METR study says Anthropic doubles model improvement rate

The author argues that current AI policy discussions conflate "weak recursion"—the historical norm of using a technology to build better versions of itself—with "strong recursion," the potential for fully autonomous AI self-improvement. While weak recursion is demonstrably occurring at frontier labs like Anthropic, the author contends that evidence for strong recursion remains thin and lacks genuine "taste." Consequently, poorly calibrated policies or panic over weak recursion risk banning vital safety and capability advancements. Open article →