Every era of computing has been defined by its interface, not its hardware.
The mainframe era was defined by the punch card. The PC era by the keyboard and the desktop metaphor. The internet era by the browser. The mobile era by touch. In each transition the machines got faster, but that is not what changed the world. What changed the world was that more humans could tell computers what they wanted.
Each interface shift has the same shape: it lowers the translation cost between human intent and machine action. Punch cards required specialists. Command lines required memorized syntax. GUIs required learned metaphors: folders, windows, trash cans. Touch required almost nothing. The curve bends toward zero.
Conversation is the end of that curve. Not because chat windows are pleasant, but because conversation is the interface humans already ship with. Nobody teaches a four-year-old to converse with a manual. It is our native protocol, with ambiguity resolution, context carrying, and repair when misunderstood all built in.
The operating system claim
Calling conversation an operating system sounds like a metaphor. I mean it more literally than that.
An operating system does three things: it schedules resources, it abstracts hardware, and it gives programs a common environment to run in. Look at what happens when language models sit at the center of a system:
- Scheduling becomes intent routing. The model decides which tool, which agent, which data source serves the request.
- Abstraction becomes semantics. You no longer address an API by its schema, you address it by what you mean. The schema is somebody else's problem, resolved at runtime.
- The common environment is context: conversation history, retrieved memory, user state. Programs that run "inside" the conversation inherit it, the way processes inherit the file system.
When I build agentic systems, this is what they actually look like. The LLM is not a feature bolted onto an app. The app has dissolved into tools that a conversation orchestrates.
What I keep getting wrong
My instinct, trained by twenty years of using computers, is to design screens first and add language on top. It is exactly backwards. The systems that work start from the dialogue (what does the user say, what should happen) and grow screens only where language is the wrong tool. Dense tables, spatial layouts, anything you scan rather than ask: screens win there, and will keep winning.
Conversation is not the interface for everything. It is the interface for intent. The distinction matters. "Show me Q3 revenue by region" is intent, so it is conversational. Reading the resulting chart is perception, so it is visual. The next operating system will speak both, but it will be organized around the first.
Why this matters to me
I watched an elderly woman use a speechbot I built, someone who had never adapted to a smartphone and never would. There was no learning curve, because there was nothing to learn. She just talked.
That is the entire thesis. For seventy years, humans have adapted to computers. The adaptation is finally reversing direction. The people building right now are not building apps. They are building the first draft of how humans and machines will talk for the next fifty years.
It is worth getting the draft right.