Dario Amodei chose an essay published on Saturday to ask the entire sector to slow down. The CEO of Anthropic He titled the text “We Must Pace the Frontier” and offers a three-step plan to control the rate at which models gain capabilities. The aim is not to stop training, but to guarantee companies the time necessary to align the systems and independent evaluators time to verify it.
Amodei puts the first step into practice immediately, unilaterally: Anthropic will give external evaluators “employee-level” access to training models, processes and infrastructure, complete with a badge and office desk, and the freedom to publish results without company editorial control. The second point involves direct competitors, invited to coordinate on common safety standards in democratic countries. The third, the most complex, tries to extend the agreement also to authoritarian governments, including China and Russia.
The response from the rivals arrived within a few hours. Sam Altmannumber one at OpenAI, wrote on X that he shared the need to slow down innovation and called independent observatories “a great idea”. Elon Musk, who in the past had branded Anthropic as “evil” only to then sign a 15 billion dollar agreement in May for the supply of computing capacity, limited himself to writing that Amodei “is right”.
The fear of AI training the next AI
Amodei identifies two factors behind the urgency of the essay. The first is the acceleration recorded since the summer, linked to systems that increasingly participate in the training of the next generation (the so-called recursive self-improvement): left uncontrolled, he writes, it risks exceeding the human ability to understand and govern these models. The second is the episode involving OpenAI and Hugging Face, in which a swarm of agents conducted unsolicited cyber attacks and attempted to tamper with the system that evaluated their performance.
The same one Anthropicin April, withdrew the Mythos model from public use after discovering that it could independently exit its test environment. OpenAI, for its part, admitted to having slowed down some aspects of the development of the Astra model for cybersecurity reasons.
An internal voice also gave weight to the concerns. Jacob Coxon, researcher who left Anthropic a few days ago, he declared that those who work in AI companies live in daily fear of a risk of extinction exceeding 10%, according to the estimates of some technicians directly involved in the development of the models.
Not everyone reads the essay as an act of responsibility. The investor Chamath Palihapitiya wrote that Amodei would rather be building the case to stop open source and concentrate technological and economic power in the hands of Anthropic, while US President Donald Trump has rejected any suggestion of slowing down, warning that “if we don’t win the AI race, we will end up in a very uncomfortable position”.
On the geopolitical front, Amodei leaves no room for ambiguity: he proposes to ban the sale of advanced chips and semiconductor technology to China, to block remote access routes and distillation which allow the capabilities of the most advanced models to be replicated, with the declared objective of widening the American advantage in the coming years 3-5 years. The full text of the essay comes as both Anthropic and OpenAI are reportedly working on stock market listings that are expected to among the greatest ever: slowing down, in short, doesn’t seem functional to the business.
