LLMs.

Large language models (LLMs) are machine learning models designed to process and generate natural language. Most modern LLMs are based on the Transformer architecture, introduced in the 2017 research paper "Attention Is All You Need." Transformers use an attention mechanism to examine relationships among tokens—the words, word fragments, or symbols that make up an input—rather than processing text only from left to right. This helps the model connect information across long passages. During pretraining, an LLM is exposed to enormous collections of text and usually learns by predicting missing or subsequent tokens. For example, given the sequence "The capital of France is," the model may assign a high probability to "Paris." Repeating this task across billions or trillions of examples enables the model to learn grammar, writing styles, factual associations, reasoning patterns, and some programming structures. The model does not normally store knowledge as a conventional database; instead, information is represented in numerical parameters adjusted during training. After pretraining, an LLM may undergo additional processes such as supervised fine-tuning, instruction tuning, and preference optimization. These stages help it follow user requests, produce safer responses, and communicate more effectively. At use time, the model generates text one token at a time by selecting from probability distributions. Sampling settings such as temperature can make output more predictable or more varied. LLMs support applications including conversational assistants, search interfaces, document analysis, translation, tutoring, customer service, content drafting, software development, and information extraction. They are important because a single general-purpose model can perform many language tasks without being separately programmed for each one. However, an LLM does not possess human consciousness or guaranteed understanding. It can produce fluent but incorrect statements, a problem commonly called hallucination. Its outputs can also reflect biases in training data, expose sensitive information if used carelessly, or be manipulated by adversarial prompts. LLM responses should therefore be evaluated, especially in legal, medical, financial, and other high-impact settings. The term "large" generally refers to the scale of the training data, model parameters, and computational resources, although there is no universal size threshold. LLMs are a major foundation of generative AI, but they are not identical to chatbots: a chatbot is an application, while an LLM is the underlying model that may power it.