A large language model (LLM) is a type of AI model trained on massive amounts of text data to understand and generate human language. “Large” refers both to the volume of training data and the number of parameters: the internal variables the model adjusts during training to learn patterns in language. LLMs predict what comes next in a sequence, which turns out to be a surprisingly powerful basis for tasks like answering questions, summarizing content, writing code, and reasoning through problems.
LLMs are a subset of foundation models, which are general-purpose models trained at scale and adaptable to many tasks. Most modern LLMs, including GPT-4 and Claude Sonnet, are made available via API, meaning other organizations can build products on top of them without training their own model from scratch. These models are first trained on enormous amounts of text to learn how language works, then go through a second phase where human reviewers rate responses to teach the model what “good” looks like. This is how they learn to be helpful rather than just fluent.
Why This Matters for Technical Writers
LLMs are the engine behind most of the AI-enabled products and features you’re likely to be documenting right now. Understanding what an LLM is — and what it fundamentally does (predict likely next tokens, not “think”) — helps you write more accurate behavior descriptions, set realistic user expectations, and explain limitations like hallucination without overclaiming or underclaiming. It also helps you ask better questions when working with the engineering teams building on top of them.
Common Confusion
LLM is frequently used as a synonym for “AI” or “generative AI” in general, but it’s more specific than either. Not all AI is generative, and not all generative AI is language-based — image generation models, for instance, are generative but not LLMs. LLMs are also often conflated with the products built on them: Claude is a product; Claude Sonnet is the underlying model, which is an LLM.
Another common confusion is treating LLM outputs as retrieval — as if the model is looking something up. LLMs generate responses based on learned patterns, not stored facts, which is why they can produce confident but incorrect information.
Related Terms
AI model, foundation model, generative AI, transformer, hallucination, RLHF