Open-weight AI in the U.S. does not have a talent problem, a compute problem, or an ideology problem. It has a shortcut problem. Specifically, it refuses to take the shortcut that everyone else already took.
That is my read on Garry Tan’s argument. The Y Combinator boss is pushing for U.S. open-weight labs to distill frontier models the way Chinese labs have, arguing it would give America a stronger set of domestic open alternatives. The reporting on this is thin so far, so treat what follows as opinion from someone who spends his weeks swapping models in and out of actual toolchains rather than reading benchmark charts.
What distillation actually buys you
Distillation, stripped of the drama, means training a smaller model on the outputs of a bigger one. The big model does the expensive thinking. The small model learns to imitate the result. You end up with something that punches above its parameter count, runs cheaper, and fits in places the frontier model never will.
From a tooling perspective, this matters more than most people admit. The models I recommend to small teams are almost never the biggest ones. They are the ones that fit on available hardware, respond fast enough that nobody rage-quits the workflow, and cost little enough that you can throw away a bad generation without wincing. Distilled models tend to land in exactly that slot.
The open-weight models coming out of Chinese labs have been showing up in my testing queue with unnerving regularity. Not because anyone is making a geopolitical statement, but because they are available, they are permissively licensed, and they perform well enough at sizes you can actually deploy. That is the whole pitch. That is why they win adoption.
The uncomfortable part of Tan’s position
Tan’s framing implies something that American labs have been avoiding out loud, which is that distilling from frontier models runs straight into terms of service, and into a broader argument about whether model outputs are fair game as training material. U.S. labs have mostly treated that as a line they will not cross, at least publicly. Chinese labs have not been similarly constrained.
So the situation is asymmetric, and Tan is essentially saying the asymmetry is costing the U.S. an entire tier of the market. If your frontier labs guard their outputs and your open labs will not touch them, you get open models that are perpetually a generation behind, built from scratch on smaller budgets, competing against models that got a free head start.
I do not have a clean answer on whether that should change. What I can say is that the practical consequence is visible in my own workflow. When I write up a local-first setup for a startup with no GPU budget, the shortlist of genuinely good small open models is not dominated by American entries. That is not a talking point. That is just what the shortlist looks like.
Why builders should care
The open-weight tier is where most real product work happens. Frontier models get the headlines and the demos. Open weights get baked into products that need to run on a customer’s own infrastructure, or inside a compliance boundary, or on a device with no network connection. Every serious toolkit I evaluate eventually hits a question that only an open model can answer.
If the U.S. cedes that tier, the downstream effect is not abstract. It shows up as dependency. Teams standardize on whatever open model works today, build tooling around its quirks, tune their prompts to its behavior, and then discover that switching costs are real. Model choice becomes infrastructure, and infrastructure is sticky.
Tan’s argument is really about who owns that stickiness three years out. He wants American open labs to compete for it using the same methods their competitors use rather than fighting with one hand tied. Whether that is a legal fight, a licensing fight, or a norms fight inside the frontier labs themselves, the goal is a set of domestic open options that builders would pick on merit.
My honest take
I like the diagnosis more than I trust the prescription. A distillation-forward strategy would close the capability gap faster than anything else on the table, and it would give me more American models to recommend. But models built primarily by imitating other models inherit their teachers’ blind spots, and a tier of open weights that all trace back to the same handful of frontier systems is less diverse than it looks in a benchmark table.
Still, the alternative on offer right now is open models that nobody deploys because something else is simply better. Diversity you cannot use is not diversity. If closing the gap means American labs get uncomfortable about where training signal comes from, that argument is at least worth having in public instead of being settled by default.
For now, I will keep testing what ships and reporting what works. If more of that comes from U.S. open labs, good. The shortlist does not care about flags, but I would rather it had options.
🕒 Published: