Large Language Model (LLM)
As of:
A large language model (LLM) is a neural network trained on very large amounts of text to predict the most probable next token. From this simple training objective emerge capabilities such as text comprehension, translation, summarisation and code generation — making LLMs the technical foundation of virtually all current generative AI applications, from chatbots to AI agents.
Technically, today’s LLMs are based on the Transformer architecture, whose attention mechanism models relationships between all tokens of an input. Text is split into tokens and processed internally as embeddings; the maximum input length defines the context window. Pretraining on raw text is usually followed by an adaptation phase (fine-tuning, for instance with instruction data and human feedback) so that the model follows instructions and avoids undesirable outputs.
In practice, one thing matters most: an LLM is not a database. It stores knowledge implicitly in its parameters, has a training cutoff date and can produce plausible-sounding but false statements (hallucination). For current or company-specific information it is therefore often combined with external knowledge, for example via RAG. A common misconception is that more parameters automatically mean better results — data quality, training and deployment context are at least as decisive.