AI chiefs endorse slower frontier development over safety

Three prominent figures in the artificial intelligence industry - Dario Amodei of Anthropic, Sam Altman of OpenAI, and Elon Musk - have joined forces to advocate for a more cautious approach to the rapid development of frontier AI technology. In an essay published on Saturday, Amodei argued that AI companies should deliberately "pace the frontier" of AI development, rather than maximizing capability gains at the expense of safety measures.

He emphasized that progress could continue at a rapid pace, but developers should prioritize alignment research, monitoring, security, and regulation to ensure the systems remain under human control. Altman concurred with Amodei's stance and committed OpenAI to adopting independent evaluators with employee-like access, in line with Amodei's framework.

Musk, who has previously expressed concerns about the risks associated with advanced AI, succinctly endorsed Amodei's viewpoint by stating, "Dario is right." While commercial pressures have made unilateral slowdowns challenging, Amodei proposed three measures to address the issue. First, frontier laboratories should grant outside evaluators the necessary access to assess safety practices, investigate incidents, and examine model behavior.

Second, companies operating in democratic countries should collaborate on common safety standards. Finally, governments should pursue international agreements on areas where shared interests are strongest. Anthropic has pledged to implement the independent-evaluator commitment, even without matching action from competitors. Amodei framed this step as a way to make safety claims more verifiable and to reduce reliance on companies evaluating their own technology.

The push for slower AI development comes in the wake of evidence that AI systems are already accelerating certain aspects of AI research. OpenAI's agents can now perform research tasks that would typically take skilled human researchers several days. However, OpenAI has stressed that fully autonomous recursive self-improvement, where AI systems contribute materially to building stronger generations of AI, has not yet occurred.

OpenAI has also acknowledged that safety considerations are beginning to influence the pace of model development, citing temporary slowdowns and enhanced protections after determining that a model reached its highest cybersecurity capability threshold. The warnings surrounding AI development have grown more urgent as AI tools become increasingly useful to researchers themselves.

OpenAI's chief scientist, Jakub Pachocki, recently noted that gains in machine intelligence could extend to recursive self-improvement, emphasizing the need for broader interventions alongside alignment work. While Amodei's proposal does not constitute a moratorium on AI research or a fixed cap on computing power, it aims to link the speed of capability advances more closely to measurable progress in safety and oversight, thereby reducing incentives for individual laboratories to race ahead at the expense of their competitors.