We’ve seen how large language models (LLMs) are pretrained on vast amounts of data and how they generate text one token at a time during inference.