Reproduce “Scaling Laws for Economic Productivity: Experimental Evidence in LLM-Assisted Translation” on OpenResearch

About this paper

Ali Merali

An empirical study quantifies economic scaling laws for LLM-assisted translation, showing that a tenfold increase in model training compute correlates with a 12.3% reduction in task time, a 0.25 point quality improvement, and a 16.1% increase in earnings per minute, with disproportionate benefits for lower-skilled workers. The findings suggest an estimated 6.95% aggregate U.S. productivity growth over the next decade from LLMs.

Key insights

• "Economic scaling laws" exist, demonstrating a continuous and positive relationship between LLM training compute and human economic outcomes like task speed, quality, and earnings per minute.
• Advanced LLMs offer disproportionately greater productivity benefits to lower-skilled workers, suggesting a potential mitigating effect on skill-based wage inequality.
• The rapid scaling of LLM capabilities, coupled with low and decreasing inference costs, implies a higher potential for aggregate productivity growth from AI than previously estimated.

Summary by alphaXiv.