This chapter offers a foundational overview of large language models (LLMs), focusing on their construction, scaling laws, emergent abilities, and the technical evolution of the GPT series models. We begin by outlining the two-stage training process of LLMs: pre-training, which establishes the model’s foundational capabilities through unsupervised learning on vast text corpora, and post-training, which fine-tunes the model using instruction tuning and human alignment techniques to enhance task-specific performance and ensure alignment with human values. The chapter then explores scaling laws, which quantify the relationship between model performance and factors including parameter size, data scale, and computational resources. Two key scaling laws are discussed: the KM scaling law and the Chinchilla scaling law. These laws provide a framework for understanding how scaling impacts model capabilities and guide the development of more efficient and effective LLMs. Subsequently, we examine emergent abilities such as in-context learning, instruction following, and step-by-step reasoning. The chapter discusses the relationship between emergent abilities and scaling laws, noting that while scaling laws predict smooth performance improvements, emergent abilities often manifest as sudden leaps in task performance, making them harder to anticipate. Finally, the chapter traces the technical evolution of the GPT series models, from the initial development of GPT-1 and GPT-2 to the recent reasoning models o1 and o3. This section provides insights into the milestones and innovations that have shaped the development of these influential models.