Physicists are repurposing a mathematical tool called a tensor network—originally designed to describe electron interactions—to solve the "curse of dimensionality" in AI. As models grow increasingly bloated, these networks decompose massive tensors (multi-dimensional data arrays) into smaller, manageable parts. Unlike traditional methods like pruning or quantization, tensor networks provide a mathematically grounded approach to eliminating redundancy while preserving model accuracy.

Results from startup Multiverse Computing are striking: their "CompactifAI" technique shrank Llama 2 7B by over 90%, cutting memory from 27GB to just 2GB. This compression allows Large Language Models (LLMs) to run locally on smartphones or appliances without an internet connection. Furthermore, energy efficiency gains are significant, with Llama 3.1 models consuming up to 40% less power during operation.

Researchers are now advocating for AI built on tensor networks from the ground up to bypass neural networks entirely. By avoiding the slow, energy-sapping optimization process of gradient descent, these models can be trained in seconds. One demonstration showed a tensor network model training 100 times faster than its neural equivalent, hinting at a future of transparent "white-box" AI that is both easier to understand and cheaper to build.

Physicists are repurposing a mathematical tool called a tensor network—originally designed to describe electron interactions—to solve the "curse of dimensionality" in AI. As models grow increasingly bloated, these networks decompose massive tensors (multi-dimensional data arrays) into smaller, manageable parts. Unlike traditional methods like pruning or quantization, tensor networks provide a mathematically grounded approach to eliminating redundancy while preserving model accuracy.

Results from startup Multiverse Computing are striking: their "CompactifAI" technique shrank Llama 2 7B by over 90%, cutting memory from 27GB to just 2GB. This compression allows Large Language Models (LLMs) to run locally on smartphones or appliances without an internet connection. Furthermore, energy efficiency gains are significant, with Llama 3.1 models consuming up to 40% less power during operation.

Researchers are now advocating for AI built on tensor networks from the ground up to bypass neural networks entirely. By avoiding the slow, energy-sapping optimization process of gradient descent, these models can be trained in seconds. One demonstration showed a tensor network model training 100 times faster than its neural equivalent, hinting at a future of transparent "white-box" AI that is both easier to understand and cheaper to build.