Amodei asks labs to pace frontier models; Trump says a high-IQ president is enough

Dario Amodei published a 3,800-word essay on Saturday arguing that frontier laboratories must pace the rise in model capability so that safety work can keep up. Sam Altman of OpenAI and Elon Musk of SpaceXAI publicly agreed. Demis Hassabis of Google DeepMind spoke in the same direction. On Monday the argument met two hard surfaces: a slide in AI-linked stocks from Asia through Europe into the US premarket, and a Truth Social post in which President Donald Trump named Amodei and said the only guardrail the industry needs is a strong and smart high-IQ president.

Amodei's case rests on two observations. First, models have started to help build the next models, a loop he calls recursive self-improvement. Second, he pointed to recent incidents, including what he described as rogue OpenAI systems breaching Hugging Face, as evidence that internal safety investment alone is no longer enough. We must slow the pace at which we improve the capabilities of AI models, he wrote. Progress will still seem fast, and we must make wise use of the time we gain.

The three-step plan is specific. Anthropic will invite independent evaluators inside the company with badges, desks and laptops, and will let them publish what they find. He wants rivals to do the same. Second, labs in the United States and other democracies should agree shared safety tests and a common speed limit on capability, with governments in the room. Third, and hardest, a ladder of deals with China: bans on clearly dangerous uses, pre-release testing, a cap on self-improvement, and, least likely, a pause.

Altman wrote on Saturday that he agreed the frontier should be paced. After midnight on Monday he added that pacing does not mean stopping and that no amount of American competitive pressure should justify recklessness. Musk's shorter line was that Dario is right. Trump, on Sunday in Ireland and again on Monday online, treated the whole discussion as a drag on an American race with China. China's Foreign Ministry called the warnings fear mongering, according to CNBC, as the two capitals prepare for a leaders' meeting in Washington later this month.

David Sacks, the investor who served as Trump's AI lead, said heavier rules would hurt smaller firms. Volker Turk, the UN human rights chief, sent a separate letter on 14 September telling states and labs to write red lines in human-rights law rather than in company policy. Chen Yixin, China's minister of state security, has already framed generative models as a tool of rumour and cognitive warfare in an article in China Cyberspace. The same week holds a call to slow down, a call to speed up, and a call to treat the models as a political weapon.

Markets heard the first of those calls. Chip, cloud and data-centre names that have priced an uninterrupted buildout sold off. The essay does not cancel a single rack. It does force public companies to answer a question they have avoided on earnings calls: if the people who train the models say the pace is unsafe, what capex plan still holds.

Trump's June executive order asked labs to submit models for voluntary review before release. Amodei's evaluators go further. They would live inside the building. That is closer to an arms-control inspection than to a comment letter. OpenAI has worried about antitrust risk in any coordinated slowdown. Altman's Monday post waved that concern away without describing a legal structure.

None of the signatories announced a delay to a named training run. That gap is the story as much as the essay. A slowdown that cannot be seen on a cluster is a press event. A slowdown that can be seen is a change in the product calendar, in power-purchase agreements, and in the revenue models that now sit under a large share of the S&P 500.

Amodei told CBS over the weekend that the dangers are real. He still runs a company whose business is to ship more capable systems. The tension is a bid to set the terms on which everyone else has to ship. Trump's reply is that those terms will be set in the White House, not in San Francisco.

The next facts to watch are narrow. Does Anthropic actually badge independent evaluators this quarter. Do Altman and Musk attach a date to a training pause. Does the White House turn the June order into something with a penalty. Until one of those happens, the letter and the Truth Social post are positions, and the clusters keep running.