AI Models: Generating 50x More CO2 for Minimal Benefit

A recent study from German researchers highlights the substantial carbon footprint associated with using large language models (LLMs) in AI. The findings show that the CO2 emissions largely depend on the model’s reasoning approach; models that employ explicit reasoning processes emit significantly more carbon, sometimes up to 50 times more than concise response models. Evaluating 14 LLMs across a variety of subjects, researchers found that reasoning-based models produced an average of 543.5 tokens per question, leading to greater emissions. While the reasoning-enabled Cogito model achieved the highest accuracy, it still had three times the emissions of concise models. Moreover, complex subjects like abstract algebra caused emissions six times greater than simpler topics. The study urges users to prompt concise answers and limit high-capacity models to necessary tasks in order to reduce emissions. Researchers hope that increased awareness about AI’s environmental impact will guide more thoughtful usage of these technologies.

Source link