How do Large Language Models predict the next token when generating text? What role do tokenization, context, probability, and transformer architecture play in this process? How does the model use previous text to generate responses that match the conversation?