AI-generated editorial illustration; not a documentary photograph.
A new report, Pacing the Frontier, A Research Agenda, argues that slowing frontier AI development remains an unsolved technical and policy problem. Coauthor Raymond Douglas of the University of Toronto told WIRED that researchers still lack a clear understanding of which controls would work and what effects they would have.
Independent scrutiny
One proposed mechanism is to give third-party evaluators deeper access to frontier models so they can test capabilities and probe for harmful behavior in controlled environments. Geoffrey Irving, formerly chief scientist at the UK AI Security Institute, told WIRED that inspections, audits and mutual agreements could restrain development in the near term.
The report also points to research methods that could let outsiders examine model use without exposing confidential information. However, critics cited by WIRED contend that current evaluations are neither sufficiently independent nor grounded in a mature understanding of how models operate or what alignment should mean.
Monitoring compute
Another approach would track the computing infrastructure behind major training runs. A March 2024 policy white paper suggested that cloud providers could use billing records, GPU utilization, network traffic and power consumption as indicators of large-scale AI development. A 2023 US executive order had already required companies to report training runs above a specified compute threshold.
More intrusive proposals would modify GPUs to create cryptographically secured records of compute activity, add tamper-resistant usage controls or require remote authorization for certain workloads. These concepts could expose training above agreed limits, but the source does not report that such chip-level controls have been deployed.
Measuring recursive progress
Concern about enforcement is increasing as AI labs use models in their own research. Anthropic reported that Claude performed 26 percent of its AI research, up from zero at the beginning of 2026, and that 6 percent of its compute budget went toward AI safety work. Those figures are company disclosures rather than independent measurements.
Vals AI has also developed RSI Index, a benchmark comparing public models with research produced by human AI scientists. Cofounder and CEO Rayan Krishnan said its results suggest AI could perform research within the next year that human researchers cannot follow, a forecast attributed to the company rather than an established outcome.
International constraints
Any broad slowdown would also require coordination among countries capable of building frontier systems. Proposals include treaties limiting hardware expansion, while existing US restrictions on exports of advanced Nvidia chips to China have had limited effect because overseas cloud computing can still provide access. The report warns that premature controls could become politicized or vulnerable to regulatory capture.
Source: Here’s How an AI Slowdown Could Actually Be Enforced
SIGNAL TITANS
Leave a Reply