Short answer
Large language models are deep neural networks trained on massive text corpora to predict tokens, enabling chat, coding help, summarization, and tool-using applications.
Why it matters
Article
What it is
A large language model is a deep neural network—typically a transformer—trained to model natural language (and often code) at scale. “Large” refers to parameter count and training data volume relative to earlier NLP models. LLMs can be prompted to perform many tasks without training a separate model for each task, though specialized fine-tunes still matter for reliability and domain fit.
LLMs are generative sequence models: they assign probabilities to possible next tokens and can sample continuations. Product layers add chat templates, tool calling, retrieval, and safety policies that shape what users experience.
Plain-English explanation
An LLM is extraordinarily good at continuing text in a way that looks coherent. If you start a sentence, it predicts likely next words; if you ask a question, it continues with something that resembles an answer. That resemblance is learned from patterns in training data and subsequent tuning, not from a guaranteed internal database of facts.
Because language encodes so much human knowledge and procedure, next-token prediction becomes a surprisingly general interface: summarize this, rewrite that, explain a concept, draft code. The same mechanism also produces mistakes that sound authoritative, which is why verification habits are part of competent use.
How it works
Most modern LLMs use the transformer architecture. Text is split into tokens; each token is mapped to an embedding vector; attention layers let tokens exchange information across the sequence; feed-forward layers transform representations; and a final head predicts the next-token distribution. Training minimizes prediction loss over enormous corpora. Context windows define how much prior text the model can consider at once.
After pretraining, models are often fine-tuned with supervised instruction data and preference optimization so they follow directions more helpfully and safely. At inference, decoding parameters (temperature, top-p) trade diversity against determinism. Advanced systems may call tools—search, calculators, code execution—to complement parametric knowledge.
Why it matters
LLMs changed software interfaces by making natural language a programming surface for knowledge work. They accelerate drafting, analysis, tutoring, and developer workflows. They also concentrate questions about copyright, labor, evaluation, energy use, and dual-use risks. Understanding token limits, hallucinations, and grounding strategies is now basic digital literacy for many professions.
Common misconceptions
- “The model looks up answers on the web by default.” Base generation uses parameters learned during training unless browsing or retrieval is explicitly connected.
- “Longer prompts are always better.” Extra context can help or distract; structure and relevance matter more than length alone, and context windows are finite.
- “Fine-tuning is the only way to specialize.” Retrieval, tools, and careful prompting often provide strong specialization with less operational cost than full fine-tunes.
Further learning
See generative AI, natural language processing, and neural networks. Stanford CS224N materials explain modern NLP and transformers in academic depth; OpenAI’s platform guides describe practical text-generation concepts used in applications.
Encyclopedia connections
This topic sits in the llms branch of the Artificial-Intelligence.info encyclopedia. Use related guides on training vs inference, responsible AI, and hands-on tooling pages to go deeper without treating any single article as complete expertise.
Examples
Instruction following in chat
Code assistance with tests
FAQ
What is a token?
Why do LLMs hallucinate?
Related glossary terms
- AI agent — A system that pursues goals by choosing actions over multiple steps, often using tools, memory, and feedback from an…
- Embedding — A numeric vector representation of an item—such as a token, sentence, image, or document—used to capture similarity and…
- Fine-tuning — Additional training that adapts a pretrained model to a narrower task, domain, or preferred behavior using specialized…
- Generative AI — AI systems that create new content—such as text, images, audio, or code—by modeling patterns from training data.
- Hallucination — When a generative model produces fluent content that is fabricated, unsupported, or inconsistent with provided sources.
- Inference — The process of running a trained model on new inputs to produce predictions or generations, usually without updating…
- Large language model — A large neural network, typically a transformer, trained on massive text corpora to predict tokens and perform language…
- Multimodal — AI systems that process or generate more than one data type—such as text, images, audio, or video—within a unified…
- Natural language processing — The AI subfield focused on analyzing, understanding, and generating human language for tasks like translation, search…
- Prompt — The input text or structured instructions given to a generative model to steer its output toward a desired task or…
- RAG (retrieval-augmented generation) — A pattern that retrieves relevant documents and conditions a generative model on that evidence to improve grounded…
- Token — A chunk of text (often a word piece) used as a basic unit for language model training, context limits, and generation.
- Transformer — A neural network architecture that uses attention mechanisms to model relationships among tokens in a sequence…
Sources
-
CS224N: Natural Language Processing with Deep Learning
Stanford UniversityUniversity course covering word vectors, transformers, and modern NLP systems underlying LLMs.
-
Text generation guide
OpenAIOfficial conceptual documentation for LLM text generation behavior.