LLM

Large language model (LLM) is a type of artificial intelligence model designed to understand and generate human language. LLMs are typically built with transformer architectures and trained on large, diverse text corpora. During pretraining, they learn to predict the next word or token, acquiring broad knowledge of language, facts, and some reasoning patterns. After pretraining, they may undergo fine-tuning or reinforcement learning from human feedback to align outputs with user expectations or safety requirements.

Capabilities include text generation, summarization, translation, question answering, classification, and code generation. They are commonly accessed via APIs or integrated into applications, and can perform many tasks with little or no task-specific data, a property known as few-shot or zero-shot learning.

Training and deployment considerations include substantial computational cost, energy use, and data curation. Models may store and regurgitate parts of their training data, raising privacy and copyright concerns. They can also generate plausible but incorrect or biased content, a phenomenon known as hallucination.

Evaluation and safety approaches rely on automated metrics and human judgments; reliability varies by domain. Safety methods include content filters, prompt protections, and alignment techniques to avoid sensitive topics and ensure consistent behavior. Open-source and commercial options differ in accessibility, licensing, and support.

Applications span industry and research, including customer support, content creation, tutoring, data analysis, and software development. The field continues to study efficiency, safety, interpretability, and alignment as models grow in size and capability.