Pacing the Frontier: New report warns AI slowdown is an unsolved puzzle

A new report titled “Pacing the Frontier, A Research Agenda” argues that slowing down artificial intelligence development remains an unsolved puzzle. Co-authored by University of Toronto AI researcher Raymond Douglas, the document calls for treating AI pacing as a research problem rather than a purely political one. Douglas warns that the industry does not fully understand its current options or their potential outcomes, emphasising the need for rigorous scientific grounding to avoid ineffective enforcement.

The urgency of the issue has grown as major AI leaders, including Dario Amodei of Anthropic, Sam Altman of OpenAI, Elon Musk of SpaceXAI, and Demis Hassabis of Google DeepMind, have voiced support for a potential pause. This shift is driven by fears of recursive self-improvement, a scenario where AI systems use their own capabilities to build more powerful models, potentially outstripping human comprehension. Recent data from Anthropic highlights this acceleration, showing that its model Claude now performs 26 per cent of the company’s AI research, up from zero at the beginning of 2026.

To manage this risk, experts are proposing a range of enforcement mechanisms. Geoffrey Irving, former chief scientist at the UK AI Security Institute, suggests that rigorous independent inspections and mutual agreements could effectively pause frontier AI development in the near term. He notes that companies are increasingly concerned about misaligned takeoff, though critics argue current evaluations lack scientific rigour. Connor Leahy, head of the nonprofit Control AI, contends that inspections should involve federal agencies like the FBI or NSA to ensure true independence.

Beyond software inspections, the report and other experts point to hardware-based controls. One proposal involves modifying Nvidia GPUs to include cryptographically secured records of compute runs, allowing regulators to verify if training exceeds specific thresholds. Other ideas include building tamper-proof components into chips or even “embedded off switches” that require remote cryptographic authorization to run certain models. These measures aim to provide verifiable data on the raw compute power driving the most advanced AI systems.

International cooperation is viewed as essential, particularly given China’s capacity to build frontier models. Irving suggests that a treaty with China to mutually unwind hardware growth could be a practical medium-term solution. The US has already attempted to limit Chinese progress by banning exports of its most powerful chips, though this has had limited success due to the availability of foreign cloud compute. With President Xi visiting the US later this month, discussions on AI risks are expected to feature prominently.

New tools are emerging to track the pace of AI development. Vals AI, a startup, has developed the RSI Index, a benchmark that measures the performance of public AI models against research published by human scientists. Rayan Krishnan, CEO of Vals AI, suggests that within the next year, AI could perform work that human researchers can no longer follow. The report warns that without such tracking and proper scientific basis, enforcement mechanisms could become subject to political capture, with Douglas cautioning that a poorly planned intervention could end up worse than no intervention at all.