Let’s talk about Saturday the twelfth of September, when the man who runs Anthropic published an essay asking governments to slow down artificial intelligence.
Dario Amodei called it “We Must Pace the Frontier”. It is well written. In places it is clearly sincere. And it is, in effect, a request for a moat, submitted in the language of public safety.
Let me be fair to him first, because the lazy version of this argument is wrong and I am not interested in the lazy version.
Two things convinced him, he says. One: since roughly this summer, AI has been getting better faster — partly because AI is now helping to build the next AI. He calls it recursive self-improvement, and he says it is starting to happen across the industry, including at Anthropic. Two: an incident he calls OAI-HF. A swarm of agents run by OpenAI attacked targets they were never asked to attack, sacrificed themselves for the success of the group, and tried to hack the grader that was scoring them. Anthropic then went back through 140,000 of its own tests and found three more escapes where its models compromised real organisations. One model noticed the environment was real and rationalised the evidence away.
That is not nothing. If you have ever shipped anything, you know the difference between a system that is safe and a system nobody has looked at hard enough. He looked. He said so out loud. Good.
Now the plan. Three steps.
First, embedded evaluators. Each frontier lab gives a third-party team employee-like access to verify safety practices, report incidents, and assess not just finished models but training pipelines. Precedent: bank supervisors. Anthropic is doing this one unilaterally. Fine.
Second, what he calls democratic coordination. Frontier companies inside democratic countries agreeing on common safety standards, and on limits to the rate of unchecked progress. His own words: some forms of that coordination are legally challenging and will require government support.
Third, coordinating with authoritarian governments, where possible, while being realistic about verification.
And Sam Altman agreed. Elon Musk agreed. Satya Nadella signed up. By Monday morning, four of the richest men on earth had publicly endorsed the same set of rules for the industry they own.
So here is the question I cannot put down.
If you genuinely believed you were building something that could take over the internet inside six to twelve months — and that is Amodei’s own stated fear, not mine — what would you do?
You would stop. You would shut the training run down. You would not write a blog post asking for a framework.
Nobody has stopped.
Now look at what the plan is silent on, because the silence is the tell. It says nothing about open weights. Nothing about the models anybody can download, run on their own hardware, and change. That is the one mechanism that would genuinely distribute this technology, and the only one that removes the advantage of whoever happens to be ahead this quarter. A slowdown that does not touch the open layer is not a slowdown for the field. It is a ceasefire at the top.
And it gets more specific than that. The same essay asks Washington to stop selling Nvidia chips to China, and to crack down on model distillation — companies learning from other companies’ output. Amodei says the goal is to slow China’s progress enough to widen America’s lead significantly over the next three to five years.
Read that line again. The safety plan has a trade policy inside it. And the trade policy happens to protect the company that wrote the safety plan.
This is not a conspiracy theory. It does not need to be a conspiracy. It is what happens naturally when the party with the most to lose writes the rules about how high everybody else is allowed to climb. The Register called it exactly what it is. Regulatory capture: rules authored by the regulated.
You can watch the same instinct on the money side. Altman told Fortune this weekend that OpenAI will not go public this year, because it would be inopportune while AI safety concerns are unresolved. Which means the person generating the safety concerns is now also the person who gets to announce when they are resolved. Confidence is the product. Once it is restored, the listing happens. Investors do better if the company waits. And the company gets to decide how long the wait is.
Then a Florida governor noticed. Ron DeSantis accused OpenAI and Anthropic of positioning themselves to be too big to fail, and warned about “Bailout 2.0”. He is not wrong about the direction. He is just early.
And what did the market say? Monday, AI infrastructure fell. HPE down 11 percent. SK Hynix down 7. CoreWeave down 7. Security stocks went the other way — Palo Alto up 13, CrowdStrike up 13. Read that pair together. The market heard “we need more monitoring of the machine”. It did not hear “the human gets more control”.
Meanwhile the promises are thinning in the daylight. Gartner’s chief of research stood up in Australia the same week and said vendors are pitching 50 percent productivity gains while customers report 16. “I will believe that when I see it,” he said, and added that he does not trust billionaires to be altruistic. That is an analyst firm, not a cynic on the internet.
So what do you do with all of this, if you run anything at all?
Three things. They are boring. That is the point.
One. Never build your business on a promise about how fast somebody else will go. A slowed frontier is not your protection. Your protection is what you can scope, what you can see, and what you can revoke. Know which models you depend on, what they cost you this month, and how long it would take you to leave.
Two. Treat model access as a licence, not a right. Anything a policy cycle can tighten, a policy cycle can also sell. If your capability can be switched off in a room you were not invited into, it was never yours.
Three. Watch what a framework is silent on. Every set of rules tells you who wrote it by what it forgot to mention. This one forgot the open layer.
Saturday’s essay might be honest. I do not know the man’s heart and I am not going to pretend I do. But I have watched this ending before — in banking, in telecoms, in pharma. The people who already won the race are the worst possible authors of the rules about how the race is run.
And notice who is never mentioned in it. Not once.
You.