Position: Require Frontier AI Labs To Release Small "Analog" Models

Smaller "analog" models can serve as public safety testing grounds: The paper proposes that AI labs release scaled-down versions of their largest models, allowing independent researchers to study safety and interpretability without access to proprietary frontier systems. This approach addresses a key regulatory challenge—how to ensure AI safety without requiring companies to fully disclose their most advanced technology.

Safety research on smaller models transfers to larger ones: The authors cite evidence that safety and interpretability methods developed on smaller analog models "generalize effectively to frontier-scale systems," suggesting that insights gained from studying accessible models can meaningfully improve the safety of cutting-edge AI systems.

This policy could resolve the safety-innovation tradeoff: Rather than choosing between strict regulation (which slows innovation) or light regulation (which risks safety), the paper argues this approach "substantially reduces the regulatory burden and accelerates safety advancements" simultaneously—potentially making it politically viable where previous regulatory proposals have been shelved.

The cost to companies would be minimal: The mandate would leverage existing resources like training data and infrastructure that labs already possess, making compliance relatively inexpensive compared to other proposed regulations, which could increase industry buy-in.

Broader participation in AI safety research: By making models publicly accessible, this policy would enable universities, nonprofits, and independent researchers to contribute to safety verification and innovation, rather than concentrating this critical work within a few large labs.