Scan and Snap: Understanding Training Dynamics and Token Composition in 1-layer Transformer | Read Paper on Bytez