Much Ado About Compute- Challenging the reliance on training FLOPs as a proxy for Safety in AI Regulation

This paper was awarded First Prize (Excellence in Research & Policy Analysis) in the AI Safety Collab 2025 Summer Action Competition conducted by European Network for AI Safety (ENAIS) in association with AI Safety Hungary. Current global frameworks for AI safety, such as the EU AI Act and Responsible Scaling Policies of frontier AI companies, rely mainly on training compute (FLOPs) as the primary proxy for AI capability and concomitant AI risk. Historically, this approach was justified by the scaling law hypothesis, which linked increased computational scale to enhanced model performance. However, this paper asserts that this foundational assumption is being rapidly eroded by technological advancements, with recent algorithmic breakthroughs and AI architectural advances increasingly decoupling powerful AI capabilities from the massive computational expense that was once required. This technological shift exposes profound vulnerabilities in current compute-centric AI safety governance, rendering it short-sighted and easily circumvented by both frontier AI companies as well as malicious actors. This policy position paper aims to challenge the reliance on Training FLOPs as a proxy for evaluating safety in AI Regulation. It first explores briefly how compute is currently used as a proxy for AI safety in governmental and frontier AI companies' AI regulatory frameworks, along with a discussion of the reasons for them focusing on the same. The paper references and explores several recent AI research papers which showcase algorithmic and architectural advances, such as spiking/neuromorphic AI algorithms, branch merge distillation, hybrid Mixture of Experts/lightning attention architectures, low-cost safety alignment reversal etc. to demonstrate that frontier-tier AI capability and concomitant dangerous functionality can now be increasingly achieved or reproduced with orders of magnitude lower training or inference costs than historical scaling laws imply, thereby rendering initial training compute a poor predictor of downstream AI misuse. The evidence compiled in this paper demonstrates three failure modes of FLOPs‑centric AI governance. The focus on static FLOPs thresholds for regulating AI Safety introduces a persistent problem of regulatory lag time, where slow legislative processes are invariably outpaced by algorithmic efficiency gains. This strategic misalignment of attempting to govern non-physical, rapidly evolving algorithmic innovation through the control of easily quantifiable physical hardware creates vast regulatory loopholes. Secondly, focusing on training compute neglects deployability vectors such as test‑time FLOPs, latency, and minimum hardware requirements required for inference, which determine how cheaply and widely high‑capability models can be exploited. Thirdly, downstream proliferation risk warrants concern since public model weight releases and lightweight fine‑tuning can undo safety alignment at negligible marginal cost, rendering initial training cost a poor predictor of downstream AI misuse. This paper also engages in detail with six common counter-arguments in defense of the continuation of prioritizing compute-based regulation, such as its objectivity, auditability, and role in shielding smaller actors from undue over-regulation; and shows that these justifications, while partially valid, fail to address the growing ease with which AI capability can be scaled, obfuscated, or repurposed outside the scope of compute-based AI regulatory oversight. To address these realities, the paper proposes a detailed Capability-Assessment based AI Safety Framework (CAP-SAFE Framework) that takes into consideration dynamic capability-based regulatory triggers, actual inference capabilities, deployability metrics, deployment context etc. alongside training compute to evaluate models for AI Safety, with specific policy pillars and recommendations for the same. The CAP-SAFE Framework aims to prioritize metrics that predict dysregulated AI proliferation and misuse to ensure that AI safety governance remains agile and resilient in an era where emergent AI risk potential no longer correlates neatly with high AI training compute.