New techniques efficiently accelerate sparse tensors for massive AI models
https://techxplore.com/news/2023-10-techniques-efficiently-sparse-tensors-massive.html "improved performance and energy-efficiency of massive machine-learning models that drive generative AI... skips over 0's, saves computation and memory... finds non-0's for wider variety of sparsity patterns, and handle case where data doesn't fit memory... still maintain flexibility and adaptability while being specialized and efficient... divides tensor values into smaller blocks, with own simple, sparsity pattern, then combining blocks into eve larger levels, remaining simple at each step" Related: Tiny AI model beats massive LLMs at logic test https://www.nature.com/articles/d41586-025-03379-9 How a simple game of 20 questions could help make AI fit for the future https://phys.org/news/2026-09-simple-game-ai-future.html