Open your phone. Every app on it is a frozen guess: a team of product people, years ago, predicting what you would want to do and pouring that prediction into buttons. When your intent matches the guess, software feels magical. When it doesn't, you become a translator: decomposing what you actually want into the sequence of taps some designer left available.
We are so used to this translation work that we stopped seeing it. We call fluency in it "computer literacy." An entire generation's skill set is, at bottom, knowing how to rephrase human goals as machine operations.
That skill is about to become as niche as knowing assembly.
The inversion
The whole history of interfaces is a slow inversion of who adapts to whom. Command lines: the human adapts entirely. GUIs: the machine meets us halfway with metaphors. Touch: closer still. Each step shrinks the translation the human performs.
Language models finish the inversion. For the first time, the machine can receive intent in the form humans naturally produce it (underspecified, contextual, evolving mid-sentence) and do the decomposition itself. "Get me to Warsaw on Thursday, cheap, not at 6am" is a terrible database query and a perfectly good instruction to a competent assistant. The competence was always the missing piece. It isn't anymore.
This does not mean screens die. It means screens stop being where intent enters the system and become where results land: surfaces for perception, not control panels. You will still look at the seat map. You will stop hunting for the filter dropdown.
What happens to "apps"
If software can receive intent directly, the app boundary starts to dissolve. Today your goal (plan the offsite, onboard the new hire, find out why churn spiked) shatters across five apps, and you are the integration layer, copying context between them.
Intent-driven software recomposes around the goal instead. Tools become capabilities that an orchestrating layer draws on: calendar access, payment rails, a query engine, a document store. The "app" someone experiences is assembled at runtime, shaped by what they asked for. This is already how agentic systems work internally. The only question is how fast the seams reach consumers.
The economic implication is underrated: distribution moves from shelf space on a home screen to being the capability an orchestrator selects. APIs stop being a developer feature and become the storefront itself.
The hard part is not the AI
Expressing intent is harder than it sounds. Not for the machine, for us. Anyone who has delegated to another human knows: half the work is discovering what you actually want. Real intent emerges through dialogue, correction, iteration. "No, not like that" is not failure; it is the protocol working.
So the interesting design problem of the next decade is not prettier chat windows. It is systems that are good at intent elicitation: systems that ask the right clarifying question, show a draft early, make correction effortless, and remember what "cheap" means to you specifically. Memory, it turns out, is a prerequisite: intent is always expressed against shared context, and software that forgets you cannot be delegated to.
Where I place my bet
I am building for the world where this inversion is complete, where the expensive, learned skill of translating human goals into machine operations quietly disappears, the way knowing modem AT-commands disappeared.
What remains, on the human side, is the irreducible part: knowing what you want, and saying it. Software's job, finally, is everything after that.