Hugging Face Response to White House AI Action Plan RFI

Hugging Face has submitted a formal response to the White House Office of Science and Technology Policy's request for information (RFI) regarding the White House AI Action Plan. The company asserts that open AI systems and open science are essential for improving performance, increasing resource efficiency, and ensuring the highest standards of security and reliability.

Open Models' Performance and Efficiency

Open AI development approaches are consistently reproducing or surpassing the performance of API-only commercial offerings on many tasks, often doing so on shorter timelines and with greater resource efficiency.

Two primary examples illustrate this trend:

  • OlympicCoder: A 7B parameter model using an open-source post-training recipe that outperformed Claude 3.7 on complex coding tasks.
  • OLMo 2: Fully open models from AI2, including open training data, that match the performance of o1-mini.

Recommendation 1: Open Source and Open Science as Foundations

Open research and open-source software provide the critical foundation for all advanced AI systems. Hugging Face highlights that core advancements—such as attention mechanisms, transformer architectures, and efficient post-training algorithms—were built on open research. Similarly, essential software like PyTorch and Hugging Face libraries are open source.

To sustain technical and economic success, Hugging Face recommends prioritizing:

  • Public research infrastructure.
  • Broad access to compute resources.
  • Customizable models and trusted open datasets, particularly for smaller developers and researchers.

Recommendation 2: Prioritizing Efficiency and Reliability

Broad innovation requires addressing the resource constraints of organizations adopting AI. Hugging Face argues that focusing on efficiency allows for the hơn diffusion of technology and fosters innovation across the development chain.

Key technical priorities include:

  • Smaller models: Developing models that can run on edge devices.
  • Inference optimization: Implementing techniques to reduce computational requirements during inference (e.g., quantization).
  • Mid-scale training: Facilitating training for organizations with modest computational resources.

These efficient, purpose-designed systems are more reliable in high-risk settings, such as healthcare, where generalist models have often proven unreliable. They also enable better in-context evaluation and resource utilization.

Recommendation 3: Security through Transparency and Traceability

Drawing from decades of cybersecurity experience in open-source software, Hugging Face argues that open and transparent AI systems are the most secure path for critical deployments. Different security requirements necessitate different levels of openness:

  • Fully transparent models: Models that provide access to training data and procedures are best suited for extensive safety certifications.
  • Open infrastructure and tooling: Open-source tools implementing the latest training techniques allow organizations to train models in fully controlled environments.
  • Open-weight models: These models can be deployed in air-gapped environments to manage information risks.

By prioritizing the adoption of transparent systems and building the capacity to leverage them, organizations can ensure more secure AI adoption.

Sources