Distill, Baby, Distill

Remember when an open-weight model from a Chinese lab suddenly had the entire AI world refreshing its group chats, checking benchmarks, and asking “wait, they built that for how little?” That moment did something important: it reminded American researchers that being open — publishing your model weights so anyone can download and build on them — isn’t a consolation prize. It’s a strategy. And now Garry Tan, the head of Y Combinator, is pushing US labs to lean into that strategy harder.

Tan is calling on smaller American open-weight AI labs to “distill” frontier models, with the goal of building a strong alternative to Chinese AI dominance. He frames it as crucial for advancing American AI research by 2026. If you’re not sure what “distill” means in this context, don’t worry — that’s exactly what I’m here for.

Distillation, explained without the jargon

Think of a frontier model — the biggest, most capable AI systems out there — as a master chef. Brilliant, but expensive to hire and impossible to fit in your home kitchen. Distillation is like having that chef train an apprentice. The apprentice watches the master work, learns the patterns and shortcuts, and eventually cooks nearly as well — while working out of a food truck instead of a five-star restaurant.

In AI terms, a smaller model learns from a larger model’s outputs. The student model absorbs much of the teacher’s capability but ends up cheaper to run, faster to respond, and small enough to operate on more modest hardware. You lose a little polish. You gain a lot of practicality.

For those of us who care about AI agents — software that actually does tasks for you rather than just chatting — this matters enormously. Agents make lots of calls to a model, over and over, all day long. A distilled model that’s 90% as smart but a fraction of the cost is often the better engine for an agent. Nobody wants an assistant that’s brilliant but bills like a corporate law firm.

Why Tan is making this a national argument

The interesting part of Tan’s pitch isn’t the technique — distillation is well understood. It’s who he’s asking to do it and why. He’s not talking to the giants. He’s talking to smaller American labs, the scrappy teams that don’t have billion-dollar training budgets but do have talent and speed.

His logic, as I read it: Chinese labs have been releasing capable open-weight models at a steady clip, and those models are becoming the default building blocks for developers worldwide. When a developer in Brazil or India or, frankly, Ohio picks a base model for their startup, they pick whatever is good, free, and available. If the best open options increasingly come from one country, the global developer ecosystem quietly reorganizes around that country’s technology.

Distillation is a way for smaller US labs to compete without matching frontier-scale spending. Instead of training a giant from scratch, they can produce smaller, excellent, openly available models — and give developers an American-made alternative to reach for.

What this means for regular people

You might reasonably ask: I’m not a lab, I’m not a developer, why should I care? A few reasons:

• Cheaper AI everywhere. Distilled models drive down the cost of running AI features, which trickles into the apps and agents you actually use.
• AI on your own devices. Smaller models can run locally — on laptops, maybe phones — which is better for privacy and works without an internet connection.
• More competition, fewer gatekeepers. A healthy open-weight ecosystem means the future of AI isn’t decided entirely by a handful of closed labs, wherever they’re located.

My honest take

I like this pitch because it’s realistic. It doesn’t ask small labs to out-spend giants; it asks them to out-hustle. Distillation is the underdog’s tool, and open weights are the underdog’s distribution channel. Put them together and you get exactly the kind of competition that keeps this field interesting.

There are open questions, of course. Distilling from someone else’s frontier model raises thorny issues about terms of service and whose intelligence you’re actually bottling. And “by 2026” is an ambitious timeline for reshaping research momentum. But as a direction? Tan is pointing at something real. The open-weight race is on, and he’d rather American labs run in it than watch from the stands.

If you’ve been following AI agents on this site, keep an eye on this one. The models that end up powering your future digital assistants may not be the giants at all — they’ll be the well-trained apprentices.

🕒 Published: September 13, 2026