Why LLMs famously miscounted the r’s in strawberry, what tokenization hides from the model, and why newer models get it right.