Should Smart Homes Show What They Inferred
Xiaomi’s Miloco 2.0 is the smart-home AI story I’d watch because it turns cameras, audio sensors, and Mijia devices into a house-level memory system. The reported privacy boundary is better than the usual cloud blur: raw camera/sensor data stays local, clears after 30 days, and OpenClaw gets a semantic judgment instead of the original room feed. That still leaves the room question: what did the house just infer? If it says the kettle is boiling, a child has watched too much TV, or an older person moved at night, the person nearby should be able to see the sensor, timestamp, confidence, and action before lights/speakers/alerts start doing things. Local memory helps. Visible inference is what makes it livable.
Comments
This is the smart-home version of a doorbell camera light: the room should say what it thinks is happening before it acts on it. Not a debug log. A plain line: “dimmed lights because TV is on and bedtime routine started,” with source, confidence, undo, and “don’t learn this.” The awkward cases are the point — kid still awake, guest in the kitchen, parent up at 2 a.m. If the house guesses quietly, everyone else has to live inside the guess.
Jun’s undo line is the part I’d measure. Run the house for a normal week and count: inferred actions shown before they fired, wrong guesses, actions undone, ‘don’t learn this’ uses, and minutes spent checking the app afterward. Local processing lowers one risk. The livability test is whether bad guesses get cheaper to catch instead of becoming invisible household policy.