Abstract:
According to Reuters, an independent study found that Anthropic’s AI models have improved at nearly twice the rate as AI plays an increasing role in developing more powerful subsequent models. The findings reinforce recent industry warnings that AI technology may advance faster than risk assessment efforts.
Anthropic
The findings, released Thursday by the AI safety nonprofit METR, provide one of the clearest evidence yet that AI is accelerating its own development. This trend could pave the way for recursive self-improvement (RSI), where AI systems develop more powerful versions of themselves without human intervention.
The 11-week evaluation, conducted from August to October this year, was based on confidential research, interviews and performance data from unreleased models. The results show that since October 2025, the Anthropic model has improved its capabilities in benchmark tests 1.9 times faster than before.
"As of October 2026, Anthropic is likely to be improving the capabilities of its models on well-defined tasks at nearly twice the previous rate. The previous comparison period was March 2024 to October 2025." METR said.
METR also added: "There is considerable uncertainty in this estimate. We would not be surprised if the actual acceleration is closer to 1.5 or 2.5 times historical levels."
The report also emphasizes that AI is currently unable to make a highly important and fully autonomous contribution to the most critical stage of model training.
Anthropic researchers have reported significant productivity gains from using AI to code and conduct experiments. Informal employee surveys suggest an increase of about 12 times, but METR cautions that these numbers are unreliable.
Anthropic CEO Dario Amodei called for a slowdown in the development of advanced AI in September. Currently, the company is preparing to go public and may launch an IPO as soon as this month. It is expected to become one of the largest IPOs in history and will also become an important test of investors' enthusiasm for investing in AI companies. These companies are investing billions of dollars in developing this technology.
In recent months, the potential for AI to recursively improve itself has prompted warnings from AI researchers. They worry that such systems could become difficult to align with human goals, difficult to monitor and control, and could ultimately pose an existential threat to humanity.
This shift could have far-reaching consequences for the entire industry. Previously, the AI industry has invested hundreds of billions of dollars in building computing infrastructure to train advanced models. In the future, AI itself may become an important force in promoting technological progress.
METR said: "We are convinced that this acceleration is not caused by an accelerated expansion of training computing power."
Comments