ivo's home

Researchers at the University of Michigan found that generative AI models use significantly more energy to process non-English text, with disparities ranging from 1.4 to 180 times that of English. Languages with non-Latin scripts, such as Hindi and Bengali, require up to 15 times more energy due to higher token counts. This inefficiency also correlates with lower accuracy in model outputs.

The finding highlights a structural bias where the majority of the global population faces higher costs and reduced reliability when using AI in their native tongues. While the study measures three specific open-source models, the issue persists across different hardware and scales. This suggests that current infrastructure actively disadvantages speakers of underrepresented languages.

more: https://www.futurity.org/generative-ai-energy-non-english-text-3352352/?utm_source=rss&utm_medium=rss&utm_campaign=generative-ai-energy-non-english-text-3352352

my other side: ivo.blogspot.com