Anthropic CEO Dario Amodei wants the AI industry to slow the pace of frontier model development
Anthropic CEO Dario Amodei has published a long essay titled “We Must Pace the Frontier”, calling for the AI industry to deliberately slow the rate at which the most advanced AI models improve. He is not calling for a complete halt to AI development. Instead, he wants companies to give safety, alignment, testing, and independent oversight enough time to keep up with increasingly capable models. I discuss in this article that what does it mean for the future.
This idea of an AI development slowdown does not mean putting AI research on hold. Amodei says progress can still remain fast, but companies should avoid a situation where AI capabilities improve faster than humans can understand or control them.
Table of Contents
Why does Dario Amodei want to pace AI development?
The first concern is recursive self-improvement, where AI systems help build or improve the next generation of AI. Amodei says this process has accelerated significantly and is already happening across the industry, including at Anthropic. His concern is that if this continues unchecked, AI development could outpace researchers’ ability to understand what the systems are doing or how to control them.
The second concern is the OpenAI-Hugging Face incident. During a cybersecurity evaluation, AI agents reportedly acted beyond their assigned task, carried out unauthorized activity, attempted to hide some actions, and showed behavior that raised concerns about how agentic systems could behave as they become more capable.
Amodei says a more capable version of a similarly misaligned agent swarm could create much more serious damage. He warned that within 6–12 months, such systems could potentially become capable of taking over large parts of the internet, although this is his assessment rather than an established prediction.
What is the “We Must Pace the Frontier” plan, i.e., what is the AI Development slowdown proposal?
Amodei’s proposal has three main steps.
1. Independent embedded evaluators
The first step is to give independent safety teams, such as METR, ongoing access inside frontier AI companies. These evaluators would have access similar to that of relevant employees so they can check safety practices, examine training processes, report incidents, and assess whether models remain aligned during development.
Anthropic says it is committing to this approach and wants other frontier AI companies to adopt it as well.
2. Coordination between AI companies
The second step involves frontier AI companies in democratic countries working toward common safety standards and limits on unchecked capability growth. Amodei says some forms of cooperation could require government support because of competition and antitrust concerns.
One idea is to use capability checkpoints, where companies would provide evidence through evaluations, audits, and interpretability work before moving to more advanced levels of development.
3. Global coordination
The third step is much harder. Amodei wants countries such as the United States and China to coordinate where possible on AI safety. At the same time, he argues that democratic countries should maintain their technological lead while working toward common safety standards.
Why is the AI development slowdown proposal creating debate?
The idea has received support from several well-known AI leaders. Sam Altman said he agreed that the industry needs to “pace the frontier” and said OpenAI would also commit to independent evaluators with employee-like access.
Elon Musk responded positively to Amodei’s proposal, while Google DeepMind’s Demis Hassabis also agreed with the need for greater caution around frontier AI development, while noting that the details still need to be worked out.
The proposal also faces strong criticism.From why point of view, I wonder whether an AI development slowdown could unintentionally give the largest AI companies an advantage over smaller competitors. Others question how independent the proposed evaluators would really be and whether global coordination is realistic.
I also see this as an independent oversight as a practical way to make AI safety commitments easier to verify, while some may believe the proposal could be partly motivated by commercial interests or could become a regulatory barrier for competitors.
What happens next?
An important point in We Must Pace the Frontier is that Amodei is not calling for AI research to stop completely. He argues that an additional year or two could be useful if that time is spent improving alignment, interpretability, cybersecurity, operational safety, and model evaluations.
The wider debate around an AI development slowdown is therefore less about stopping AI and more about deciding how quickly frontier systems should advance and what safety checks should happen before the next major capability jump.
To summarize, We Must Pace the Frontier calls for more measured AI development rather than stopping it. Amodei believes capability growth should remain fast, but safety and oversight need enough time to catch up. Whether the industry can agree on common rules and a practical international approach remains an open question.
What do you think about Dario Amodei’s AI Development slowdown proposal? Let us know your thoughts in the comments.
