MetaBox

Inference should be distributed

The web won because nobody owned it. Personal AI will go the same way.

1. Inference belongs where the person is

Today's AI is a cloud service you rent, the way the web was once a part of AOL. The magic of the web was distributed: anyone could run a server, anyone could write a page, and the protocol was public. A personal agent that runs your life should run in your house, on your hardware, with the cloud as an on-ramp rather than a landlord. Muse's own Secure VM is already the right shape. It is just in the wrong building.

2. You own the data and the context

Your memory, your history, your preferences, your files, the graph of how you actually live: this is the most valuable asset in personal AI, and today every vendor keeps it in their own account. Muse lets you read and edit a memory file. Claude lets you paste it out as text. Nobody exports it in a format another agent can read. The user should hold the keys, and the format should be open. Meta's Confidential VM is the promise. A home box is the delivery.

3. Local models handle most of the day

The gap between the best open model and the best closed model is about four months, and a 30 billion parameter model in 2026 beats last year's 70 billion. A Glimmer-class model on a box at home handles the drafts, the summaries, the reminders, the classification, the background work at zero marginal cost and without a byte leaving the house.

4. Route to the frontier only when the task needs it

When the job is hard, the box calls Muse Spark in Meta's cloud, pays for it, and comes back. Routing is already a product category: OpenRouter serves 25 trillion tokens a week across 400 models, GPT-5 ships with a router, and Apple pays Google about a billion dollars a year to escalate Siri. Meta gets a metered revenue line on top of the subscription, and the user gets frontier intelligence exactly when it matters.

5. Apps connect to your agent, not the other way around

Today an app calls a model vendor's API and marks up the tokens. The right direction is the reverse: an app exposes what it can do, and you connect your agent to it with a button, the way you sign in with Google. The protocols exist. MCP is the wire, OAuth 2.1 is the consent screen, A2A is how two agents talk, and Skills are the packaging. Developers sell apps and skills. They stop reselling tokens.

6. One memory, shared across your life

The trip planner should know about the work deadline. The meal planner should know about the doctor's advice. This only works if there is one agent with one context, and apps plug into it with scoped permissions, rather than a dozen apps each holding a sliver. That is a platform decision, and only the platform owner can make it.

7. Proactive, ambient, in the home

An agent that works while you sleep needs somewhere to run while you sleep. A phone goes in a pocket. A cloud VM costs Meta money every hour it thinks. A box on the shelf draws 60 to 100 watts, is always on, and is the natural hub for the TV, the glasses, the speaker, and every phone in the family. Two thirds of Americans say they would pay more for a device that protects their privacy. This is that device.

Meta sells attention and devices, not tokens. A free model everyone builds on expands the market and blocks a rival walled garden.The argument Meta made for open weights in July 2024. It is still right.