A Large Language Model (LLM) is a type of artificial intelligence system trained to understand and generate text. It is the technology behind tools like ChatGPT, Claude, and Gemini. Despite the technical-sounding name, the core idea is straightforward: an LLM learns patterns in language by reading enormous quantities of text, then uses those patterns to predict what should come next in a given sequence of words.

How LLMs are trained

Training happens in two main phases:

  • Pre-training. The model reads billions of documents from the internet, books, and other sources. At each step, it is given a piece of text with the last few words hidden, and it learns to predict what those words should be. This process runs on thousands of specialised computer chips (GPUs or TPUs) for weeks or months. By the end, the model has learned grammar, facts, reasoning patterns, and much more — purely by being trained to predict the next word.
  • Fine-tuning. The raw pre-trained model is useful but unpredictable. Developers then train it further on curated datasets of human conversations and use a technique called Reinforcement Learning from Human Feedback (RLHF) — where human raters score the model's answers and the model adjusts to produce higher-rated responses. This is what makes ChatGPT or Claude feel like a helpful assistant rather than a random text generator.

Key concepts you will encounter

Tokens

LLMs do not process text character-by-character or word-by-word. Instead, they split text into tokens — chunks that are roughly 3–4 characters on average. The word "understanding" might be one token; a less common word like "epistemology" might be split into three. Pricing for most AI tools is quoted in tokens (e.g. "$0.003 per 1,000 tokens").

Context window

The context window is how much text an LLM can "see" at once — both your input and its own previous output. Early models had context windows of 4,096 tokens (about 3,000 words). Modern models like GPT-4 and Claude 3 support hundreds of thousands of tokens, making it possible to process entire books in a single prompt. A larger context window is directly useful for tasks like summarising long documents or maintaining long conversations.

Parameters

Parameters are the numerical values inside the model that encode everything it has learned during training. Model sizes are described in billions of parameters: GPT-3 had 175 billion, while more recent models are rumoured to be in the trillions. More parameters generally mean more capability, but also more compute cost to run.

Temperature

Temperature controls how "creative" or unpredictable the model's output is. At temperature 0, the model always picks the most statistically likely next token — outputs are deterministic and repetitive. At higher temperatures, the model samples from a wider range of possible tokens, producing more varied and sometimes more creative (but also more erratic) text. Most AI writing tools let you adjust this slider.

What LLMs are good at

  • Summarising and rewriting text
  • Translating between languages
  • Answering questions based on context you provide
  • Writing and editing code
  • Drafting emails, reports, and marketing copy
  • Explaining complex topics in simple terms

What LLMs are NOT good at

  • Reliable factual recall. LLMs often "hallucinate" — they produce confident-sounding text that is factually wrong. Never rely on an LLM for citations, statistics, or anything requiring verified data without checking independently.
  • Real-time information. A model's knowledge has a cut-off date. It cannot tell you today's stock price or yesterday's news unless it has a web-search tool attached.
  • Consistent reasoning over very long chains of logic. For tasks that require many sequential steps of precise reasoning (e.g. complex mathematics), models still make errors.

The major models in 2026

The landscape has converged around a handful of frontier providers:

  • OpenAI — GPT-4o and the o-series reasoning models
  • Anthropic — Claude 3 and 4 series (focused on safety and long context)
  • Google — Gemini 1.5 and 2.0 (integrated with Google Workspace)
  • Meta — Llama 3 (open-weights, runs locally)
  • Mistral — Efficient open-source models

Most AI tools you evaluate on this site use one of the above models as their underlying engine, then build a specialised product on top — a writing assistant, an SEO tool, a coding helper, and so on.

Summary

An LLM is a statistical engine trained to predict plausible text. Its intelligence is real but different from human intelligence: broad pattern recognition rather than deep understanding. The practical implication is that LLMs are powerful accelerators for writing, coding, and information work — but require careful prompting and human review to be used reliably. The next guide covers exactly that.