The LLM is the new browser: a runtime that interprets that format and renders something useful. Tool use is the new JavaScript: dynamic behavior, triggered conditionally, acting on the world. Context windows are the new DOM. Memory and RAG are the new URLs.
The HTML-to-AI era map: six parallel layers from primitive to governance, with the paradigm waves forming a spider web of overlapping concepts.
HTML did not create one wave. It created six.
This is the part of the story people forget. HTML’s impact was not a single moment of disruption. It was a sequence of compounding paradigm shifts, each one building on the last, each one birthing industries that did not exist before. Understanding that sequence is the only honest way to extrapolate what AI will do next.
| Era | HTML did this | AI is doing this |
|---|---|---|
| Wave 1 1993–98 | Static sites. Brochureware. The world put its pamphlets online and called it a revolution. | Chatbots and copilots. Assistants bolted onto existing software. Early adopters, uncertain ROI. |
| Wave 2 1998–2004 | AJAX and dynamic content. Pages that could update without reloading. Gmail in 2004 was the watershed. | RAG, memory, and grounding. Models that reach out to live data and stay current. The static prompt giving way to the dynamic context window. |
| Wave 3 2004–10 | Rich landing pages and conversion science. Design became a growth lever. A/B testing, funnels, analytics. | AI-native landing pages. Products that demo themselves, answer objections in real time, and personalize the pitch per visitor. The funnel becomes a conversation. |
| Wave 4 2008–14 | Web apps and SaaS. Software moved to the browser. Salesforce, Basecamp, Dropbox. The CD-ROM died. | Agentic apps. Software that acts, not just displays. Book the flight, draft the contract, file the ticket. The GUI becomes optional. |
| Wave 5 2013–20 | MVC, React, SPAs. Componentized front-ends. JavaScript frameworks. The DOM as a state machine. | Multi-agent systems and MCP. Orchestrated models, tool use, structured hand-offs. The prompt as a composable unit. AI components you assemble like Lego. |
| Wave 6 2016–now | Mobile-first and PWAs. The screen got smaller and more personal. The web met the phone in your pocket. | Ambient and embedded AI. Intelligence disappearing into the device, the earpiece, the car, the home. The interface becomes invisible. The model is just there. |
We are sitting somewhere between wave one and wave two on the AI timeline. The chatbot-as-pamphlet era is already feeling tired. The dynamic, grounded, personalized wave is just beginning to break. Waves three through six exist only as demos and roadmaps. Anyone who tells you they know what wave six looks like is guessing, but the arc is real.
A new interface paradigm, not just a new tool
Here is what separates this moment from the introduction of, say, the smartphone, or the cloud. Those were delivery mechanism changes. HTML and AI are something different. They each introduced an entirely new way for humans to interact with computers.
HTML gave us the visual interface as the universal default. Before the web, interacting with software meant CLIs, proprietary GUIs, application-specific paradigms. After the web, the page became the contract. Click here, fill this in, follow this link. Every piece of software, for thirty years, has essentially been a page.
AI is introducing the conversation as the new default. Not a chatbox bolted onto a webpage. A fundamentally different contract between user and machine. You do not click to a destination. You describe a destination and the system gets you there. The interface becomes the request itself.
This is why the transition feels disorienting to designers and developers trained in the visual paradigm. The skills do not transfer cleanly. Knowing how to lay out a page beautifully does not tell you how to design a conversation gracefully. Knowing how to write CSS does not tell you how to write a system prompt. The craft is new, even if some instincts carry over.
The jobs that did not exist yet
Every HTML wave created professions that could not have been named in 1993. Webmaster, front-end developer, UX designer, SEO specialist, growth hacker, React engineer. These were not old jobs with new names. They were genuinely new cognitive work, requiring new mental models, new tools, new communities.
| HTML created | AI is creating |
|---|---|
| Webmaster The generalist who owned the site | AI product lead Owns the model’s behavior end to end |
| Front-end dev Turns design into rendered UI | Prompt engineer Turns intent into model behavior |
| UX designer Architects user flows through screens | Conversation designer Architects user flows through dialogue |
| SEO specialist Makes content findable by machines | LLM optimization Makes brands visible inside AI answers |
| Growth hacker Optimizes the funnel scientifically | Agent orchestrator Designs multi-step autonomous workflows |
| CMS manager Structures content for web delivery | Eval engineer Structures feedback loops for model reliability |
Most of these AI-era roles barely have agreed-upon names yet. Some of them are still being practiced by people who do not know they are pioneering a profession. That is exactly where front-end development was in 1997. The job existed. The title did not.
The W3C problem: nobody is minding the spec
Here is where the HTML analogy breaks down most visibly, and where the risk is greatest.
HTML had the World Wide Web Consortium. A body of consensus, slowness, and genuine interoperability standards that made the web buildable. You could write HTML once and trust that it would render recognizably across machines, operating systems, and decades. The spec was public. The process was open. The browser wars were brutal, but there was always a standard to converge toward.
AI has no W3C. It has no IETF. It has no agreed-upon spec for what a model should do, how it should behave, what “safe” or “aligned” or “capable” means in a way that is testable, portable, and vendor-neutral. It has government hearings, voluntary commitments, and a proliferation of competing evals that measure different things in incompatible ways.
The consequences are already visible. A prompt that works reliably on one model breaks silently on another. An agent built for one API version stops working when the model is updated without notice. Behavior that is safe on one deployment is unsafe on the next. There is no “View Source” for alignment. You cannot audit a model the way you can audit a webpage.
This is not a small problem. The web’s interoperability was not a gift. It was hard-won through years of standardization work that felt boring and bureaucratic precisely because it was doing something important. The AI industry is currently racing to ship while that work goes undone. At some point, the absence of standards will become the story, not the features built on top of the absence.
The most durable opportunity in AI is not building the next application. It is building the infrastructure that makes the next thousand applications trustworthy. Evaluation frameworks, behavioral contracts, interoperability layers, audit tools. The picks and shovels of an intelligence gold rush, and the standards body that does not yet exist.
We are, unmistakably, at the beginning of something. The primitive has arrived. The waves are forming. The jobs are being invented. The interface paradigm is shifting beneath our feet. What we do not yet have is the governance layer that will make this sustainable. Whoever builds that will matter as much as whoever built the browser.