Anthropic's chief urges the industry to deliberately slow the growth of SI capabilities; Altman, Musk and Hassabis agree the same day
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
The leaders of all the major labs have agreed that the race needs to slow down and outside checkers should be let inside: the growth of SI capabilities is, for the first time, publicly called a problem rather than a victory. The key is that the "pace" does not become a brake that pushes breakthroughs to later.
In the essay "We Must Pace the Frontier," Dario Amodei proposed slowing the pace at which labs increase model capabilities, not by halting training but by leaving time for alignment and protection. The plan has three parts: external evaluators such as METR get permanent "employee-level" access (workspaces, badges, laptops, the right to publish findings without Anthropic's editing), which Anthropic is introducing unilaterally; common safety standards for labs in democratic countries, brokered by the US government because of antitrust risks; and international agreements, including narrow limits together with China. Amodei estimated that in 6 to 12 months swarms of misaligned agents could theoretically take over internet infrastructure, with damage in the hundreds of billions of dollars. The same day, Sam Altman wrote that he agrees and OpenAI will introduce the same embedded evaluators, Elon Musk replied "Dario is right," Demis Hassabis backed the direction but proposed an industry-wide standards body instead of separate agreements, and Hugging Face launched the Open Alignment Initiative and asked to join the evaluator program.