26 percent of Anthropic's AI development is now led by AI itself

Anthropic, the company behind the AI chatbot Claude, has for the first time published concrete figures on how much the AI itself contributes to developing future versions of itself. According to the new figures, Claude now "leads" 26 percent of the company's research and development, up from less than 1 percent in February this year.

Behind the figures are up to 30,000 so-called AI agents running simultaneously at Anthropic. An agent is an AI program that can carry out an entire task independently, step by step, from a high-level instruction, unlike an ordinary chatbot that you simply ask questions. Anthropic's employees use Claude in around 90 percent of their work.

What "leads" actually means

The word "leads" sounds more dramatic than it is, though. Anthropic has created its own scale for how independently the AI works. Claude "leading" a task means the AI can solve it from start to finish based on an instruction, while humans still oversee and make the final decisions. So it's not the same as the AI working entirely on its own.

Questions have also been raised about how precise the figures really are. Claude itself helps judge which level each piece of work should be placed at. When employees were asked to assess their own work, they agreed with each other only about a third of the time, and Claude's own assessments matched the employees' in only 59 percent of cases. This shows that the boundaries between the levels are hard to draw sharply.

Oversight of the many agents

To keep track of the many agents, Anthropic uses two forms of oversight: a system that can block actions in real time, and a subsequent review of what the agents have done. In August, the real-time system blocked about 0.002 percent of more than a billion decisions, while the subsequent review covered all of the agents' actions.

Anthropic itself points out that it is becoming harder for humans to understand and control the language models as the AI takes on a larger share of its own development. That's one of the reasons the company is now publishing the figures, so the development can be followed from the outside.

For ordinary Danes and Danish workplaces, the story is a sign of how fast AI development is moving right now. Even though Claude is far from working entirely on its own, the figures show that AI is increasingly involved in the development of new technology itself, not just as a tool you use in everyday life. It's worth keeping an eye on, both because it could mean faster progress in the AI tools you use at work, and because it raises questions about how well humans can keep up and maintain oversight.