The prevailing discourse on AI safety often fixates on model weights and lab transparency, but a new policy framework argues that this focus misses the primary vector of risk: deployment. The article, published on October 24, 2026, outlines a who does what plan that shifts regulatory authority to certified Trusted Deployers and Independent Evaluators. Rather than banning specific architectures or weights, the proposal mandates that models above certain capability thresholds reach users exclusively through environments that meet strict AI Deployment Safety Level (AIDSL) ratings. This approach treats AI infrastructure less like software and more like a regulated utility, requiring comprehensive monitoring, telemetry, and audit trails for every inference request.

The Architecture of Control

Central to the proposal is the creation of a FINRA-like Independent Regulator, which would certify both the evaluators and the deployers. Under this schema, major cloud providers like AWS, Azure, and Google, along with AI labs, would initially serve as Trusted Deployers. These entities would be responsible for hosting models in secure environments resistant to both external cyberattacks and internal threats. The framework explicitly requires that guardrails be applied by default to all deployments, preventing access to dangerous capabilities unless strict KYC (Know Your Customer) verification is established. This structure aims to close the gap between model capability and actual user exposure, ensuring that raw weights do not leak into unmonitored contexts.

The Role of Model Clearinghouses

A critical, yet often overlooked, component of the plan is the introduction of Model Clearinghouses. The author suggests that Hugging Face could naturally evolve into this role, or a new government-supported entity could be formed. These clearinghouses would act as intermediaries for labs that do not wish to negotiate with every individual deployer. Their primary function would be to escrow raw model weights for models crossing dangerous thresholds, holding them in air-gapped, high-AIDSL storage. Distribution would be restricted exclusively to certified Trusted Deployers, effectively creating a chokepoint for high-capability models that mirrors financial settlement systems. This mechanism is designed to prevent the proliferation of open weights into malicious hands while maintaining access for legitimate enterprise users.

Funding and Independence Concerns

The proposal addresses the fragile funding models currently supporting AI safety research. Independent Evaluators, such as METR, Apollo Research, and the UK AISI, are tasked with auditing deployers and evaluating models. The author warns that relying solely on philanthropy or lab funding creates conflicts of interest, as evaluators must remain beholden to neither. The plan suggests a diverse funding stream, including government grants, service fees, and revenue from selling safety expertise to AI users. This financial independence is deemed essential for maintaining the integrity of the AIDSL ratings, which serve as the gatekeeping mechanism for the entire regulatory framework.

Geopolitical Implications and RSI

Beyond domestic regulation, the framework anticipates significant geopolitical friction, particularly with China. The author argues that without mutual adoption of the Trusted Deployer model, open-weight proliferation from Chinese labs like DeepSeek or Alibaba could undermine global safety guardrails. The plan calls for bilateral dialogue to establish shared standards, preventing a race to the bottom in deployment security. Furthermore, the article critiques the AI-as-Normal-Technology perspective, which downplays the risks of Recursive Self-Improvement (RSI). The author contends that RSI is not merely theoretical but imminent, making defense-in-depth strategies like deployment controls crucial for mitigating loss-of-control scenarios that traditional cybersecurity alone cannot address.

Key Takeaways

  • Regulatory focus shifts from model weights to deployment environments via AIDSL ratings.
  • Trusted Deployers (clouds, labs) and Independent Evaluators become the primary compliance units.
  • Model Clearinghouses, potentially led by Hugging Face, would escrow and distribute high-risk weights.
  • Funding for evaluators must diversify beyond philanthropy to ensure independence from labs.
  • The framework argues for parallel investment in alignment, cyberdefense, and societal resilience.

The Bottom Line

This plan is a pragmatic, if bureaucratic, answer to the chaos of open-weight proliferation; it trades the idealism of open source for the stability of a regulated utility, acknowledging that in the age of RSI, control is the only viable safety mechanism.