Anthropic CEO Dario Amodei published an essay on his personal website, We Must Pace the Frontier, arguing that the AI industry needs to deliberately slow advances in model capabilities. The essay says progress would still feel fast and calls for using the additional time wisely. It is a notable position from the leader of a company at the forefront of the race to build more capable models.
A three-stage plan
The essay sets out three stages with increasing levels of coordination. The first, which Anthropic says it is adopting on its own, gives independent external evaluators ongoing access comparable to that of an employee. The text cites the safety evaluation organization METR as a reference. Evaluators would check compliance with safety practices and commitments, report incidents, and assess both finished models and training pipelines. Amodei compares this approach to regulatory supervisors working within banks. The second stage calls for coordination among AI companies; the third requires global coordination, including governments.
The context behind the essay
The essay appeared in the same week as the resignation of an Anthropic researcher who raised concerns about potentially lethal AI risks by the end of the decade and criticized the industry’s race toward self-improving superintelligence. In a CNN interview, Amodei said he agreed more than he disagreed with the former employee.
The essay discusses the risk of swarms of autonomous AI agents compromising critical internet infrastructure within six to twelve months, potentially causing hundreds of billions of dollars in damage. These are risk scenarios discussed in the source, rather than established predictions.
Why this matters for people building with AI
For engineering and product teams relying on frontier models, the debate about capability growth has practical implications. The laboratories supplying AI infrastructure for thousands of applications are publicly discussing whether and how to slow development.
Anthropic’s announced commitment to give external evaluators ongoing access to training pipelines, as well as finished models, also sets a transparency precedent that other companies may face pressure to follow. Governments could eventually turn such auditing into regulatory requirements. The topic connects with Anthropic’s threat intelligence report, covered in an earlier story.