What a Large Language Model Is
We unpack what an LLM is, how it 'thinks', and why its answers are statistics rather than knowledge.
What an LLM is
A large language model (LLM) is a program trained to predict the next word in a piece of text. Show the model billions of pages of text, and it learns to continue any phrase statistically. ChatGPT, Claude, Gemini, and dozens of other assistants are all built on this simple mechanism.
What matters from the very start
- The model doesn't know facts. It knows statistical patterns of language. If the correct answer appeared often in the training data, it will probably surface. If it appeared rarely, the model will confidently invent a plausible but false answer. This is called a hallucination.
- The model doesn't remember your past conversations by default. Each chat is a fresh context window. Whatever you wrote at the start of a conversation stays inside it; a new chat starts from zero.
- The model has a "cutoff date". It was trained on data up to a certain point. It doesn't know recent events (news, releases) unless you give it search or paste the context in yourself.
Where the "smartness" comes from
The model doesn't "understand" in the human sense. It shines wherever the task reduces to the combinatorics of language: rewrite, summarize, classify, continue. The more a task is tied to stable patterns in text (writing, code, law, documentation), the better the result.
For tasks that need precise, fresh information or independent fact verification, you need extra tooling: search, RAG, tool use, human review. Treat the raw model as a fast first draft, not as a source of truth.
Take a task from your routine that eats 30+ minutes. Describe it in one sentence. Hand it to ChatGPT/Claude and compare the result against your own reference output.
Copy and adapt to your context. Text in angle brackets should be replaced.
You are my assistant. I have a task I do often. Before doing it, ask me 3-5 clarifying questions so you don't invent facts. If you don't have enough information, say so explicitly. Task: <paste the description>
- People assume the model "knows" things — in reality it completes them statistically.
- They don't specify audience or format — and get a generic, faceless answer.
- They accept the first answer as final, even though the prompt could be improved.
- If an answer feels uncertain, ask the model to explicitly list what it does NOT know.
- Ask the model to ask clarifying questions first, before answering.
- For important tasks, ask it to respond as a skeptic who hunts for holes.
Any language task: rewrite, explain, build a list, turn notes into an email, brainstorm options, answer a routine question.
When you need verifiable real-time facts and have no search tool. When the answer drives a legal/medical decision with no human check.