What is a large language model (LLM)?

Every answer an LLM gives is built one small piece of text at a time, each piece chosen because it is a likely continuation of everything before it. That single mechanism, repeated thousands of times, produces essays, code and conversation. Understanding it explains both what these models do well and why they fail.

How does a large language model work?

A large language model works by predicting the next token, a word or part of a word, given all the text so far. It turns the input into numbers, passes them through many layers of a neural network called a transformer, and outputs a probability for every token in its vocabulary. The application picks one token from that distribution, appends it to the text, and runs the model again. A 500-word answer takes several hundred of these steps.

The model has no separate database of facts and no memory between requests. What it "knows" is stored as patterns in its parameters, the billions of numbers adjusted during training. What it can see right now is limited to its context window.

How are LLMs trained?

Modern LLMs are trained in stages, and each stage changes what the model is good at.

1. 1
   Pretraining

   The model reads trillions of tokens of public text and code and learns to predict the next token. This is where it picks up grammar, facts, reasoning patterns and coding ability. It is by far the most expensive stage.
2. 2
   Instruction tuning

   The model is fine-tuned on examples of instructions paired with good responses, so it answers questions instead of just continuing text.
3. 3
   Preference training

   Humans or other models compare pairs of answers, and the model is trained to prefer the better one, a family of methods that includes RLHF. This shapes tone, helpfulness and refusals.
4. 4
   Reasoning training

   Newer reasoning models are further trained with reinforcement learning on problems with checkable answers, such as maths and code, so they learn to work through problems step by step.

The effect of the later stages is large. In OpenAI's 2022 InstructGPT paper, people preferred the outputs of a 1.3 billion parameter model trained with human feedback over the original 175 billion parameter GPT-3, even though it was more than 100 times smaller.

What makes a language model "large"?

"Large" refers to both the number of parameters and the amount of training data. GPT-3, published in 2020, had 175 billion parameters and showed that a model of that size could perform new tasks from a few examples in the prompt, without retraining. Since then, frontier models have grown further, and labs often no longer publish exact sizes. Scaling laws describe how performance improves predictably as parameters, data and compute grow together.

Parameter

One of the numbers inside the network that training adjusts; large models have billions

Pretraining

The first, largest training stage: predicting the next token on a huge text corpus

Inference

Running a trained model to produce output, which is what an API call does

Foundation model

A large model trained on broad data that many applications build on

Open-weight model

A model whose parameters are published so anyone can run it, such as Llama or Mistral

What can LLMs do well, and where do they fail?

LLMs are strong at tasks where the answer is a plausible, well-formed piece of language: drafting, summarising, translating, classifying, extracting fields from messy text and writing code in popular languages. They are weak wherever being plausible is not the same as being right.

The most important failure is hallucination: a fluent, confident answer that is false. It follows directly from the mechanism. The model is trained to produce likely text, and a false sentence can be very likely.

How do engineers make LLMs reliable?

Engineers make LLMs reliable by controlling what goes into the model and checking what comes out, rather than trusting the model alone. The main tools are clear prompts, retrieval to supply current facts, tool calling for calculations and live data, structured output for machine-readable answers, and evals to measure quality before and after every change.