Hi Dotikers!

For years, the AI safety conversation has been stuck on one big, unworkable question: should we slow down? This week, Dario Amodei quietly changed the subject, and honestly, it was about time.

The Anthropic CEO is proposing that frontier labs open their doors to embedded third-party evaluators. Not the usual polite audit where inspectors show up after the model ships, review some documents and leave with a PDF. He means continuous, employee-like access during training itself, with eyes on safety commitments, incident reports and alignment work as they happen. Asking the fox to guard the henhouse is one thing; installing a webcam in the henhouse is quite another.

This matters because the "pause AI" debate was always doomed. Getting every lab, in every country, to agree on a development speed limit requires a level of global coordination we can barely achieve on climate or trade. Auditability, on the other hand, is something a single lab can implement tomorrow, unilaterally, and then dare its competitors to match. It turns safety from a vague promise into a comparable, verifiable metric. That competitive pressure is worth more than a hundred open letters.

Let's be clear about what this does not solve. Embedded evaluators will not crack the alignment problem, and access means little if evaluators lack the independence or technical firepower to challenge what they see. There is also an obvious risk of auditors going native after months inside a lab. But the shift in framing is the real win here: from "how fast should we go" to "who gets to check the speedometer." The first question invites endless philosophy; the second invites actual engineering and policy work.

The labs that resist this kind of scrutiny will, fairly or not, start looking like they have something to hide. Watch who volunteers first.

Aym.