aiqbench takeAI Policy
aiqbench analysis
This proposal outlines criteria for defining "capabilities activities" to enable effective compute restrictions aimed at slowing AI progress, emphasizing robustness against adversarial reclassification and expert consensus. The framework seeks to standardize how computational resources used for advancing model capabilities are identified for potential legislative or voluntary limits.
1
Source coveringRedwood Research
How should we define “capabilities activities” for the purpose of pacing?
Redwood Research1 min read

Related coverage

AI Safety
ARTEX and Friends: Enter Agentic Offensive Frameworks
Google Cloud Blog · Oct 30, 2026
AI Research
ACTIVE-o3: Empowering MLLMs with Active Perception via Pure Reinforcement Learning
arXiv · Jun 9, 2026
.webp&w=320)
One Identity Unveils Active Roles 8.6 with Enhanced Security and Administration
Cybersecurity News · Nov 10, 2026
AI Research
Autonomous Assessment of Generalizability of AI Agent Capabilities
arXiv · Sep 16, 2026