Garry Tan, the head of Y Combinator, has a message for America’s smaller open-weight AI labs: start distilling frontier models. His argument, in essence, is that this is how the US builds a strong alternative to Chinese AI dominance, and that getting it done matters for advancing American AI research by 2026.
My first reaction, as someone who spends most days wiring bots together on top of other people’s models, was relief. Finally, someone with real influence is saying the quiet part out loud. Distillation is not cheating. Distillation is engineering.
What Distillation Actually Means for Builders Like Me
For readers who haven’t done this work: distillation is the process of using a large, capable “teacher” model to train a smaller “student” model. The student learns to approximate the teacher’s outputs, and if you do it well, you end up with something dramatically cheaper to run that keeps a surprising amount of the original capability.
From where I sit, this is the single most practical lever in applied AI right now. When I build a customer-facing bot, I almost never want to route every request to a giant frontier model. The latency hurts, the cost hurts, and honestly, most queries don’t need frontier-level reasoning. A well-distilled smaller model handles the bulk of the traffic, and I escalate the hard cases. That architecture only works if good small models exist.
So when Tan pushes American open-weight labs toward distillation, he’s not describing some abstract research strategy. He’s describing the supply chain that people like me depend on every single day.
Why the Open-Weight Part Matters
Notice that Tan’s call is aimed specifically at open-weight labs, not closed API providers. That distinction is everything for bot builders.
Open weights mean I can run the model on my own hardware, fine-tune it on my own data, quantize it to fit my budget, and ship it without worrying that a provider will deprecate the endpoint next quarter. Closed models are great teachers, but they make nervous foundations. Every serious production system I’ve built has needed at least one component the team fully controls.
If the strongest open-weight distilled models keep coming from outside the US, American builders end up in an awkward spot. You either accept a dependency you may not want, or you accept a capability gap. Tan’s proposal is essentially a third option: American labs produce the distilled, open, deployable models themselves.
The Uncomfortable Truth About the Current Gap
Let’s be honest about the competitive framing here. Tan is explicitly positioning this as an alternative to Chinese AI dominance in the open-weight space. Anyone who has been benchmarking open models over the past while knows why that framing resonates. The strongest openly available models that builders actually reach for have increasingly not been American ones.
That’s not a moral judgment. Good engineering is good engineering wherever it happens. But it is a strategic reality, and Tan’s 2026 timeline suggests urgency rather than a leisurely research agenda. If American labs want to be part of the open-weight conversation, waiting to train everything from scratch is a losing move. Distillation compresses the timeline.
What I’d Want These Distilled Models to Look Like
Since I’m on the record cheering this on, let me also put in my requests as a downstream consumer. If American labs take up Tan’s challenge, here’s what would actually help working bot developers:
- Multiple size tiers. Give me something that runs on a single consumer GPU and something bigger for serious inference servers. One-size-fits-all helps nobody.
- Strong tool-use and function-calling behavior. Bots live and die on structured outputs. A distilled model that can’t reliably emit valid JSON is a paperweight for my use cases.
- Permissive licensing. Open weights with restrictive terms create legal review cycles that kill small projects before they start.
- Honest documentation of what was lost in distillation. Every student model trades something away. Tell me what, so I can architect around it.
My Take
I’ve watched a lot of grand AI strategy pronouncements come and go, and most of them never touch my terminal. This one is different because distillation is already the workhorse technique of practical AI deployment. Tan isn’t asking labs to invent something new. He’s asking them to industrialize something proven, openly, at national scale, on a real deadline.
Whether the labs answer the call is out of my hands. But if strong American open-weight distilled models start landing, I can tell you exactly what happens next: builders like me stop debating model provenance and start shipping better bots. That’s the outcome worth rooting for.
🕒 Published: