A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in making AI systems more energy-efficient. The research focuses on reducing the computational power required for large language models (LLMs) during their inference phase, which is when the trained model generates responses. By implementing a novel method called ‘early exiting,’ the system allows simpler queries to bypass deeper, more complex layers of the neural network. This selective processing can cut energy consumption by over two-thirds for certain tasks without a noticeable drop in output quality. The approach addresses growing concerns about the substantial environmental footprint of running powerful AI models, offering a pathway toward more sustainable AI deployment. Read the full article at: https://technologyreview.com/2024/05/15/1090000/mit-ai-energy-efficiency-early-exiting.