
Large language model (LLM): what it is and how it works
In short: A large language model is a neural network trained on huge amounts of text to predict the next token, so it can write, answer and reason.
A large language model (LLM) is a neural network, almost always a transformer, the architecture introduced in 2017, trained on huge amounts of text to predict the next token in a sequence. Through that objective at scale, it learns grammar, facts, styles and patterns of reasoning, which lets it write, summarise, translate, answer questions and generate code. Further training on instructions and human feedback turns it into a helpful assistant.
Large language model in plain words
An extremely well-read autocomplete that has seen so much text it can continue almost any conversation, document or program in a useful way, and sometimes a confidently wrong one.
How it works
- Pre-training: the model reads trillions of tokens and learns to predict the next one, adjusting billions of parameters.
- Post-training: instruction tuning and reinforcement learning from human or AI feedback make it follow requests and behave safely. The InstructGPT paper reported in 2022 that people preferred a 1.3B-parameter model tuned this way over the 175B-parameter GPT-3.
- Inference: given a prompt, it generates a response one token at a time, each choice based on everything before it.
Example
As an illustration, give it "Write a polite reminder to a client about an invoice that is 15 days overdue" and it drafts the email. Give it a stack trace and it suggests a likely cause. The mechanism is the same in both cases; only the context changes.
Where you'll meet it
- Assistants such as ChatGPT, Claude and Gemini
- Coding tools, search engines and office software
- APIs from model providers and open-weights models you can host
- Every term in this field guide
What it changes for you
For developers, an LLM is a new kind of component: powerful, probabilistic and in need of tests, guardrails and cost control. For knowledge workers, drafting, summarising and first-pass analysis become fast and cheap, while judgment and verification become the valuable part. For industries, the change is less about the model and more about workflows built around it.
LLM vs foundation model
A foundation model is any large model trained broadly and adapted to many tasks, including image, audio or multimodal ones. An LLM is a foundation model focused on language.
Common mistake
My view: treating an LLM as a database of facts. It is a pattern engine. Give it the facts through context or tools, and use it for what it is good at: language, synthesis and reasoning over what you provide.
Related terms
Token · Context window · Foundation model · Hallucination · Inference
From the archive: How to choose an LLM for your actual workload
An LLM is a pattern engine, not a database: bring the facts, then let it reason.


