EFFICIENT LLMS WITHOUT MATRIX MULTIPLICATION

Summary from the TLDR AI newsletter

Researchers have found a way to eliminate matrix multiplication (MatMul) from large language models while maintaining strong performance, even at the billion-parameter scale.