A large language model (LLM) is a neural network trained on very large amounts of text to understand and generate natural language. Technically, LLMs are built on the transformer architecture, which processes sequences of words in parallel and uses attention mechanisms to capture statistical relationships between words.
At its core, an LLM is a probability machine: it repeatedly predicts the next token in a sequence, producing text that follows the patterns learned during training. Training is largely self-supervised on unlabeled data; models are then often refined for specific tasks through fine-tuning and alignment with human feedback.
LLMs handle a wide range of tasks — summarization, translation, classification, code generation and dialogue — without being reprogrammed for each one. However, their outputs are not fact-checked: when the required knowledge is missing, the model may hallucinate. Capability typically scales with model size, training-data volume and compute.