I Don’t Want an AI House. I Want a House That Works.

I have Home Assistant running enough things in my house that occasionally I have to remind myself the objective was convenience.

This is one of the occupational hazards of being an engineer with access to Raspberry Pis. You begin with “it would be nice if this light turned on automatically” and eventually find yourself debugging why a Bluetooth sensor, an automation, a weather integration and a smart plug have formed a tiny distributed system whose only apparent objective is preventing you from making coffee.

I love home automation. I also think we are rapidly confusing an automated house with an AI house.

I don’t particularly want an AI house.

I want a house that works.

The light switch is already a pretty good interface

There are places where intelligence genuinely helps. Natural-language control is useful. Pattern recognition can identify unusual energy use. A model can summarize what changed while I was away. Weather With Regerts can tell me the desert has once again chosen violence in language considerably more entertaining than “106°F.”

But if I walk into a dark room and the occupancy sensor sees me, I do not need a language model to deliberate about whether illumination aligns with my broader life goals.

Turn on the damn light.

The best home automations are often boring deterministic rules. If this sensor changes, do that. If the door has been open for five minutes, notify me. If the temperature exceeds a threshold, change the HVAC state. Those rules are understandable, testable and—most importantly—predictable.

Local first isn’t nostalgia

One reason I like Home Assistant is its local-first philosophy. My house should not stop functioning because somebody else’s cloud service has an outage, changes its API, decides my five-year-old device is obsolete, or gets acquired by a company whose strategic vision consists mostly of raising the subscription price.

Cloud services are useful. I use plenty of them. Remote access is useful. Voice services are useful. Integrations with outside data are useful.

They should be enhancements rather than structural dependencies for basic household functions whenever practical.

If the Internet goes down, my light switches should continue their distinguished career as light switches.

AI belongs where ambiguity lives

This is the architectural distinction I keep finding useful. Deterministic systems should handle things that have deterministic answers. AI becomes interesting where the input or desired output is ambiguous.

“Temperature is 106°F” is a fact supplied by a sensor or weather provider. “Outside is currently a hostile environment apparently designed by Satan’s facilities-management department” is presentation.

That is why Weather With Regerts is, accidentally, a decent AI architecture lesson. The joke does not decide the weather. The underlying system determines the facts and the personality layer makes those facts more fun to consume.

If the personality layer fails, I should still know the temperature.

If an AI summary service fails, my thermostat should still operate.

If voice recognition misunderstands me, there should still be a button.

Graceful degradation is a feature

Enterprise engineers have names for this stuff: fault tolerance, graceful degradation, separation of concerns. Somehow consumer technology regularly forgets all three because putting “AI-powered” on the box is easier to market than “continues functioning when Comcast is having a bad day.”

A smart home should become progressively less fancy when components fail, not progressively less habitable.

That means local control where it makes sense, manual fallbacks, automations that do not depend on unnecessary services, and enough observability that when something does break I can determine why without conducting a séance.

I already have a separate machine for séances.

Make the intelligence optional

I am excited about what AI can add to home automation. Better natural-language interfaces alone can remove a tremendous amount of configuration friction. Context-aware notifications can make alerts less noisy. Models can find patterns I would never think to encode manually.

But intelligence should sit on top of a reliable house rather than underneath it.

The future I want is not a home that constantly demonstrates how clever it is. I want one that quietly handles the boring things, tells me when something genuinely needs my attention, lets me experiment with ridiculous integrations when I feel like it, and fails in ways that leave the lights working.

My house does not need consciousness.

It needs uptime.