Anthropic announced that its AI model, Claude, is playing an increasing role in developing its next generation; a trend that raises new questions about the industry's move toward recursive self-improvement capabilities.
According to ITNA, Anthropic said on Thursday in a statement that Claude now directs 26 percent of the company's research and development activities for its models.
Anthropic said this means Claude can carry out most of a given task solely from a high-level instruction, “end-to-end” and under human supervision; although the model is not yet fully autonomous.
According to the report, about 90 percent of the company’s R&D activities are also conducted in “collaboration” with Claude; a situation that Anthropic says means the model can perform large portions of the work but under close human guidance.
The growing role of Claude
A notable point in the report is the rapid increase in Claude’s share of research and development. Anthropic stated that in February Claude’s share of independently leading projects was zero, but just six months later, in August, that figure rose to 26 percent.
The company also said that by August about 30,000 AI agents were engaged in research and engineering work.
Safety concern context
This announcement came as some prominent figures in the AI industry, including Dario Amodei, Anthropic’s CEO, have called for slowing the pace of development due to safety concerns.
Anthropic emphasized in its statement that it must do everything it can to reduce the gap between leading AI labs’ knowledge and public awareness.
However, the statement does not make clear how close Anthropic believes it is to achieving “recursive self-improvement” (that is, an AI model’s ability to independently improve its future versions without human intervention). Still, the speed at which Claude’s role in research processes has increased has drawn the attention of observers in the field.
According to NBC News, the company also released details of its oversight measures for agents to show how precisely monitoring systems can detect potential agent misbehavior.