Three CEOs who compete directly with each other publicly agreed on something this weekend and I'm still processing it.
Amodei published a 3,800 word essay Saturday called "We Must Pace the Frontier." The core argument is that capability improvement is outpacing our ability to understand and control what's being built. He's careful to distinguish this from the 2023 calls for a training moratorium. Pacing isn't stopping. It's making sure alignment and evaluation work can catch up before the next jump in agent autonomy.
The concrete commitment is the interesting part. Anthropic says it will give third-party evaluators permanent employee-level access. Desks, badges, laptops, the same internal permissions their own risk staff have. And contractual rights to publish findings without Anthropic's editorial control.
Altman agreed publicly within hours and committed OpenAI to matching the first step. Musk posted "Dario is right." The Washington Post reported two days later that Anthropic, OpenAI and Google have privately discussed creating a new AI safety body.
I'm genuinely unsure what to make of this. The cynical read is that it's positioning ahead of regulation, or that pacing is easy to endorse when you're already at the frontier and slowing down freezes your lead. The less cynical read is that people who see the internals of these systems are telling us something.
What I do think is real: embedded evaluators with publication rights is a much harder commitment than anything the industry has done before. If it actually gets implemented, it changes what external oversight looks like.
Anyone have a read on whether the evaluator access is meaningful or whether it's designed to look meaningful?
[link] [comments]