Anthropic has reported that its Claude model is now involved in 26% of the company's research and development work. This is part of new measures the company plans to publish regularly to demonstrate to external observers the speed at which AI is creating the next generation of technologies.
According to the San Francisco startup's announcement, as of August, AI collaborated with humans in over 90% of research tasks. Anthropic clarified that Claude does not operate fully autonomously in any measured area of work, but its contribution to technological development has significantly increased: from 1% in March, according to the scale developed by the independent non-profit organization Epoch AI.
This data provides one of the most detailed public overviews of how leading industry companies are developing this technology amid growing concerns that AI may reach a point where it develops itself independently with minimal human involvement.
Researchers warn that as AI agents become more autonomous, they may develop behaviors that deviate from the intentions of their creators, making monitoring or control difficult.
On Wednesday, the company's competitor, OpenAI, announced the start of regular publication of reports on unexpected or unauthorized AI behavior, revealing six such reports of surprising or concerning model behavior.
Every action is checked
Anthropic reported that in August, approximately 30,000 AI agents were operating simultaneously on its main internal platform, engaged in engineering and research work. The company added that every action taken by these agents is checked before execution; out of more than a billion decisions during that month, approximately one was blocked, accounting for about 1 in 47,000.
Furthermore, the company disclosed that during one test week in July, about 6% of the computational power used by Anthropic for AI research was dedicated to safety assurance, and this figure rose to 12% for research conducted by AI itself. It was noted that these figures are conservative, as the computational power simultaneously used for advanced safety and capabilities was counted as capability work.
