What is a large language model (LLM) and how does it work

The definition, straight up

A large language model (LLM) is an AI system trained on enormous amounts of text to predict and generate human language.

When you type a question to ChatGPT, Claude or Gemini, you're interacting with an LLM. The model doesn't "think" like a human — it calculates, with very high precision, which words or phrases are most likely to be a coherent and useful response to your input.

How does an LLM work internally?

Without going into the maths: an LLM learns from patterns in text. A lot of text.

1. Training

The model is fed billions of pages of text: books, articles, websites, code, conversations. During training, it learns which words tend to appear together, which structures are coherent, and how to respond to different kinds of questions.

GPT-4o, Claude Sonnet 4.5 and Gemini 2.0 were trained on amounts of data exceeding everything a human could read in thousands of lifetimes.

2. Tokens, not words

LLMs don't process whole words — they process tokens, which are fragments of text. The word "intelligence" might be 2 or 3 tokens. One token is roughly 0.75 English words.

The context window size (how many tokens the model can "remember" in a conversation) determines how much text it can process at once: ChatGPT handles 128K tokens, Claude Pro up to 200K, Gemini up to 1 million.

3. Transformers: the key architecture

All modern LLMs are based on the Transformer architecture, published by Google in 2017. Transformers process text in parallel and use a mechanism called "attention" to understand which parts of the text are relevant to each other.

You don't need to understand the technical details to use them — but knowing they exist helps you understand why LLMs are so good at context and so bad at exact maths.

Why do LLMs sometimes get things wrong?

Hallucinations

An LLM can generate text that sounds completely confident and plausible but is wrong. This is called hallucination. It's not a bug — it's a consequence of the prediction mechanism: the model generates what "sounds right", not what "is true".

That's why Perplexity and Gemini cite sources: they add a web search layer to anchor responses to verifiable data.

Knowledge cutoff

LLMs have a training cutoff date — the point up to which data was processed. GPT-4o and Claude Sonnet 4.5 have a cutoff of August 2025. What happened after that, they don't know — unless they can search the internet.

Maths and exact reasoning

LLMs are bad at exact calculations. They are language prediction systems, not calculators. That's why ChatGPT Plus includes Code Interpreter: it delegates calculations to Python, which is exact.

The most important LLMs in 2026

LLM vs generative AI: are they the same?

Not exactly. Generative AI is the broad concept: systems that generate new content (text, images, video, audio). LLMs are a specific type of generative AI focused on text and language.

• ChatGPT is an LLM (with added image and voice capabilities)
• Midjourney is generative AI but NOT an LLM (it generates images, not text)
• DALL-E 3 is generative image AI, integrated inside ChatGPT

Why does knowing this make you a better user?

Understanding what an LLM is makes you better at using these tools:

1. You know when to trust it: if the model can't verify something, it might hallucinate it. Ask for sources when data matters.
2. You know how to ask: LLMs respond better to clear instructions and abundant context. More context = better response.
3. You know its limits: don't ask for exact calculations without Code Interpreter. Don't expect today's news if the model doesn't have web access.

FAQ

Do I need to know how to code to use an LLM?

No. ChatGPT, Claude and Gemini interfaces are conversational — you write in natural language and the model responds. Programming is only needed if you want to access models directly via API to build your own applications.

Which is the most powerful LLM in 2026?

It depends on the task. Claude Sonnet 4.5 is best for writing and reasoning. GPT-4o is the most versatile. Gemini 2.0 is best for up-to-date information. There's no "best" in absolute terms — there's the best for each use case.

Do LLMs learn from my conversations?

By default, commercial models like ChatGPT or Claude may use your conversations to improve their future models, but you can disable this in privacy settings. If you handle sensitive data, review the provider's privacy policy and consider using local models or privacy-guaranteed models like Mistral.

→ Want to pick the right LLM for your needs? Read How to choose the best AI chatbot.