Smart homes spent a decade being disappointing in a specific way: everything worked, and nothing worked together. You could turn a light on with your voice, which was marginally slower than a switch. The promised intelligence never arrived because the systems could only follow rules you wrote yourself.
That is the part machine learning has genuinely changed, and it is worth separating from the marketing.
From rules to inference
The old model was strictly conditional. If motion is detected after sunset, turn on the hall light. Every behaviour had to be anticipated and written down, and the result was brittle: the light came on for the cat, and stayed off when you arrived early.
The current generation infers instead. A camera distinguishes a person from an animal from a passing car. A thermostat learns the building’s thermal behaviour and starts heating before you ask. A robot vacuum recognises a cable and drives around it rather than eating it.
The difference is not that the devices are cleverer in some abstract way. It is that they can categorise the world well enough to stop needing exhaustive instructions.
Where it has actually helped
- Cameras. The old motion alert fired on every shadow and rainstorm, so people switched notifications off entirely. Person, vehicle, animal and package classification made alerts worth reading again. This is the biggest single improvement in the category.
- Heating. Learning thermostats model how quickly a building warms and cools, and schedule around that rather than around a timer.
- Vacuums. Obstacle recognition and room mapping turned a device that got stuck into one that finishes.
- Voice. Assistants now cope with natural phrasing and follow-up questions rather than requiring an exact command.
- Energy. Some systems identify individual appliances from the electrical signature at the meter, without a sensor on each device.
The processing moved back into the house
This is the most consequential change and it gets the least attention.
Early smart devices sent everything to a server because the analysis needed more computing power than a doorbell contained. Chips capable of running recognition models locally are now cheap enough to put inside the device.
The practical results are significant. Response is faster, because nothing waits on a round trip. Devices keep working when the internet is down. And video can be analysed without leaving the building, which is a genuine privacy improvement rather than a claimed one.
When comparing products, “on-device processing” or “local processing” is a specification worth prioritising, and it usually costs nothing extra.
What has not been fixed
Interoperability remains the central problem. The Matter standard was meant to solve it, and has improved basic control across ecosystems, but advanced features still tend to work only inside one manufacturer’s app.
Subscriptions have spread. Features that seem to be part of the hardware, particularly camera recognition and video history, frequently sit behind a monthly fee, and can be withdrawn from a device you already own.
Longevity is worse than for ordinary appliances. A device dependent on a company’s servers stops working when that company loses interest, and this has happened repeatedly to products that were selling well.
Buying sensibly
- Check what needs a subscription before buying, not after. Ask specifically which features stop without one.
- Prefer local processing and local control, so the device survives an outage and an acquisition.
- Check Matter support, but do not assume it covers everything the product does.
- Put smart devices on a guest network. Cheap connected hardware is a recurring source of security problems, and a separate network limits what a compromised device can reach.
- Solve an actual problem. The devices people keep using address a genuine annoyance. The ones in a drawer were bought because they were clever.
The useful test after a year is simple: would you notice if it stopped working? Heating that manages itself and a camera that only alerts when it matters pass. A voice-controlled kettle does not.
Join the discussion