Large language model (LLM)

A large language model (LLM) is the AI behind chatbots such as ChatGPT, Claude and Gemini. It has learned from huge amounts of text how language works, so it can summarise a document or draft an email. It writes by predicting, piece by piece, what fits next. Its answers read well but are not always right, and it only knows your prices or customers if you tell it.

How does it work?

An LLM is a neural network with billions of adjustable numbers, called parameters. The survey by Zhao et al. describes LLMs as transformer language models with hundreds of billions of parameters or more, trained on very large amounts of text, and notes that there is no formal agreement on how big a model has to be to count. The transformer design comes from Vaswani et al. (2017). Meta's Llama 3 (July 2024) comes in three sizes, with 8, 70 and 405 billion parameters, and the largest was pre-trained on 15.6 trillion tokens of text (Llama 3 paper).

Training happens in stages. In pre-training the model reads text and learns to predict the next token. A model trained only that way does not reliably follow instructions. OpenAI's InstructGPT work fine-tuned GPT-3 on answers written by people and then with RLHF (reinforcement learning from human feedback). In its tests, people preferred the answers of the 1.3-billion-parameter InstructGPT model over those of the 175-billion-parameter GPT-3 (Ouyang et al., 2022).

When you use the model (inference), it computes a probability for every possible next token, picks one, adds it to the text and repeats (autoregressive generation). Jay Alammar shows this loop step by step for GPT-2. Everything the model takes into account has to fit in its context window.

What can it do?

In 2020, GPT-3, with 175 billion parameters, showed that a large model can do a new task from a few examples in the prompt, without further training, and that larger models do this better (Brown et al., 2020). This is called in-context learning. Wei et al. (2022) call an ability "emergent" when it is present in larger models and absent in smaller ones. Schaeffer et al. (2023) argue that many of those sudden jumps come from the metric researchers chose: measured on a continuous scale, the improvement is gradual.

Where does it go wrong?

• It can state false things with confidence, a hallucination: it produces likely text, and likely text can be wrong.
• Its knowledge stops at a date. In September 2026, Anthropic gives Claude Haiku 4.5 a reliable knowledge cutoff of February 2025 and Claude Sonnet 5 one of January 2026 (models overview).
• It does not know your own data, such as orders or stock levels, until you supply it: in the prompt, through RAG or through function calling.

Which LLMs are there?

Some LLMs can only be used as a service, through an app or an API, such as Anthropic's Claude and Google's Gemini. Others are published as open weights that you can download and run on your own server, such as Meta's Llama 3 (see open weights vs open source). Makers usually offer several models at different prices. In September 2026, Anthropic charges $1 per million input tokens for Claude Haiku 4.5, which it describes as its fastest model, and $4 for Claude Opus 5.5, which it recommends for most work.

Example

Suppose a webshop wants help with customer emails. An LLM can read a complaint and draft a polite reply in Dutch or English straight away. It cannot know whether order 1234 has shipped, because that is not in its training data. So the shop connects it to the order system through function calling and adds its returns policy to the prompt. A staff member checks the drafts at first, because the model can state a delivery time or refund rule that sounds plausible and does not exist.