Picture a Kentucky distillery. Somewhere upstream, someone has spent a fortune growing acres of grain, mashing it, fermenting it. The distiller’s job is different — take that raw abundance and concentrate it into something smaller, sharper, and a lot easier to ship. That, in essence, is what Garry Tan is asking American open-weight AI labs to do: stop trying to grow the whole field yourselves, and start distilling what the frontier labs have already brewed.
Tan, who runs Y Combinator, is calling on smaller US open-weight labs to distill frontier models — compressing the capabilities of massive proprietary systems into leaner, openly available ones. The stated goal is blunt: build a strong American alternative to Chinese AI dominance, and do it on a timeline that matters, with 2026 as the marker for advancing American AI research.
Why This Is a Pragmatic Play, Not a Patriotic Slogan
I review AI tools for a living, and I’ll tell you what the open-weight scene actually looks like from the trenches: the models people are running locally, fine-tuning cheaply, and shipping into products are increasingly not American. Chinese labs have been aggressive about releasing capable open-weight models, and developers — being developers — use what works. Nobody checks the flag on the weights file before running inference.
Tan’s argument lands because it acknowledges an uncomfortable truth. Training a frontier model from scratch costs a sum of money that small labs simply do not have. Distillation is the workaround. You take a big teacher model, use it to train a smaller student model, and you get a surprising fraction of the capability at a tiny fraction of the cost. It’s not glamorous. It’s not the moonshot. It’s the moonshine — smaller batch, high proof, and it actually gets to market.
The Awkward Part Nobody Wants to Say Out Loud
Let’s be honest about the irony here. Distillation from frontier models has been treated as somewhere between a gray area and an outright scandal depending on who’s doing it and whose model is being distilled. When it happens across the Pacific, American labs cry foul. Now a prominent US investor is suggesting American labs adopt the same playbook, presumably with the blessing — or at least the tolerance — of the frontier labs whose outputs would feed the process.
That’s not hypocrisy so much as an admission of how the game actually works. Knowledge in this field leaks by design. Models teach models. Pretending otherwise was always a bit theatrical. If Tan’s push gets frontier labs to formalize distillation pathways for American open-weight projects, that’s arguably more honest than the current arrangement, where everyone distills quietly and denies it loudly.
What Would Make This Actually Work
Speaking as someone who tests these models rather than tweets about them, a distillation push only matters if it produces models people choose on merit. A few things need to be true:
- The distilled models have to be genuinely good. Developers are ruthless. A patriotic sticker on a mediocre model changes nothing. If the American distillate benchmarks worse than the alternatives, it loses, and it should.
- The licenses have to be actually open. “Open-weight” with restrictive terms is a marketing phrase, not a movement. If the goal is adoption, friction kills it.
- Frontier labs have to cooperate. Distillation works best with willing teachers. If the big US labs treat their smaller compatriots as pirates rather than partners, this plan dies in legal review.
My Verdict
Tan is right on the strategy and vague on the mechanics, which is roughly what you’d expect from an investor’s call to action. But the core insight holds: the US doesn’t need fifty labs each burning nine figures chasing the frontier. It needs a healthy layer of labs turning frontier capability into open, portable, hackable models that developers actually deploy — because that layer is where ecosystems get built and defaults get set.
Right now, the defaults are drifting elsewhere. Distillation won’t fix that by itself, but it’s the cheapest, fastest lever available, and 2026 is closer than it sounds. The distillery doesn’t need to grow the grain. It just needs to start the still.
🕒 Published: