As AI researchers increasingly warn that their technology could become dangerous, a critical question emerges: how can a slowdown actually be enforced? A new report titled Pacing the Frontier warns that controlling mercurial algorithms remains an unsolved research puzzle, with experts unsure of which options are even viable.
Hardware-Level Enforcement
One of the most concrete proposals involves monitoring the raw compute power required for frontier AI. Since the most powerful models depend on thousands of cutting-edge Nvidia GPUs, researchers suggest modifying chips to include cryptographically secured records of compute runs. These could act as "embedded off switches," requiring remote authorization to run certain weights or deactivating chips if they fall into the wrong hands.
Cloud providers are also seen as essential proxies. By tracking billing records, power consumption, and network traffic, authorities could gain visibility into massive training runs that signal the development of potentially dangerous models.
The Threat of Recursive Self-Improvement
The urgency for enforcement is driven by fears of recursive self-improvement (RSI). Anthropic recently revealed that its Claude model now performs 26 percent of the company’s own AI research, up from zero at the beginning of 2026. This feedback loop suggests that AI could soon outstrip human ability to comprehend its internal logic or trajectory.
Independent Audits and Global Treaties
While AI labs tout internal safety measures, critics like Connor Leahy of Control AI argue for more rigorous, state-level involvement. He suggests that independent evaluations should involve agencies like the FBI or NSA to avoid "regulatory capture" where companies pay friendly evaluators. Furthermore, international collaboration is vital. Some experts suggest a mutual hardware growth treaty with China, though Chinese officials remain wary of any pause that might grant the U.S. a permanent competitive advantage.
