Anthropic, an American generative artificial intelligence (AI) development company, has urged global AI research laboratories to consider decelerating the pace of AI development. This call stems from a pressing concern that AI systems are advancing too rapidly, potentially reaching a state of "recursive self-improvement" (재귀적 자기 개선), where AI can independently design and develop its own models without human intervention.

In a blog post dated June 4th, authored by Marina Favaro, head of the Anthropic Institute, and co-founder Jack Clark, Anthropic presented internal data illustrating the rapid growth of AI models, positing that a global slowdown in AI development would be beneficial. They elaborated on the concept of recursive self-improvement (RSI), defining it as the hypothetical capacity of an AI to autonomously architect and implement advancements to its own codebase, algorithms, and even its foundational models, thereby initiating a self-sustaining cycle of exponential capability enhancement. While Anthropic clarifies that current AI has not achieved RSI, and its attainment is not an inevitability, they caution that it could materialize sooner than most institutions are prepared for. Such a development, while potentially yielding "immense benefits" across scientific and medical domains, concurrently elevates the risk of "humans losing control over AI systems."

To mitigate this burgeoning risk and afford humanity sufficient time to adapt to the profound implications of RSI, Anthropic proposes a coordinated slowdown in technical development. The core methodology for achieving this involves a multilateral agreement among AI research entities: specifically, multiple research laboratories would collectively commit to pausing or significantly decelerating their development efforts under mutually agreed-upon conditions. Crucially, this framework necessitates the implementation of robust, reciprocal verification mechanisms, enabling each participating laboratory to independently confirm the adherence of others to the agreed development moratorium.

Anthropic posits that such a collaborative treaty is feasible, drawing an analogy to the 1987 Intermediate-Range Nuclear Forces (INF) Treaty between the United States and the Soviet Union. This historical agreement prohibited the production, testing, and deployment of ground-launched ballistic and cruise missiles with ranges between 500 and 5,000 kilometers, symbolizing an era of arms control despite geopolitical tensions. However, Anthropic acknowledges that while historical treaties like INF required decades to build the requisite infrastructure and trust, the rapid pace of AI evolution means "we don't have that time." It is also noted that the INF Treaty was abrogated in 2019. Looking forward, Anthropic intends to establish an organization that will engage with policymakers and civil society stakeholders to foster discussions and coordinate efforts pertaining to full recursive self-improvement management.