Do LLMs understand language?

They model statistical patterns in language extremely well, which produces understanding-like behaviour. Whether that constitutes real understanding is debated; for practical purposes, treat it as a very capable pattern engine with real limits.

Why do they have a knowledge cut-off?

A model only knows what was in its training data up to a point in time. To use newer or private information, you supply it at run time — for example through retrieval-augmented generation.

What are Large Language Models (LLMs)? — Getting Digital glossary

Large Language Models

Also: LLM, foundation model, language model

A large language model is a neural network trained on vast amounts of text to predict the next piece of text, which turns out to be enough to make it write, summarise, translate and reason in language.

One simple objective, surprising range

An LLM is trained on a deceptively narrow task: given some text, predict what comes next. Do that across a huge slice of human writing, with enough parameters, and the model absorbs grammar, facts, styles and a good deal of reasoning as a side effect. The result is a single generative AI model that can be pointed at translation, drafting, coding or question-answering without being retrained for each.

Almost all of them are built on the Transformer architecture, the design that made training at this scale practical. You steer them at run time through prompt engineering, and you keep them factual by feeding in trusted context with retrieval-augmented generation.

Why "foundation" model

Because one pre-trained model serves as a base for many downstream uses — adapted by prompting or light fine-tuning rather than trained from scratch each time.

Key takeaways

→LLMs learn from a single next-token objective, which yields broad language ability as a by-product.
→They are general-purpose: steered by prompts and grounded with retrieval rather than retrained per task.

Related concepts

Broader topicGenerative AI
LLMs are the text branch of generative AI.
Learn firstTransformers
LLMs are built on the Transformer architecture.
Leads on toRetrieval-Augmented Generation
RAG exists to ground large language models.
Related

FAQ

Do LLMs understand language?: They model statistical patterns in language extremely well, which produces understanding-like behaviour. Whether that constitutes real understanding is debated; for practical purposes, treat it as a very capable pattern engine with real limits.
Why do they have a knowledge cut-off?: A model only knows what was in its training data up to a point in time. To use newer or private information, you supply it at run time — for example through retrieval-augmented generation.

Large Language Models

One simple objective, surprising range

Key takeaways

Related concepts

Courses that teach this

FAQ