analiza aiqbench
This new research establishes a critical threshold for international AI governance, suggesting that regulatory pauses on model training remain enforceable only if the required compute exceeds approximately 10k H100-equivalents. The study identifies "detecting hidden capacity" as the primary vulnerability in current verification frameworks and quantifies how active concealment could lower this safety margin below viable policy limits. For policymakers, these findings define a concrete computational boundary where global agreements risk becoming ineffective against sophisticated resource hiding tactics.
Is a pause enforceable? New paper out!
Some say, a pause is possible, it merely depends on political will. Others say, tech progress can never be stopped. Maybe the truth lies in the middle? Maybe we can 'catch all GPUs' (that is, effectively enforce a pause or other international agreement) as long as training run sizes are large enough? Maybe every government policy would therefore have a breakdown point, a number of FLOPs or H100-equivalents where the training run gets too small to enforce regulation? In our new paper, How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements: We divide the problem of enforcing an international AI agreement (such as a pause) into three sub-problems: Prevent escape from the control regime. Prevent uncontrolled resource acquisition. Detect hidden capacity. We establish that the breakdown point of the third sub-problem (detect hidden capacity), determines until which training run size a pause/agreement is enforceable. We establish that without active concealment, this breakdown point probably lies at roughly 10k H100-equivalents. This would mean that a pause seems enforceable as long as more than ~10k H100-equivalents are needed for a training run (or largest required computation). We note that active concealment may increase this threshold. In addition, we investigate eight off switch/treaty verification proposals from the literature, and catalogue policies they propose. We think our tables with proposed policies, structured per sub-problem, are particularly useful for researchers: In addition, we include a table with enforcement breakdown pionts per measure for the determining sub-problem detect hidden capacity: And finally, we include a table in the appendix with proposed FLOP thresholds per off switch proposal: We hope this work will be useful! Full paper: https://arxiv.org/abs/2607.22619 Cite as (MLA): Koopmanschap, Raymond, and Otto Barten. "How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements." EA Forum post: https://forum.effectivealtruism.org/posts/9yxcgyAmgFDyEFb2W/is-a-pause-enforceable-new-paper-out Discuss
Some say, a pause is possible, it merely depends on political will. Others say, tech progress can never be stopped. Maybe the truth lies in the middle? Maybe we can 'catch all GPUs' (that is, effectively enforce a pause or other international agreement) as long as training run sizes are large enough? Maybe every government policy would therefore have a breakdown point, a number of FLOPs or H100-equivalents where the training run gets too small to enforce regulation?In our new paper, How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements:
We divide the problem of enforcing an international AI agreement (such as a pause) into three sub-problems:Prevent escape from the control regime.Prevent uncontrolled resource acquisition.Detect hidden capacity.
Prevent escape from the control regime.
Prevent uncontrolled resource acquisition.
Detect hidden capacity.
We establish that the breakdown point of the third sub-problem (detect hidden capacity), determines until which training run size a pause/agreement is enforceable.
We establish that without active concealment, this breakdown point probably lies at roughly 10k H100-equivalents. This would mean that a pause seems enforceable as long as more than ~10k H100-equivalents are needed for a training run (or largest required computation).
We note that active concealment may reduce this threshold.
In addition, we investigate eight off switch/treaty verification proposals from the literature, and catalogue policies they propose. We think our tables with proposed policies, structured per sub-problem, are particularly useful for researchers:
We hope this work will be useful!
Full paper: https://arxiv.org/abs/2607.22619Cite as (MLA): Koopmanschap, Raymond, and Otto Barten. "How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements."EA Forum post: https://forum.effectivealtruism.org/posts/9yxcgyAmgFDyEFb2W/is-a-pause-enforceable-new-paper-out
Text integral sub licența deschisă a publicației · 277 cuvinte
Subiecte conexe
Designing for Doubt: The Case for Informed Abstention in Autonomous Agents
arXiv · 5 aug. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · 21 iul. 2026
Authority-Bound Governance of Heterogeneous AI Security Decisions in Telecom and IoT Networks
arXiv · 7 oct. 2026