Claude "leads" more than a quarter of the AI research at Anthropic

AI developer Anthropic stated that, within the company itself, its AI model Claude “oversees” 26% of the research and engineering work aimed at creating the next generation of artificial intelligence. Back in March, this figure was estimated at less than 1%, the company noted. Anthropic presented the new data as part of its regular reporting, which it plans to publish to demonstrate the pace of its AI development process. The announcement appeared on the company’s blog.

Thus, the Claude chatbot can handle most of the task after receiving a general prompt from an employee, while the human’s role is limited to overseeing the process, the companies said. Anthropic also revealed that engineers used AI in more than 90% of the research projects conducted at the lab. However, the company emphasized that Claude did not operate completely autonomously in any of the tasks evaluated.

To assess AI’s contribution to corporate objectives, Anthropic uses a scale specifically developed by Epoch AI, an independent nonprofit organization that tracks technological developments. At level zero (AL0), there is a complete absence of AI involvement in the work, while at the highest level, level five (AL5), “AI operates completely autonomously, without human involvement.” Currently, the company mostly reports AI involvement at the intermediate level, AL3—that is, AI “collaborates” with humans and can perform large volumes of work under direct supervision.

The data released by Anthropic could help determine how close the AI industry is to “recursive self-improvement,” which the AI company describes as “a model capable of creating its own successor completely autonomously,” writes Business Insider. The release of this information comes amid growing concerns that AI could reach a point where it is capable of developing itself with minimal human involvement. Researchers warn that as AI agents become more autonomous, their behavior may diverge from their creators’ intentions and become more difficult to control and monitor.

Against this backdrop, Anthropic CEO Dario Amodei warned over the weekend that, given the accelerating pace of AI development—as early as 6–12 months from now—a swarm of agents, such as those that attacked the startup Hugging Face, “may be capable of taking over the entire internet,” potentially causing hundreds of billions of dollars in damage. In light of this, Amodei called for slowing the pace of AI development. He was supported by the head of OpenAI and Elon Musk, CEO of the aerospace and AI corporation. However, not everyone in the market took these warnings seriously; for example, well-known short seller Michael Burry stated that he considers the Anthropic CEO’s remarks to be self-promotion ahead of the IPO.

This article was AI-translated and verified by a human editor