Large language models, explained
Large language models, often shortened to LLMs, sit behind many writing, question-answering, and conversational AI tools. They turn text into tokens and work with the relationships among those tokens.
A large language model is a deep learning model trained on large amounts of data so it can understand prompts and generate natural language and other content.
An LLM predicts the next token
An LLM does not prepare a complete answer before it starts writing. It predicts one token, adds that token to the sequence, and repeats the process.
The prediction draws on patterns learned during training and the material in the current prompt. That process can produce fluent answers without guaranteeing that each statement is true.
The surrounding application adds control
The model is only one part of an AI application. The application can add instructions, supply business documents, limit available actions, or ask a person to approve a result.
Those controls change what information the model receives and what can happen after it responds. They do not alter the basic next-token process inside the model.
Large language model terms
- Token
- A unit of text, such as a word, part of a word, or punctuation mark, that a model processes.
- Transformer
- The neural network design that helps a model track relationships across a sequence of text.
- Inference
- The process of using a trained model to produce an answer from a new prompt.
What shapes an LLM response
The prompt
The current instructions and context tell the model what response to attempt.
The training
Patterns learned from training data shape which words the model predicts.
The context limit
The model can consider only the material that fits within its context window.
Questions about large language models
It processes a prompt and generates a sequence of tokens. This supports tasks such as drafting, summarizing, classifying text, and answering questions.
An LLM uses statistical relationships learned from data. Its fluent language does not prove human understanding or factual knowledge.
It predicts likely text rather than checking every claim against a source. An application can add trusted information and review steps to reduce that risk.
Related glossary terms
Put plain language to work
Caffeine apps run on the Internet Computer, a public network built for this.