Anthropic model development speeds up two-fold despite calls to slow down

A tension exists between Anthropic leadership and the pace of its technological development. CEO Dario Amodei has publicly urged the industry to reduce the speed of AI advancement to allow safety measures to catch up. However, data released Thursday by the independent monitoring group Model Evaluation and Threat Research, or METR, indicates that Anthropic is developing its systems roughly twice as quickly as it was a year ago.

METR analyzed internal data provided by Anthropic over an 11-week period spanning from mid-August to late October. The nonprofit, based in Berkeley, California, also interviewed company staff during this time. The group stated that it was given access to internal data that it described as unprecedented for an external review. This access allowed researchers to track how well the company's models perform on specific, well-defined tasks.

The findings suggest that AI itself is a primary driver of this acceleration. METR confidently ruled out increased computing power as the cause, a factor that traditionally correlates with model performance. Instead, the researchers attributed the speedup to the use of AI to improve AI. Anthropic has noted that its Claude chatbot increasingly assists in developing newer, more advanced versions of the system.

Amodei called for a slowdown in a September 12 essay. In that piece, he urged the AI world to slow the pace at which they improve model capabilities. He argued that advances have been driven by recursive self-improvement, which he defined as AI's growing ability to build the next generation of AI. At the same time, he championed the idea of having external evaluators embedded within leading labs, specifically naming METR as an appropriate third-party partner.

Sam McCandlish, an Anthropic co-founder and responsible scaling officer, responded to the report in a post on X. He stated that many of METR's conclusions align with his own observations. McCandlish wrote that while AI is speeding up research, it is difficult to say exactly how much. He acknowledged that the company's redactions to its public risk report make it harder for others to check the work, but he argued that sharing exactly what drives the speedup is commercially sensitive and risks providing a roadmap for competitors.

Anthropic is preparing to go public later this month in what could be the largest initial public offering in history. The company plans to list on Wall Street as debate continues over the potential for AI models to elude human control. METR said its assessment focused on whether the growth rate of internal capabilities has accelerated and how much of that change is attributable to AI assistance.

The nonprofit said its public report overall understated the level of acceleration occurring within the company. METR noted that Anthropic's redactions limited how the group could communicate detailed estimates. The researchers found no bottlenecks, such as restrictions on computing power or data, holding back progress. They warned that using AI to improve itself could start to compound the acceleration, although the report stated that AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the training process.