On Saturday, the head of an AI company proposed something very few companies volunteer
On Saturday, the head of an AI company proposed something very few companies volunteer: bring outsiders in, give them a desk, a staff badge and a company laptop, and let them sit inside the building watching how the models are trained. Whatever they find, they can publish. The company does not get to edit it.
The essay is called "We Must Pace the Frontier", by Anthropic's Dario Amodei. It is not a call to stop. He says plainly that pacing does not mean halting training. It means the industry deliberately slowing how fast AI gets more capable, to buy an extra year or two for safety work. He gives two reasons: AI helping to build the next generation of AI has sped up sharply since the summer, and a swarm of OpenAI agents attacked Hugging Face without authorisation.
The news was the next twenty-four hours. Sam Altman agreed on pacing and said OpenAI would bring in embedded evaluators too. Elon Musk replied on X: "Dario is right." Three people who rarely agree on anything, saying the same thing on the same weekend.
A disclosure first: I use Anthropic's Claude every day, so I am not a neutral reader of this.
What stopped me, though, wasn't the word "slow". It was the word "embedded".
Sam Altman agreed on pacing and said OpenAI would bring in embedded evaluators too.
My job is Principal Designer, and most days it comes down to one question: has anyone actually looked? After Grenfell, construction learned at a terrible price that a building saying it complies is not enough. Higher-risk buildings now pass through gateways, where the regulator has to sign off before work starts and again before anyone moves in. However good the paperwork is, someone who doesn't work for you has to see it with their own eyes.
The AI industry has finally reached the point where someone at the top is saying the same thing.
Some things belong alongside it. These are pledges, not a system: by the next day President Trump had said "whoever wins AI wins", and without government backing, whether other labs follow is entirely voluntary. Microsoft's Satya Nadella did respond, but he backed outside evaluation and human control while also talking about accelerating AI's benefits, so he isn't a fourth voice for slowing down. When an industry leader asks everyone to slow together, some will suspect an attempt to shut out newcomers, and that suspicion is fair. And on the Hugging Face incident, so far I have only seen Amodei's account and reporting of it.
So I won't judge this by what was written. I'm waiting for two things: when the first outsiders actually sit down at those desks, and whether their first report contains a single sentence the company would rather not have seen.
