AI Pace: Anthropic development speed doubles
Independent auditors found the San Francisco firm is improving model capabilities at twice the rate of a year ago, citing recursive self-improvement.
By Sarah Reynolds ·
Anthropic is advancing its artificial intelligence models at a rate approximately twice as fast as it was one year ago, according to a report released Thursday by the independent nonprofit Model Evaluation and Threat Research.
The finding contradicts public statements by the company’s leadership. In an essay published on Sept. 12, CEO Dario Amodei urged the industry to reduce the speed of innovation. He argued that the pace of capability growth must slow to allow safety measures to mature.
METR conducted its assessment to provide an external check on the technology, as details regarding AI risks have historically come primarily from the companies building them. The group was granted access to Anthropic’s internal data and interviewed staff members. This review period coincided with Anthropic’s preparation for an initial public offering scheduled for later this month, which could become the largest stock market debut in history.
The auditors determined that the acceleration is driven largely by AI systems helping to build newer versions of themselves. METR ruled out increases in computing power as the primary cause, a factor that traditionally dictates model performance. Instead, the researchers attributed the speedup to the use of AI in the development process.
The report includes caveats regarding the scope of the acceleration. METR noted that the doubling of speed applies to well-specified tasks that can be graded automatically. It is unclear if this trend extends to messier, less defined tasks. The group also acknowledged internal disagreements and uncertainty about whether Anthropic understated its previous progress, making the current figures appear higher.
A central concern identified in the analysis is the lack of bottlenecks. METR found no restrictions on compute or data holding back the work. This allows the recursive self-improvement cycle to continue without friction. The report states that AI is not yet able to make fully autonomous contributions to the most critical parts of model training, but researchers observed signs that this capability is approaching.
Amodei described this recursive process as the primary driver of recent advances. He warned that if left unchecked, the technology could outpace human control. The CEO previously supported the concept of independent evaluators with employee-like access, a role filled by METR for this specific analysis. OpenAI CEO Sam Altman has also expressed support for such third-party oversight.
