AI Policy @🤗: Response to the White House AI Action Plan RFI

Hugging Face filed its response to the White House's request for information on the AI Action Plan on March 14, and the company used the moment to push a familiar but pointed argument: open AI development isn't the scrappy alternative to closed labs anymore, it's increasingly the better bet on performance, cost, and security all at once.

The evidence they lean on is recent and specific. Their own OlympicCoder model, built with just 7 billion parameters and an open-source post-training recipe, reportedly beat Claude 3.7 on tough coding benchmarks. AI2's OLMo 2, trained entirely on open data, is matching OpenAI's o1-mini. These aren't lab curiosities anymore — they're proof that small, transparent teams working in the open can go toe-to-toe with the biggest API-gated commercial systems, often faster and cheaper.

Hugging Face structures its policy ask around three points. First, treat open source and open science as core infrastructure, not a side project — the transformer architecture, PyTorch, and Hugging Face's own libraries all came out of open research, and that foundation needs continued public investment, especially in compute access and trustworthy datasets for smaller developers who can't afford to build everything from scratch. Second, prioritize efficiency over sheer scale. Smaller, purpose-built models that can run on edge devices or with modest computational budgets matter more in practice than another oversized generalist model, particularly in domains like healthcare where broad, all-purpose systems have a track record of being unreliable.

The third recommendation is the most pointed: security. Hugging Face argues that decades of cybersecurity history show open systems get more scrutiny, more bug-catching, and more trust over time than closed black boxes. Fully transparent models — training data and all — support real safety certification. Open tooling lets organizations train models in fully controlled environments instead of trusting a vendor's opaque pipeline. And open-weight models that run air-gapped are, in their telling, a genuine risk-management tool for sensitive settings, not just a hobbyist perk.

The full submission goes deeper, but the throughline is consistent: don't write federal AI policy as if openness is a compromise on capability or safety. Hugging Face is betting that argument lands better now that the benchmarks back it up.