What is a large language model?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
A large language model, or LLM, is an AI model trained on extensive collections of language to recognize patterns and produce text. A model is the mathematical component that turns information it receives into an output. LLMs can support tasks such as answering questions, summarizing documents, translating languages and writing computer code.
Many LLMs are built as neural networks: connected layers of calculations whose internal numerical settings change during learning. These settings are called parameters. “Large” usually refers to the scale of the model and its training, but there is no single parameter count that universally separates a large language model from a small one.
Text is divided into tokens: small pieces that can be words, word fragments or punctuation. This does not mean the model must read the input one piece at a time. Transformers, a common design for language models, can process many input tokens in parallel, although many generate responses one token after another. A common training task is predicting the next token from the preceding text. The text itself supplies the target, an approach called self-supervised learning. Repeating this task across many examples teaches patterns that can support much more than sentence completion.
Additional training can teach it to respond more usefully to instructions. For example, developers can supply demonstrations of desired answers or use people’s comparisons of alternative responses. These steps help explain why a conversational assistant behaves differently from a system that merely continues unfinished text.
When you use the model, your instructions and supplied information influence the answer. This material is called context. The model combines patterns learned during training with the information available in that interaction. It does not automatically know everything that has happened since training or have access to your documents.
An application can connect it to search or other tools, meaning software functions that retrieve information or perform actions. Those connections are additional capabilities. A model generating a sentence about checking a website is not the same as an application actually visiting that website and reading it.
In a discharge-summary task, an LLM may produce fluent wording even when the source record omits an important medication change or follow-up test. The appropriate output is to flag the gap, not complete the story with a typical clinical course. This is why supplying the actual record, preserving uncertainty and checking the summary against its sources matter. A convincing clinical narrative is a language output; whether it accurately represents this patient is a separate question.
Sources: Vaswani and colleagues — Attention Is All You Need; Google — Introduction to Large Language Models; Ouyang and colleagues — Training Language Models to Follow Instructions with Human Feedback.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy