There's a corner of my desk now that would confuse anyone who knows me as a designer: a microcontroller the size of a stick of gum, a tangle of jumper wires, a cheap mic, and a soldering iron I am still not entirely qualified to own.
People assume this is a sudden pivot. It's the opposite. It's the oldest thread in my life finally getting picked back up, because the thing that was blocking it quietly disappeared.
01 It started with a screwdriver
Before I had any language for design, I took things apart. Remotes, radios, a walkman that never worked again, whatever was nearby and unsupervised. I wasn't trying to fix anything. I wanted to see the inside, because the case only ever tells you what someone decided you should think about the thing.
Then I grew up, and the industry routed all of that curiosity onto screens. Screens are where the jobs were. And screens are forgiving in a way that shaped a whole generation of designers: every mistake is an undo away, every experiment is free, nothing you do has weight. Nothing about hardware is forgiving like that. Solder the wrong pin and the smell tells you immediately. I missed that, for years, without knowing what I was missing.
02 A decade behind the glass
Here's the part that makes the pull make sense. For most of my career, at Meta especially, I've designed software for people who touch hardware for a living: data center operators walking the halls, swapping drives, racking machines, keeping the physical substrate of the internet alive. My job was the console in their hand while they did it.
I got good at that seam, the exact pixel where software ends and a human reaches into a machine. And if I'm honest, I spent a decade being quietly jealous of the other side of it. They got the atoms. I got the glass. I'd design the diagnostic flow, and someone else got to stand there and hear the fans spin down when it worked. The best I ever shipped was a toast notification telling them it had.
You can't spend ten years designing for people who fix machines without eventually wanting to build one.
03 The excuse collapsed
So why now? Because the two things that always stopped me stopped being true in the same year.
The first was skill. Hardware used to demand a stack I didn't have: firmware in C, board layout, driver documentation written by and for people who already knew the answer. That wall is mostly gone. The same AI-assisted workflow I use to ship production interfaces sits patiently next to me and debugs firmware at eleven at night. It has read every datasheet ever scanned. I describe what I want the way I'd brief an engineer, and we get there together. A designer with AI tooling is no longer locked out of the stack below the screen, any of it.
The second was that intelligence got small and cheap enough to put inside an object. A few dollars of silicon and a Wi-Fi call to a model, and suddenly a dumb box on my desk can listen and answer. The first time one of my builds did that, out loud, in a physical thing I had assembled with my own hands, I laughed like a kid. It's the taking-apart instinct running in reverse. I spent my childhood opening gadgets to find out how they think, and it never occurred to me that one day I'd be able to close one up with actual thinking inside.
# the whole "product", more or less.
# everything hard about it is design, not code.
while True:
heard = mic.listen()
reply = model.ask(heard) # intelligence, rented by the token
speaker.say(reply) # the design problem starts here
Six lines. The parts cost less than lunch. That's the collapse: the distance between "I wonder if" and "it's on my desk and it works" went from a career change to a weekend.
04 The interface is leaving the screen
That's the fun half of the story. Here's the half I'd actually defend in a meeting.
Almost every AI product right now is a chat box, and I've written before about designing those surfaces well. But a chat box is AI wearing software's old clothes. The genuinely unsolved interface problems show up when the model gets a body, even a tiny one:
- What does trustworthy look like on a device with no screen to explain itself?
- How does a physical object admit uncertainty, without a confidence bar to hide behind?
- When is presence a feature, and when is a listening object in your home just creepy?
- What's the hardware equivalent of undo, once an action touches the physical world?
These are exactly the trust questions I've been working on for years, transposed into a harder key. Nobody has the pattern library yet. The people deciding what AI hardware feels like to live with are mostly firmware engineers and supply-chain people, and I mean no disrespect when I say that a designer belongs in that room. The way to earn a seat in it is the same as it was for software: stop hanging back with the spec and build the thing yourself.
So that's why there's a soldering iron on my desk. Same instinct as the kid with the screwdriver, same job as the designer at the console, one layer closer to the machine. The prototypes are ugly and I love them. When one of them stops being ugly, you'll read about it here.