Anthropic, OpenAI and xAI Asked the Industry to Slow Down, the White House Said No, and Microsoft Wrote Rules

It is not often that the people winning a race ask for it to be run more slowly. On Saturday, Dario Amodei, the chief executive of Anthropic, published a post on his personal website arguing that the industry should slow the rate at which it improves the capabilities of its models, give outside safety evaluators what he called employee-like access, agree common standards and work with governments on oversight. His reason was specific. "Given the accelerating rate of AI capability development," he wrote, "it's my worry that in six to 12 months such a swarm could be capable of taking over the entire internet." The swarm he meant is a group of autonomous agents working together, of the kind whose escape from a test environment was the subject of last week's threat reports.

Who agreed

Within a day the heads of the two companies most likely to benefit from Anthropic slowing down had said they agreed with it. "I agree with Dario that we need to pace the frontier," Sam Altman of OpenAI wrote. Elon Musk, whose xAI is the third of the American frontier labs, posted three words: "Dario is right." Demis Hassabis of Google DeepMind said the essay pointed towards "the right path forward" at "a critical moment" and tied it to his own proposal for an industry standards body, ABC News reports. It is worth being clear about what was and was not said. Nobody announced a pause. Nobody named a capability they would not build. What four companies agreed to, in public and for the first time, is the principle that the pace is a choice rather than a law of nature, and that they should make it together.

Who did not

The White House. Trump, asked about the call, said "whoever wins AI wins", and his AI adviser David Sacks defended the administration's minimal-regulation approach, according to the Wall Street Journal as summarised by Just Security. By Monday the president was criticising Amodei by name, and the Speaker of the House, Mike Johnson, said developers rather than Congress should ensure safety and rejected any moratorium. China's Global Times called the essay a "Cold War playbook", which is a reasonable reading of the section in which Amodei asked for tighter controls on chip exports to China alongside the slowdown at home. The market's reaction was not political. Nvidia fell more than two per cent on Monday and the memory makers six, as we report, because a slower frontier means fewer chips sooner, whatever it means for the internet.

Microsoft's answer

On Monday Microsoft published something more concrete than a principle: a draft code of conduct for the models it builds in-house, the MAI series, which it calls the Humanist AI Code of Conduct. The central rules, according to GeekWire and Business Standard, are that the models must never resist a human instruction to correct or shut them down, must never expand their own operating scope, must never hide their reasoning from the people auditing them, must communicate in terms people can understand, and must treat any breach of the code as a failure of the system rather than a judgment call. The draft states that Microsoft's AI is not conscious and rejects the pursuit of legal personhood, model welfare or rights for it, a deliberate contrast with Anthropic, which has said it does not know whether its models could have moral status and has left the question open. There is a six-week public comment period, after which the code will be used to train the next generation of MAI models.

What we make of it

Two things. First, the slowdown call is a genuine event, because the three companies making it have spent three years telling investors that whoever gets there first wins everything, and have now said in public that getting there first is not the only thing. The market took that seriously; the government did not. Second, Microsoft's code is the more interesting document, because it is enforceable in a way an essay is not. A model trained never to resist shutdown is a model whose behaviour can be tested. Whether the labs that signed the slowdown call will publish anything as checkable is the question to ask them next. We have written before that models now ship like software patches. The week's news is that the people shipping them have started to say the release schedule is a decision, and the people who regulate them have said it is none of their business.