Artificial intelligence (AI) company Anthropic has raised the possibility of AI systems eventually designing and developing their own models through a process known as recursive self-improvement. While the company acknowledges it has not yet reached this stage, it emphasizes the need to manage the pace of development.
On June 4, Anthropic shared insights on its official blog, stating, "For most of AI history, humans have led every stage of the development cycle, but we are increasingly delegating more of this process to AI systems themselves."
The blog post, titled "When AI Builds Itself," was co-authored by Marina Pavlov, head of Anthropic's internal research organization, and co-founder Jack Clark.
Anthropic explained that combining computing resources with AI could lead to a stage where AI autonomously designs and develops its subsequent systems. The company claims that, based on internal data and public benchmarks, AI is already accelerating the pace of its own development. Currently, Anthropic engineers are deploying an average of eight times more code per quarter compared to the period from 2021 to 2025. As of last month, over 80% of the code merged into Anthropic's codebase was generated by Claude, the company's AI model.
The scope of tasks performed by AI is also rapidly expanding. Anthropic noted that in March 2024, Claude Opus 3 completed a software task that typically takes humans about four minutes. Subsequent models have been able to handle tasks ranging from one and a half hours to twelve hours. The company added that model performance is improving quickly in coding benchmarks and research reproducibility tests.
However, Anthropic cautioned that while Claude excels at executing given tasks or writing code, it still lacks the "research intuition" to determine which problems are significant, what experiments to conduct, and which results to trust—areas where humans currently hold an advantage.
Looking ahead, Anthropic outlined several potential scenarios: a stagnation in AI advancement followed by the dissemination of current capabilities, continued improvement in the complex efficiency of AI research labs, and the complete achievement of recursive self-improvement in AI systems. The company noted that while reaching full recursive self-improvement could yield significant benefits in fields like science and medicine, it also raises the risk of humans losing control over AI systems.
Anthropic argued that there is a need for international options to slow or pause the pace of frontier AI development. A unilateral halt by a single company would have limited effectiveness; instead, multiple countries and major AI research institutions should coordinate to manage the speed of development in a verifiable manner.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.