The Anthropic logo. [Reuters]
The Anthropic logo. [Reuters]

Anthropic, the company behind the Claude AI model, has urged the world to consider slowing the pace of AI development globally.

The call reflects growing concern that AI systems are advancing faster than expected and may soon reach a stage where they can improve their own performance without human intervention.

In a blog post published Wednesday, Anthropic released internal data showing the rate of progress among its most advanced AI models, saying it would be "highly desirable" for the world to develop the capacity to regulate the speed of AI development.

Anthropic's head of research, Marina Favaro, and co-founder Jack Clark co-wrote the post.

The two said recent trends in AI development appear to be moving toward a stage known as "recursive self-improvement."

Recursive self-improvement refers to a process in which AI systems enhance themselves without human input, in turn producing even more capable AI.

Some experts warn that if this stage becomes a reality, it could trigger unpredictable risks and widespread social disruption.

"We believe it is desirable for the world to have the option to slow down or pause frontier AI development so that societal institutions and AI alignment research can keep pace with technological progress," Favaro and Clark wrote.

They proposed international agreement on how to slow AI model development and called for mechanisms to verify that competitors comply with any such arrangement.

While recursive self-improvement has not yet occurred and is not inevitable, they warned it "could arrive sooner than most institutions are prepared for."

Clark has previously spoken often about the potential risks of AI and the implications of AI models reaching the recursive self-improvement stage, saying such a development would transform the world.

"That kind of technology has never existed before, but I believe it could happen within the next two years — maybe sooner," he said.

Anthropic said it plans to continue discussions with policymakers, researchers and other stakeholders on recursive self-improvement and verification systems in the months ahead.


mokiya@heraldcorp.com