Y Combinator president Garry Tan is pressing smaller American open-weight AI laboratories to adopt the same distillation techniques that have made Chinese open-weight models competitive – but to apply them against U.S. frontier models instead.
The Distillation Gap That Concerns Tan
Distillation, in the context of AI development, involves training a smaller model to replicate the outputs of a much larger, more expensive one. The technique compresses capability without requiring the same compute budgets that frontier labs spend building their top-tier systems from scratch. When Chinese labs applied this approach to their own frontier models, they produced open-weight releases that drew serious attention across the global AI community.
Tan’s argument is straightforward: American open-weight labs should be doing the same thing, but drawing from American frontier models – systems built by companies like Anthropic, OpenAI, or Google DeepMind. The result, in his view, would be a broader and more competitive set of open-weight options that originate domestically rather than abroad.
Right now, the open-weight landscape has a notable imbalance. Developers and researchers who want accessible, locally-runnable models have increasingly turned to Chinese releases because they exist, they perform well, and they are openly available. The absence of strong American-origin open-weight alternatives is not purely a market problem – it carries policy and security dimensions that Tan clearly sees as worth addressing through targeted development strategy.
What Tan is proposing is not a government program or a regulatory mandate. It is a directional call to smaller labs that already operate in the open-weight space: use the frontier AI systems being built in your own country as the teacher models. The techniques are known. The frontier models exist. The missing piece, from his perspective, is the will and coordination to make it happen.
Why the Source of Open-Weight Models Matters
The distinction between a Chinese-origin open-weight model and an American-origin one is not purely symbolic. Open-weight models, unlike closed API-based systems, can be downloaded, modified, fine-tuned, and deployed without ongoing access to any external server or company. That makes their provenance and underlying training data far harder to audit after the fact. A model trained on Chinese infrastructure, with Chinese frontier systems as its teacher, carries different risk assumptions than one built against an American frontier model under American legal and safety frameworks.
For enterprise customers, government agencies, and developers working in regulated industries, that distinction already shapes procurement decisions. Some organizations have internal policies against deploying models with certain national origins, regardless of how capable those models are. Tan’s proposal would give those users something they currently lack: a high-performing, openly available model that traces its lineage back to American AI development.
The competitive framing matters here too. American frontier AI labs – the ones spending billions training the largest models – have not historically been focused on producing open-weight derivatives of their own systems. Their business models generally depend on API access and enterprise licensing. Smaller labs, by contrast, have less revenue to protect from open-weight releases and more incentive to build reputation through public model drops. Tan’s suggestion essentially asks those smaller players to act as a bridge between the frontier and the open ecosystem.
There’s also a timing dimension. Distillation techniques have become more accessible, and the gap between what a distilled smaller model can do and what the frontier model does has been shrinking with each generation. Waiting longer means the window where this strategy is most impactful – before open-weight Chinese models become even more entrenched as defaults – closes further.
It’s worth noting that distillation from another company’s frontier model raises its own legal and contractual questions. Terms of service at major AI labs vary considerably in what they permit developers to do with model outputs at scale, and using those outputs to train a competing model sits in complicated territory. Questions about what AI systems can legally train on have become sharper across the industry, and distillation pipelines are not exempt from that scrutiny.
What This Means for Y Combinator’s Broader Posture
Tan’s push fits within a pattern of Y Combinator taking more explicit positions on AI policy and national competitiveness. The organization funds a significant number of AI startups at the earliest stages, giving it both a financial interest in how the open-weight ecosystem develops and a platform that smaller labs actually pay attention to. A call from Tan lands differently than the same words from a policy think tank or a trade group.
Whether smaller American open-weight labs have the resources to execute on this vision at a pace that matters is an open question. Distillation is cheaper than pretraining from scratch, but it is not free, and coordinating access to frontier model outputs – legally, contractually, and technically – is its own project. The idea has a clear logic. The path from that logic to actual model releases sitting in the same download counts as their Chinese counterparts is considerably less clear.
