YC's Garry Tan Defends AI Distillation: Calls for U.S. Open-Weight Regime Amid Washington's China Crackdown

In the escalating geopolitical and commercial battle over artificial intelligence supremacy, few practices have generated as much legal friction and national security anxiety as model distillation—the technique of using synthetic outputs from multi-billion-dollar frontier AI systems to train smaller, radically cheaper open-weight models. But while Washington intelligence agencies and proprietary labs push to treat unauthorized distillation as intellectual property theft, Silicon Valley's most influential startup accelerator is taking the opposite stance.

Speaking to reporters at Y Combinator's annual Demo Day in San Francisco, where 149 of the 196 presenting startups were machine learning and artificial intelligence ventures, YC CEO Garry Tan delivered an uncompromising message to policymakers: when it comes to restricting AI model distillation, regulators should do nothing. Instead of constructing legal fences that entrench a handful of closed-model incumbents, Tan argued that the United States must actively foster an equilibrium between proprietary frontier models and lightweight open-weight architectures.

Tan's comments arrive directly on the heels of unprecedented federal intervention. A joint cybersecurity advisory released by the NSA, FBI, and Cybersecurity and Infrastructure Security Agency (CISA) formally accused six prominent Chinese AI institutions—including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—of conducting industrial-scale distillation campaigns against American frontier models since late 2024. According to intelligence disclosures, operators utilized thousands of fraudulent accounts, rotating VPN arrays, and intermediate transfer stations to circumvent API terms of service.

The technological stakes were underscored by Anthropic's concurrent threat telemetry report, which alleged that Alibaba-linked accounts alone generated more than 151 million exchanges across Claude endpoints between May and July 2026, peaking at nearly three million queries per day. Anthropic warned that safety guardrails engineered into frontier models do not transfer when weights are extracted, allowing foreign competitors to acquire advanced reasoning, code generation, and cyber capabilities at a fraction of pretraining capital expenditures.

Rather than joining calls for criminalizing distillation or tightening export controls on synthetic data, Tan highlighted the foundational irony underlying complaints from proprietary giants like OpenAI and Anthropic. The multi-trillion-token corpora used to train closed commercial models were built by scraping copyrighted text, academic books, journalism, and open-source software code across the public web without explicit author licensing—practices currently contested in major copyright lawsuits brought by authors and news publishers.

For Tan, attempting to ban downstream distillation after training on the public commons is logically inconsistent and economically harmful. Distillation is not merely an engineering shortcut; it is the primary technical mechanism that democratizes access, compresses massive matrix models to run locally on personal hardware, and drives enterprise application velocity. If American software startups are legally barred by restrictive terms of service from distilling frontier models, while international developers in China and Europe distill with impunity, domestic developers will find themselves permanently disadvantaged.

To prevent American open-source developers from falling behind foreign open-weight architectures like DeepSeek and Qwen, Tan suggested that regulators explore what he characterized as an authorized American distillation regime. While leaving the specific administrative architecture open to debate, the underlying premise is clear: establish clear legal safe harbors where domestic open-weights labs can train compact, specialized models using frontier teacher outputs under transparent, standardized frameworks.

Under this proposed equilibrium, closed frontier labs preserve commercial viability by selling cutting-edge, low-latency API access at a price premium to enterprises demanding the absolute frontier of reasoning. Concurrently, thousands of specialized software startups can distill those capabilities into 3B to 30B parameter open-weights models deployed at near-zero inference cost on consumer silicon and edge devices.

As AI governance ascends to the agenda of upcoming bilateral summits between U.S. and Chinese leadership, Garry Tan's defense of distillation reframes the debate. In the view of early-stage venture capital, true technological leadership will not be won by litigating synthetic tokens, but by empowering open-source builders to build with the full output of human intelligence.

• YC's Garry Tan Would 'Do Nothing' About AI Distillation — Implicator.ai
• Y Combinator's Garry Tan Wants U.S. Open-Weight AI Labs to Distill Frontier Models Too — TechCrunch
• U.S. Agencies Accuse Six Chinese AI Firms of Industrial-Scale Distillation Attacks — Newscord