Anthropic AI development speed has doubled, researchers find
An independent review suggests Anthropic is developing AI systems twice as fast as last year, despite the CEO's public calls to slow the industry's pace.
A report from the research nonprofit Model Evaluation and Threat Research (METR) indicates that Anthropic has doubled the speed of its AI model development compared to the previous year. The findings arrive as the company prepares for an initial public offering amid growing industry debate regarding AI safety and the control of autonomous systems.
Researchers at METR, who were granted access to Anthropic's internal data and staff, concluded that the primary driver of this acceleration is the use of existing AI models to assist in building newer, more capable versions. This process, often described as recursive self-improvement, has allowed the firm to advance faster than conventional computing power upgrades would typically allow.
The findings challenge the narrative presented by Anthropic’s leadership. While CEO Dario Amodei has publicly advocated for slowing the pace of capability improvements to ensure safety protocols can keep up, the METR report suggests that the company’s internal progress has significantly exceeded its own public risk disclosures.
As Anthropic moves toward a major Wall Street debut, the report highlights the technical concerns regarding the lack of bottlenecks in AI development. METR warned that the compounding effect of AI-assisted development could outpace human oversight, noting that Anthropic researchers have observed significant progress in the machines' ability to contribute to their own training processes.