Balancing Environmental Impact and Accuracy in AI Models

A recent study from Hochschule München University of Applied Sciences highlights the significant environmental impact of large language models (LLMs), revealing that about 52% of American adults use them regularly. The researchers evaluated 14 LLMs, testing their responses to 1,000 “benchmark” questions and measuring the associated greenhouse gas emissions. They found that reasoning-enabled models produced up to 50 times more CO2 emissions compared to concise response models. Notably, while more advanced models like Deep Cogito 70B achieved higher accuracy (84.9%), they resulted in triple the emissions of simpler counterparts. Alibaba’s Qwen 7B model emerged as the most energy-efficient but had lower accuracy (31.9%). Findings suggest a trade-off between accuracy and sustainability; users can minimize emissions by opting for concise outputs or simpler models for applicable tasks. The study aims to encourage more thoughtful use of AI technologies to reduce their carbon footprint as they become integrated into daily life.

Source link