Large language models (often abbreviated LLM) are advanced artificial intelligence programs trained on an enormous quantity of textual data so that they understand human language and can use it. Their main job is to process and generate text in a way that is natural and meaningful to a human. Thanks to their extensive training they learn grammar, facts, various writing styles and the logical connections between words. This enables them to answer questions, translate, summarise long documents or even write creative texts.
Put simply, a large language model is like an extraordinarily well-read assistant who has read nearly everything ever written on the internet. When you put a question or a task to it, it does not merely look up keywords but tries to grasp the context and sense of your query. On that basis it then formulates an answer by statistically estimating which word should logically follow the previous one. This is what allows it to hold a fluent conversation, explain complex concepts, or produce texts that feel as though a person wrote them.
Large language models
Technical definition: Large language models are a class of artificial intelligence model, specifically deep neural networks, designed for natural language processing (NLP). They are characterised by an enormous number of parameters (weights), often on the order of billions to trillions, and by training on massive text corpora. Most modern LLMs are based on the transformer architecture, which uses the attention mechanism to capture contextual relationships between words in the input sequence efficiently. The models are typically trained by self-supervised methods, such as predicting the next word in a text. This process lets them learn complex syntactic, semantic and factual representations of language, which can then be specialised for particular tasks (classification, translation, text generation) through a process called fine-tuning.
Etymology and context: The term “large language models” has three parts. Language models are statistical tools for modelling the probability distribution of word sequences. The qualifier large refers to the exponential growth in size (number of parameters) and volume of training data that has been characteristic of models since roughly 2018. That shift was made possible by the ground-breaking transformer architecture introduced in 2017. Large language models such as those in the GPT (Generative Pre-trained Transformer) family represent a major advance in artificial intelligence and are the foundation of modern conversational agents, chatbots and advanced tools for automating work with text.