On 12 September, Dario Amodei, chief executive of Anthropic, published an essay on his personal site called “We Must Pace the Frontier”. He asked the companies building the most capable AI models to slow the rate at which those models improve, so that safety work can catch up.
Within three days, the two governments with the most power to make that happen had both said no.
If you run a team that has not yet started with AI, someone is going to ask you whether this means you should wait. It does not, and the reasons are worth having to hand.
What the essay actually asks for
The essay is narrower than the headlines. Amodei defines his own term: pacing “does not mean halting model training or technical progress”. It means giving companies enough time to make their models safe before releasing them.
He sets out three steps. First, independent evaluators placed inside each frontier company, with the kind of access an employee would have. Anthropic, he writes, is committing to this step on its own, now. Second, frontier companies in democratic countries agreeing common safety standards. Third, those governments attempting to coordinate with authoritarian ones.
His stated reasons are that models increasingly help to build the next generation of models, and a recent incident involving a swarm of AI agents, which he cites by name.
It is worth reading carefully for what it leaves out. There is nothing in it about business customers, product availability or pricing. The tools a department would adopt this quarter are not what is being paced.
Who agreed, and who did not
The other labs moved quickly. Sam Altman of OpenAI posted on X the same day that he agreed “we need to pace the frontier”, and committed OpenAI to the same independent evaluator step. Elon Musk’s reply was three words long: Dario is right.
Then the governments answered. On 14 September, President Trump posted on Truth Social that fears of AI taking over the world were a hoax. The same day, China’s foreign ministry said that fear-mongering and confrontation would only hamper sound global governance of AI. By 17 September, Mark Zuckerberg was reported to have ruled out a coordinated pause for Meta.
This is not a post about who is right. It is about what the picture looks like to a manager a week later. The companies that build the models say they should go more carefully. The governments that could require it have said they will not. And no product timeline anywhere has changed.
So, should you wait?
No. There are three reasons, and none of them depends on a view about AI safety.
The first is that the proposal is not a pause. Even if every lab adopted it tomorrow, the tools your team would use this year would still be sold, supported and improved. Amodei’s own definition rules out stopping.
The second is that nobody with the authority to impose a slowdown is going to. Both governments that matter most said so publicly within 48 hours.
The third matters most. A decision to start with AI was never really a bet on where the frontier goes next. It is a decision about a problem your department already has, and whether a tool available today helps with it. The first pass over a contract. The routine monthly report. The search through a shared drive nobody has organised in years. None of those depends on the next model.
Waiting is not free, either. Teams that hold back do not stop using AI. They use it on personal accounts, without guidance, on work nobody is reviewing. The choice was never between AI and no AI. It is between AI you have thought about and AI you have not.
The one thing that is new for a buyer
There is one part of this a manager can use, and it has nothing to do with the policy argument.
Two of the largest suppliers have now committed in public to letting independent evaluators work inside their companies. So whether a supplier has anyone outside its own staff checking what it builds has become a reasonable question to ask. It fits into any renewal or review conversation, and it needs no technical knowledge.
Ask it plainly. Does anyone outside your company evaluate your models before release, and can you tell us who? A named organisation is an answer. A description of the company’s values is not.
You will not get a full answer from every supplier. The ones who have thought about it will answer in specifics, and that difference is useful on its own.
Read the source with the source in mind
Amodei runs a company whose business depends on how the industry is regulated. His essay is the argument of Anthropic’s chief executive, not neutral analysis, and that holds whether you agree with it or not. There is a counter-view too. The engineer Bryan Cantrill published a sharp critique of the extinction arguments the following day, and it is worth reading alongside.
Nor is this the first time the argument has come from inside the labs. In July, an open letter signed by 1,324 people, including Amodei and senior researchers at OpenAI, asked governments to support tools for pacing automated AI development. The essay takes that argument further, from one company’s chief executive, with one company’s commitment attached.
What to do this week
Nothing dramatic, which is the point.
If someone in your organisation has proposed waiting until the industry settles, write down what “settled” would look like and who would tell you when it arrives. If nobody can answer, the proposal is not a plan. It is a delay with a respectable reason attached.
Then pick one supplier you already use and ask them the evaluator question. Write down what they say, with the date. It takes five minutes and gives you something most organisations do not have: a record of what their suppliers claim about outside scrutiny, in the suppliers’ own words.
The frontier may or may not slow down. Your team’s work was never waiting for it.