The AI lab Anthropic has announced significant advancements in their research and development efforts, particularly highlighting the contributions of its AI model, Claude. This update was shared on Thursday, detailing Claude’s role in the creation of a more advanced version of the system.
Currently, Claude is responsible for leading 26% of the development initiatives at Anthropic, indicating that it can handle a majority of tasks from start to finish based on higher-level prompts, albeit under human oversight. The AI is not yet capable of full autonomy at this stage.
Nevertheless, approximately 90% of Anthropic’s R&D is conducted in conjunction with Claude, which the company states allows the model to perform significant portions of work while still receiving direct human guidance.
This announcement coincides with growing concerns about AI safety, as prominent figures in the industry, including Dario Amodei, CEO of Anthropic, are advocating for a measured approach to AI development.
In a recent blog post, the company emphasized the need for transparency in AI advancements, stating that it is essential to bridge the gap between the knowledge of AI research labs and the public. They suggested that improving the measurement and reporting of AI development can empower society to better understand and engage with these technologies.
Anthropic expressed concern that unchecked advancements in self-improving AI models might pose challenges for human oversight and comprehension. They argued that by sharing development metrics publicly, there can be greater awareness of how close AI laboratories are to achieving recursive self-improvement—where an AI can independently create a more advanced version of itself.
The company has also encouraged fellow AI developers to regularly disclose similar metrics using a standardized methodology, allowing for better comparisons over time and across different labs.
While it remains unclear how close Anthropic is to reaching the milestone of recursive self-improvement, the progress displayed by Claude is noteworthy. In February, Claude did not lead any research or development efforts; within just six months, however, the model had advanced to directing a quarter of the work by August.
Additionally, Anthropic provided insights into the oversight systems they have implemented, revealing that nearly 30,000 agents were engaged in research and engineering tasks as of August. This oversight is vital for monitoring agent behavior and detecting any instances of misconduct. The company has also committed to integrating external third-party evaluators to enhance their safety monitoring efforts.
The conversation around AI safety gained further traction recently when an Anthropic researcher resigned, issuing a stern warning regarding the potential dangers posed by AI. In response, leaders in the tech industry, including Amodei, OpenAI’s CEO Sam Altman, and Elon Musk, have supported calls for a slowdown in AI development. This sentiment stands in contrast to perspectives from other tech figures and former President Donald Trump, who have advocated against such a pause.
