METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July

Anthropic researchers have identified that current large language models are nearing the threshold of Recursive Self-Improvement, a state where software could theoretically modify its own code to become more capable. While the company reports that current systems do not yet possess this autonomous ability, the findings highlight an increasing need for safety protocols to prevent potential runaway development. This assessment serves as a baseline for monitoring how future model iterations handle complex programming and iterative task optimization.