A new study from MIT and Google researchers demonstrates a significant advancement in making AI systems more energy-efficient. The research focuses on a technique called ‘inference skipping,’ which allows large language models to dynamically bypass unnecessary calculations for simpler queries. This approach could reduce the computational cost—and therefore the energy consumption—of running models like GPT-4 by up to 50% for certain tasks without sacrificing accuracy. The method works by having the model assess the difficulty of an input and selectively activating only the necessary parts of its neural network. While promising for reducing the environmental impact of AI, the technique currently requires specialized training and is not yet ready for widespread deployment. Read the full article at https://technologyreview.com/2024/05/15/ai-energy-efficiency-breakthrough.