Building AI, not just using it

Two things became clear while we were building our new Edge AI Lab for teens, and both of them are reasons the program looks the way it does.

The first is that the real learning happens in the trade-offs. The second is that responsible AI lands differently when it is your own project. Neither of those is the kind of thing you can lecture into a kid. You have to build the conditions where the kid runs into them on their own. So that's what we built.

The thing most programs skip

Almost every teenager we meet can already use AI. They can prompt a chatbot, generate an image, get homework help they probably shouldn't. Fluency with AI as a finished product is no longer rare, and it is no longer the valuable thing.

The valuable thing — the thing that is genuinely scarce and genuinely future-proof — is understanding how these systems actually work, well enough to build one. And there is a specific corner of the field where a motivated teenager can go from zero to building something real in a few weeks: edge AI, artificial intelligence that runs directly on a physical device instead of in a distant data center.

This is not a niche. The industry is moving this way fast. The AI a kid talks to on a phone lives in the cloud — a question travels over the internet, a giant server does the thinking, an answer comes back. That works beautifully for chat and breaks down completely for anything that has to move, see, or react in the real world. A robot can't wait for the internet to decide whether to stop. A camera in a hospital can't send private video to an outside company. A machine in a tunnel has no signal at all. So the intelligence is moving onto the machine itself.

The Edge AI Lab, part of our Robotics & Physical AI zone, teaches that. Small teams of teenagers set up a real NVIDIA Jetson Orin Nano — the same class of on-device AI computer used in industry, about the size of a paperback — and over six sessions they learn how it works, run real AI models on it, and build and demo a project of their own.

Where the learning actually happens

Here is the part we didn't fully anticipate. The deepest learning in the room does not happen during the explanation. It happens in the argument.

A Jetson has a fixed memory budget. You cannot run an enormous model on it; you have to choose. So a team sits there and has to decide which model fits, what they're willing to give up for speed, whether a smaller model that runs at thirty frames a second beats a smarter one that crawls. That is a real engineering trade-off, the same shape of decision a professional makes, and the kids work it out by arguing with each other.

Or a team's object detector keeps calling a coffee mug a bowl. Why? They have to reason about it — the lighting, the angle, the training data the model never saw, the fact that the model is guessing from patterns and sometimes guesses wrong. In ten minutes of being annoyed at a misclassified mug, a fourteen-year-old internalizes more about how machine learning actually behaves — its limits, its errors, its confidence that isn't the same as correctness — than an hour of slides could deliver.

We didn't design those moments as lessons. We designed a situation with real constraints and let the constraints do the teaching. The mentors in the room are there to keep hands on keyboards and to resist the urge to just fix it, because the fixing is the kid's to do.

Responsible AI you can hold

The second surprise was how naturally the ethics conversation arrives, and how much heavier it lands, when the AI is running on a device the student owns for the afternoon.

Talk to teenagers about AI ethics in the abstract and you get the expected, slightly bored answers. Put a working camera in their hands that runs entirely on their own machine, and the question changes character. "What should this camera record, and what should it never save?" stops being a debate-club prompt and becomes a design decision they have to make, with their own hands, before the thing works. Because everything runs locally — nothing is uploaded, nothing leaves the room — privacy is not a policy they're told about. It's a property of the system they're building, and they're the ones who decide it.

That is also why the program is built the way our whole approach to safety and age-appropriate AI is built: supervised throughout, data kept on the device, and a standing rule that we build things that help rather than things that surveil. When a team builds a people-counter or a security camera as their project, the responsible-use rule isn't an add-on at the end. It's part of the build, written down before the demo.

The judgment we want them to have

The line we keep coming back to: we want this generation to have more than fluency using AI. We want them to have judgment about building it.

Fluency is using the tool well. Judgment is knowing what the tool actually is — where it's strong, where it quietly fails, what it costs, what it should and shouldn't be pointed at. You don't get judgment from a model that lives in someone else's data center and arrives as a finished answer. You get it from a model you set up, ran, watched make mistakes, and shaped with your own constraints. You get it from building, not from using.

That's the whole bet of the Edge AI Lab. Every student finishes by demoing a working AI project they built and can explain — an object detector, a voice assistant that runs offline, a self-driving robot, a privacy-first camera. The demo is the point. So is everything they argued about to get there.

Come build

We're opening a small first cohort, for ages 13–16, and keeping it intentionally quiet while we finalize dates and seats. No prior coding experience is required; we teach the minimum needed, every team is supervised, and the responsible-AI framework runs from the first session.

If that sounds like your kid — the one who would rather build than watch — the best move is to add your name to the interest list. You'll be first to hear when enrollment opens.

Join the interest list →