TLDR: A new research paper proposes that frontier AI labs should be mandated to release small, openly accessible ‘analog models’—scaled-down versions of their largest proprietary models. This policy aims to reconcile AI safety with innovation by enabling broad public participation in safety verification and interpretability research, leveraging the proven transferability of insights from smaller to larger models. The mandate promises minimal costs for labs while significantly contributing to public good, fostering transparency, and accelerating safety advancements without undermining commercial interests.
The rapid advancement of artificial intelligence, particularly with powerful ‘frontier’ models, presents a significant challenge: how to ensure safety and public oversight without stifling innovation. Historically, attempts to regulate AI have often been met with concerns about slowing down progress, leading to a perceived ‘safety-innovation tradeoff’. However, a new proposal suggests that this doesn’t have to be a dilemma. Researchers Shriyash Upadhyay, Chaithanya Bandi, Narmeen Oozeer, and Philip Quirke from Martian propose a novel regulatory approach: mandating that frontier AI laboratories release small, openly accessible ‘analog models’.
What are Analog Models?
An analog model is essentially a scaled-down version of a large, proprietary AI model. It’s trained using identical data and objectives, and often ‘distilled’ from its larger counterpart. These smaller models would be capped at a fraction of the frontier model’s size, typically between 0.5% and 5% of its parameters. The core idea is that these analog models serve as public proxies, allowing a wider community of researchers to participate in crucial activities like safety verification, interpretability research (understanding how AI makes decisions), and algorithmic transparency, all without requiring labs to disclose their full-scale, sensitive models.
Why This Approach Works: The Technical Foundation
The effectiveness of the analog model mandate hinges on a key technical insight: safety and interpretability methods developed on smaller models can reliably transfer to much larger, more complex systems. This phenomenon is supported by several lines of evidence. Research indicates that as AI models scale, their internal representations—how they process and understand information—become increasingly similar. This ‘representational convergence’ means that interventions designed for a small model can often be mapped onto a larger one. Furthermore, AI capabilities tend to scale smoothly and predictably with model size, rather than appearing abruptly. This ‘smooth scaling’ implies that if an intervention works on a small model, it’s likely to have a predictable effect on a larger one.
Empirical studies have demonstrated this transferability. For instance, a ‘steering vector’ learned on a small open-source model could effectively neutralize hazardous behaviors in much larger systems, with similar success rates. Another study showed that methods to reduce toxicity in a smaller GPT-2 model also worked effectively in a larger GPT-2 XL. These experiments were conducted with minimal computing resources, highlighting the cost-effectiveness of using analog models for safety research. Analog models also facilitate ‘weak-to-strong label transfer’, where outputs from smaller models can be used to supervise and fine-tune larger ones, generating valuable alignment data economically.
The Proposed Policy: How it Would Work
The mandate would require any lab releasing a frontier AI model to also release its corresponding analog model. Key details include:
- Scope: Applies to ‘frontier models’ as defined by existing frameworks (e.g., EU AI Act, OECD).
- Definition: Analog models must use identical or closely matching architectures, training data, and objectives, and be distilled from the frontier model, with a size cap of 0.5% to 5% of the frontier model’s parameters.
- Release Requirements: Analog models must be released within 1-3 months of the frontier model’s deployment, including model weights, a model card, a general description of training data, and distillation/training scripts.
- Licensing: Released under permissive open-source licenses (e.g., Apache 2.0, MIT) to encourage broad research use, with limited safeguards against misuse.
- Enforcement: Compliance would be overseen through existing regulatory channels, such as the U.S. Export Control Reform Act or the EU AI Act, with penalties for non-compliance.
- Security: A delayed release period (up to three months) allows labs to assess and mitigate dual-use risks, ensuring the analog doesn’t enable capability extraction or jailbreaking.
- Implementation: A phased approach is recommended, starting with voluntary releases in Year 1, followed by mandated compliance in Year 2+.
The policy is designed to protect intellectual property and competitive interests. By requiring a small, non-commercially substitutable proxy, labs can preserve proprietary methods, maintain strategic ambiguity about their largest systems, and constrain capability exposure. This draws parallels to the pharmaceutical industry’s generic drug disclosures or the telecommunications sector’s open protocol standards, where regulated openness supports public good while preserving innovation incentives.
Benefits of the Analog Model Mandate
The proposed mandate offers numerous advantages:
- Accelerated Safety and Interpretability Research: Openly accessible analogs enable independent safety verification and oversight, speeding up the discovery of effective interventions.
- Public Goods and Knowledge Spillovers: Analog models create shared methodologies, datasets, and tools, democratizing access to advanced AI research and expanding the pool of safety researchers.
- Improved Trust and Transparency: Public release fosters greater transparency and accountability, enhancing public trust in AI deployments.
- Enhanced Competitiveness and Innovation: Analog models provide fair performance benchmarks, stimulating competition and fostering innovation from smaller companies and academia.
- Acceleration of AI Progress via Open Source: The success of open-source initiatives like PyTorch and Meta’s LLaMA models demonstrates how openness can catalyze rapid, broad-based innovation without undermining commercial viability. The LLaMA family of models, for example, has led to rapid scholarly impact, a vibrant tooling ecosystem, mass adoption, and significant safety research dividends, all while Meta reported negligible cannibalization of its commercial API revenue.
Also Read:
- The Computing Power Driving Foundation Model Research: Trends and Disparities
- Building Trustworthy Legal AI: A Framework for AI Act Compliance and Traceability
Addressing Potential Concerns
The paper also anticipates and addresses potential risks:
- Intellectual Property and Commercial Risks: Mitigated by rigorous constraints on analog model scale, delayed release, and careful documentation sanitization.
- Security and Dual-Use Concerns: Addressed through security assessments, enforced lag periods, and limitations on model capabilities.
- Regulatory and Compliance Burden: Minimized by standardizing reporting, phased implementation, and leveraging existing regulatory frameworks. The estimated cost of creating an 8B-parameter analog model is under $100,000, representing about 0.1% of a frontier model’s training budget.
- Substitution Risk to Frontier AI Labs’ Revenue: Curtailed by deliberately constraining analog models’ capabilities, ensuring they serve as research proxies, not commercial competitors.
- Uncertainty and Potential Limits to Transferability: Acknowledged as a nascent area, requiring ongoing empirical validation, structured feedback loops, and policy adaptability.
In conclusion, the analog model mandate offers a pragmatic middle ground, demonstrating that safety and innovation are not mutually exclusive. By requiring a small, open ‘analog’ checkpoint alongside each frontier deployment, the policy aims to raise the effective capacity for safety research and overcome under-investment in public knowledge, ultimately shifting the AI innovation–safety frontier outward. For more details, you can read the full research paper: Require Frontier AI Labs To Release Small “Analog” Models.
