Claims about AI productivity improvements (marginalrevolution.com)

A recent study led by Ali Merali from Yale University reveals significant insights into the productivity improvements linked to Large Language Models (LLMs). The paper, titled “Scaling Laws for Economic Impacts,” presents empirical data indicating that each subsequent year of LLM advancement could reduce task completion times by an average of 8%. Notably, the research involved over 500 professionals across various fields who performed tasks using one of 13 different LLMs, highlighting that approximately 56% of productivity gains stemmed from enhanced computational resources, while 44% originated from advancements in algorithms. This research is pivotal for the AI/ML community as it quantifies the relationship between LLM scaling and tangible economic benefits. The findings suggest that as LLMs continue to improve, they could potentially increase U.S. productivity by around 20% in the next decade, particularly in non-agentic analytical tasks. However, the study notes that gains were less pronounced in agentic workflows that require tool use, indicating a nuanced impact of AI on different types of professional tasks.