A quantum trick helps trim bloated AI models

Physicists have developed a technique using tensor networks—mathematical structures originally created to describe quantum particle interactions—to dramatically compress bloated AI models without sacrificing accuracy. The method can reduce large language models like Llama 2 by over 90 percent in memory usage while cutting energy consumption by 30-40 percent, potentially allowing AI to run on smartphones without internet connections. Researchers are also exploring tensor networks as a complete alternative to neural networks, which could eliminate energy-intensive training processes and create more understandable AI systems.