What is an LLM? A practical explanation
The mental model you need before you read benchmarks or model cards
What an LLM is (plain English)
An LLM (Large Language Model) is software trained to predict the next token. A token is a small chunk of text (often part of a word).
When it keeps predicting tokens again and again, you get sentences, paragraphs, code, tables, or structured output.
A helpful mental model: “super autocomplete”
Think of it like autocomplete that learned from a massive library of writing. By default it doesn't “look up” facts like a database does—it generates what usually comes next based on patterns in your prompt.
Why it can feel intelligent
Language has a lot of structure. If a model learns enough patterns, it ends up learning grammar, common knowledge, typical reasoning steps, and standard formats (emails, docs, JSON, Python) as a side effect.
What it's great at
- Writing and rewriting in a consistent tone (friendly, formal, concise, etc.).
- Summarizing and reorganizing long content into something easier to scan.
- Turning messy notes into clean structure (headings, bullet points, tables).
- Producing code scaffolds, then iterating with you to refine them.
Where it commonly goes wrong
- It can sound confident while being incorrect (especially on niche facts).
- It may fill in missing details you didn't confirm (assumptions).
- It can produce code that looks right but doesn't match your real constraints (schema, business rules, edge cases).
A reliable workflow
1. Ask for the answer in the format you want (bullets, JSON, steps). 2. Ask: “List assumptions you made.” 3. Ask: “List what you're uncertain about and how to verify it.” 4. Treat verification as part of the task: check sources, run tests, and validate against your domain rules.