MetaBox

Chat is not the interface

Muse talks to you like a person in a message thread. That is a great way to start a task. It is a bad way to live inside one. The interface of the future is generated on demand, and Meta should own the means of generating it.

The problem with a text box

A text box has no affordances. It puts the whole burden of articulation on the person, and it makes every adjustment a re-prompt. Jakob Nielsen called intent-based AI the first new interface paradigm in sixty years and then warned that half the population is not articulate enough to get good results from prose alone. A chat thread is a command line with better manners.

Muse already knows this. Its Ideas tab, its Feed, its Goals view, and the Artifacts it builds inside the thread are all moves away from chat. They are the right instinct expressed as tabs. The finished version is an agent that renders the interface each task deserves, at the moment you need it, on whatever screen you are looking at.

What generative UI looks like

You ask about a trip and get a live itinerary you can drag. You ask about the budget and get a dashboard you can edit. You ask what the kids have this week and the TV shows a board the whole family can see. The model does not write a paragraph about your data. It builds the tool for it, from a design system it knows, and the tool stays as long as you want it.

This is shipping elsewhere. Gemini generates a custom interface per prompt and Google's own evaluation ranked it just below human-designed pages and far above text. A2UI, open sourced by Google, lets an agent describe an interface as JSON that native apps render. MCP Apps, co-authored by Anthropic and OpenAI, lets any tool return a live component inside Claude, ChatGPT, Copilot and Cursor. Anthropic reports over half a billion artifacts built. Android will generate home screen widgets from a sentence this year.

Three tiers, three audiences

Generated on demand. For everyone. The agent builds the interface for this task, right now, from Meta's design system, on the phone, the box's admin surface, the TV, or the glasses. Disposable when the task ends, saved when you like it.

Templates. For people who want a starting point. A meal planner, a training log, a homework tracker, a shared shopping list. Adapted by the agent to your household, malleable at the point of use rather than by an engineering team far away.

Built by professionals. For the long tail. Designers and developers publish interfaces and full apps to the marketplace, priced and reviewed, that plug into your agent and your context. This is the tier where Meta earns a margin, and it is the subject of the next page.

The demand is already measured

Lovable reports 500 million dollars in annual revenue with 80 percent of its builders non-technical and about a million new projects a week. Replit went from under three million dollars in 2024 to a run rate above 500 million. People are paying real money to build small personal software today, on cloud tools, with their data in someone else's database. Give them the same power on a box at home, with local models that cost nothing per token and an agent that already knows them, and the addressable audience is every family with a Meta account.

Industrial software can only target the biggest problems for the most people. The long tail of small, personal needs has never been served.The argument from Ink & Switch and Maggie Appleton, compressed.