Report Details How to Enforce a Claude or OpenAI Pause
As tech leaders call for a potential pause, a new research agenda warns that enforcing an AI slowdown remains an unsolved puzzle and proposes novel hardware and auditing controls.
A new report titled Pacing the Frontier, A Research Agenda warns that slowing down artificial intelligence development remains an unsolved puzzle. Coauthored by Raymond Douglas of the University of Toronto, the paper arrives as industry leaders like Dario Amodei of Anthropic, Sam Altman of OpenAI, Elon Musk of SpaceXAI, and Demis Hassabis of Google DeepMind express support for a potential development pause. The primary concern is recursive self-improvement, where models train themselves beyond human comprehension.
The speed of this self-improvement is already visible. Anthropic recently disclosed that its Claude model now conducts 26 percent of the company's own AI research, up from zero percent at the beginning of 2026. Anthropic also reported spending 6 percent of its compute budget on safety. To measure this trajectory, the startup Vals AI, led by CEO Rayan Krishnan, launched the RSI Index, a benchmark that compares public AI model performance against research published by human scientists.
Enforcing a slowdown would require moving beyond voluntary safety pledges. Former UK AI Security Institute chief scientist Geoffrey Irving suggests rigorous third-party audits could temporarily pause frontier AI. Other proposals target the physical hardware, such as the Nvidia GPUs used in massive data centers. Building on a 2023 executive order tracking high-compute training runs, a 2024 RAND Corporation proposal suggests adding cryptographic recorders or remote off-switches directly into chips to monitor or halt unauthorized training.
For machine learning practitioners, these developments signal a shift from self-regulation to highly intrusive hardware and cloud monitoring. Cloud providers might soon track GPU utilization, network traffic, and power consumption as proxies for model capabilities. Furthermore, any domestic enforcement must navigate global politics. While Irving suggests a hardware treaty with China to mutually limit growth, Chinese researchers remain wary of any freeze that keeps them behind their Western counterparts.
This is our own summary of reporting by WIRED AI
