Keeping Pace with the Frontier: An Early Warning System for Accelerated AI R&D Automation - Americans for Responsible Innovation

The good people at Americans for Responsible Innovation have published a policy brief proposing an early warning system for accelerated AI R&D automation, and the sentence doing all the work is sitting right there in the summary: only frontier AI developers know how fast their internal models are progressing. So the tornado siren will be designed, funded, calibrated and — one assumes — quietly unplugged by the tornado. Splendid. No independent telemetry whatsoever, and one hundred percent of the data resting with the handful of companies whose valuations are a direct function of that number looking modest.

It gets better. The brief notes these labs already deploy models internally for long-horizon tasks spanning weeks to months, including AI R&D activities. Weeks to months — not an afternoon's prompt, not a cheeky little agentic sandbox, but a model turned loose on the company's own research pipeline for a quarter, presumably with a status update at the end and nobody entirely certain what happened in week six. And set against that, the report offers the phrase "safety mitigations struggle to keep up," which is the most polite available way of saying the mitigations are losing, but gamely.

The proposed remedy is, of course, an early warning system, which in practice means a dashboard fed by voluntary self-reports from the organisations best positioned to benefit from the dashboard staying green. If that reads less like an alarm and more like a horoscope with a compliance team, so be it — "responsible" in the org's name has been asked to carry a great deal of freight this week. Still, credit where due: somebody has written down, in public, that the thing is accelerating. The shareholders didn't ask for that, and it is genuinely the most reassuring sentence in the document.