New transformer architecture can make language models faster and resource-efficient

New technique can accelerate language models by 300x

Language models can use steganography to hide their reasoning, study finds

Decomposing Language Models Into Understandable Components