Dynamic Context Pruning for Efficient and Interpretable Autoregressive Transformers | Read Paper on Bytez