Field Notes from a Paradigm Shift

My laptop dragged this afternoon, and my first suspicion was that I'd done it to myself. For a while now I've had the computers around my house pitching in on background AI work, and the obvious suspect was the laptop I was typing on - quietly working against me. I checked. It had already stepped aside.

Here's the setup. Three ordinary machines on the same home Wi-Fi - a little always-on Mac mini, my work laptop, and a small Windows PC - each able to run AI models locally, on the machine itself, with no cloud bill ticking. In front of them sits a shared helper: any job that needs a bit of AI hands its work over, and the helper picks whichever machine is free and sends it there. It isn't fussy about the task. The one I keep it busiest with is big and boring - reading a firehose of news articles and pulling out the names, places and companies inside them - but that's just one of many. On their own these machines would sit idle most of the day; a laptop spends its life either asleep or waiting for you to type. I built a way to borrow that dead time for whatever needs doing.

The interesting part isn't the sharing. It's the manners. The laptop only joins in when three things are true at once: it's plugged into the wall, you haven't touched the keyboard for three minutes, and it isn't already busy with your own work. Miss any one and it politely declines, and the job goes to the always-on mini instead. Touch a key and it drops the background work instantly. Your own apps never wait behind it. On battery it doesn't lift a finger, and it's left free to fall asleep like a laptop should.

The flip side of a polite laptop is that it vanishes without warning - you sat down, so it's gone. So the thing that hands out the work expects that. It sends each job to whoever's least busy, and if a machine goes quiet mid-task it passes the work to another one within seconds and benches the sleeper for a short rest. One machine going down never stops the rest. From the outside it looks like a single reliable worker. Underneath, it's three unreliable ones, coming and going as the household does.

The future of cheap AI might not be a bigger building. It might be the hardware already sitting in your living room, doing nothing.

Which is the part I can't stop turning over. The whole cultural picture of AI is a data centre the size of a town, drinking a river and a power station. And here I was watching real AI work get done, for no extra money, on three machines I'd already bought for other reasons. The clever bit wasn't scale. It was timing - being smart about when, not how big.

None of this is new in spirit. Folding@home has borrowed idle home computers to fold proteins for science for over twenty years. But look at what's idle now: the graphics card in a gaming PC, the single most capable AI chip most households will ever own, sitting dark twenty-three hours a day. There are millions of them. Pool that across strangers and you've reinvented Napster, except the thing being shared is raw compute - a vast, cheap AI cloud made of hardware that already exists, in bedrooms and living rooms, bought and paid for. Gamers could rent their cards out between matches and get paid for the silence. Cheap AI for the masses, without pouring a single new data centre.